Manuscript revision and pending experiment work, snapshot before restructuring

Clarity pass over the main text (36-item audit), Discussion rewrite and cut,
acknowledgements, Souly et al. as ref 62, lettered SI panels, model section
moved under Results; plus the untracked curriculum/society/compose/smol
configs, runners, figures, stats and tests that the SI already cites.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y64o8FKP7rCuXzC48pxpMm

tasks/clarity-audit-2026-09-13.md

@@ -0,0 +1,110 @@
+# Clarity audit of paper/pnas/main.md (2026-09-13)
+
+Standard: an interpretive sentence must state the concrete formula, number or mechanism it refers
+to; figure citations must match what the figure plots; no herald sentences; no process ghosts.
+Three parallel audits (Intro+model; Results 1–4; Results 5–6 + Discussion). Line numbers refer to
+main.md at the time of the audit. Nothing below has been applied yet.
+
+## Tier 1 — factual or self-contradictory (fix regardless)
+
+1. L131–132 "no amount of merging can recover it (Fig. S3)". Fig. S3 is re-baselining (E6); it has no
+   merging arm. It shows a population that adopts its own collapsed output as reference never regains
+   lost items. Recite it for that.
+2. L595–598 Discussion: replay fractions "sit where the inheritance model's operational threshold
+   lies". Contradicts the count-not-fraction closed form. Rewrite: they bracket 5% at n=200, and each
+   delivers tens to thousands of replayed examples per step, past the ten copies that hold 95%.
+3. L617–620 "past a threshold, become irreversible". Reintroduces the retracted threshold. The
+   irreversibility condition is concrete: every copy gone from every parent and source (Fig. S3).
+4. L276–277 "(Fig. 3B–C)" cited for a 2×2 (two sizes × easy/hard) comparison; 3B is 0.5B easy
+   merge-vs-specialist, 3C is 7B hard routing-vs-average. Cite precisely; point to Table S2 for the rest.
+5. L121–122 autoencoder "collapses faster (Fig. 2)". The comparison with drift is Fig. S2A–B; Fig. 2A
+   shows the collapse. Cite both.
+6. L107–110 three closed forms named, one written. The union formula `T[ρq + (1−ρ)(1−(1−q)^K_T)]`
+   appears nowhere in the paper, yet Table 1 labels that row "closed form". Write it (here or in the
+   merging section).
+7. L541–544 the in-sample rank correlation of the conflict predictor is never given (only its CI and
+   the held-out 0.35–0.40). Insert the value from Table S2.
+8. L263 "at any rarity (Fig. S8)" — supported only at the tested rarities.
+
+## Tier 2 — result named but not stated / revelation lands vague
+
+9. L246–256 Jenkin/blending null: "exact description" asserted; the conservation law announced
+   without saying what is conserved (expected rare-item mass q·p in the child, independent of K).
+10. L277–283 headroom definition is near-tautological ("routing wins when routing would score
+    higher"). State what sets it: 7B-easy at ceiling (1.00 on two families) → nothing to recover;
+    7B-hard average 0.41 vs routing 0.50 in every seed; 0.5B same gap on easy tasks.
+11. L235–237 "twice the fraction ... the measured price of the estimator bias" → 10% vs 5%, because
+    sharpening loses rare modes faster than sampling alone.
+12. L636–637 "the ablation shows what removing it does" → 0.48 vs 0.78, confident and wrong.
+13. L607–612 "equilibrium theory / failure theory / prediction" labels → give the three contents
+    (m ≈ 1/p per step, 2m/(2m+1) kept; self-replay = g=0; functional disagreement predicts where
+    weight distance does not, with shared-data control).
+14. L492–494 "what moves the cliff" never stated → share of shared prompts under contradictory
+    conventions (Fig. 5B).
+15. L447–452 six-generation ceiling never named → what one adapter holds (0.80 for a single model
+    taught the whole syllabus).
+16. L577–579 and L633–635 Discussion "headroom"/"composed" carry the result → give the margins
+    (routing +0.09, screened offspring +0.07 on hard tasks, every seed; parity at ceiling).
+17. L513–515 "paid this floor" — floor undefined in main text → obligate-merge arm from generation 3.
+18. L148 Table 1 "(the headroom rule)" used before defined (L281).
+19. L147 Table 1 "Consequence-level only" is internal shorthand.
+20. L318–321 conformity term never motivated → consensus = learning from own outputs when there is
+    no verifier; g=0 rewards agreement with itself.
+
+## Tier 3 — heralds, process ghosts, jargon
+
+21. L73–76 three programme sentences announcing the paper.
+22. L82 "Population genetics prices each decision."
+23. L83–86 Fig. 1B "society in time rather than in space" — say what the shift buys.
+24. L221–225 "What these runs add is the comparison ... exposes two departures".
+25. L254–256 herald before the proposition.
+26. L496 "The pre-registered emergent test constrains the claim most."
+27. L331–333 "partly built in ... could alter the picture" (review-response ghost) → separability point.
+28. L560–563 and L656–659 "does not validate a specifically population-genetic mechanism" said
+    twice; keep one, as a statement about the subject.
+29. L468–471 semicolon chain, "count-to-effect link"; the quadratic snowball is Orr (43), not Fig. 5E–F.
+30. L292–294 clonal interference in shorthand → spell out: two variants in different individuals
+    never meet in an asexual descendant.
+31. L346–348 "the Lamarckian channel biology forbids and engineering permits" → one clause of anchor.
+32. L273–275 nonlinearity caveat leaves out why the analogy holds (1/N scaling of an update, 67).
+33. L103–106 Wright–Fisher used before defined; "learner, not organism" antithesis.
+34. L93–94, L98–99, L111–112 housekeeping in reader-facing text ("I use the word throughout",
+    "allele frequency of the dictionary", "standing tests in the codebase").
+35. L372–413 fourteen numeric pairs in one paragraph; split at "Three things rise with generation".
+36. Fig. 3D legend colour order vs text order (conflict, overlap, duration) — check they match.
+
+
+## Proposed rewrites (from the audits; GG's voice to be checked before applying)
+
+1. "Once every copy of a rare item is gone from all parents and all sources nothing can rebuild it, and the inheritance model shows the trap in its commonest form: a population that adopts its own collapsed output as its new reference never recovers the items it had lost, whatever real data it is fed afterwards (Fig. S3), so remedies must act while copies still survive somewhere."
+2. "The replay fractions the field settled on empirically (about 1% for instruction tuning, 89; 5% to 25% in continual pretraining, 90) bracket the 5% found here at 200 samples per generation, and the closed form says why they scatter: at typical batch sizes each delivers tens to thousands of replayed examples per step, well past the ten copies per generation that hold 95% of a source's diversity, so the number that matters is the count of replayed examples of each skill, not the fraction."
+3. "Evaluation that averages over capabilities hides exactly the losses drift predicts first, the rare ones, and a rare capability is recoverable only while some parent or source still holds a copy (Fig. S3). Monitoring the tail, the accuracy on the rarest items rather than the mean, is therefore the leading indicator."
+4. Cite "(Fig. 3C for 7B on hard tasks; the 0.5B and 7B-easy comparisons in Supplementary Information, Table S2)" — verify the SI location.
+5. "(Fig. 2A; the comparison with drift in Fig. S2)".
+6. "(the heterozygosity decay `E[H_t] = H_0(1 − 1/n)^t`; the stationary diversity under real data, written in the next subsection; and the expected number of rare items held by at least one of `K_T` parents, `T[ρq + (1−ρ)(1−(1−q)^K_T)]`, used in the merging section)".
+7. Insert ρ and CI from Table S2.
+9. "Averaging two models does to a rare capability exactly what Jenkin said blending would do to a rare variant: a child fit to the mean of `K` parents sees the item `K` times more often in the mixture and at `1/K` of its mass when it does, and for a rare item the two cancel. Blending inheritance is therefore the null model of merging, and the proposition below states the cancellation exactly." Proposition lead: "In the inheritance model the dilution is a conservation law: the expected mass of a rare item in the child is `q·p` whatever the number of parents, so averaging over more parents neither helps nor harms a rare item's survival."
+10. "Routing wins by exactly the amount averaging loses to dilution, and two things set that loss. A strong base on easy tasks has none: after averaging, the 7B model scores at ceiling (1.00 on two of three families), and routing has nothing to recover. Hard tasks restore it: at 7B the average falls to the level of the best single specialist (0.41), because each specialist's own skill is diluted, and routing among the intact specialists scores 0.50, ahead in every seed. A weak base (0.5B) shows the same gap on easy tasks. The variable is headroom, the distance between the average and the ceiling, and neither model size nor task difficulty alone."
+11. "The autoencoder needed about 10% real data where the inheritance model needed 5%; the difference is what its sharpening bias costs, since a learner that concentrates mass on common modes loses rare ones faster than sampling alone would."
+12. "...and removing it is the one ablation that fails outright: a population selected on agreement with its own consensus instead of on the verifier settles at 0.48 against 0.78 for the full society (Fig. 4D–F), confident and wrong."
+13. "For continual learning the results give three things. The replay ratio has a formula: `m ≈ 1/p` examples per step of the rarest skill one refuses to lose, and 2m/(2m+1) of the diversity is kept. Replaying a network's own output (33, 91–93) is grounding with `g = 0` and collapses on the timescale of Fig. 2, one hop being too short to see it. And a pre-merge test (functional disagreement on shared probes) predicts interference where weight distance does not, with a control for shared training data that the regression (86) and distance (87, 88) studies lack."
+14. "That no single model can answer one prompt two ways is a matter of information, not of training (SI Text S1, Proposition S2). What the population view adds is where the cliff sits: it moves with the share of shared prompts under contradictory conventions (Fig. 5B), and in a population that share grows whenever lineages adopt conventions independently."
+15. "Under a curriculum that delivers every skill to every lineage the ceiling is what one adapter can hold (0.80 for a single model taught the whole syllabus), and sex and selection each reach it sooner without raising it."
+16. "*Route or screen rather than average whenever the average falls short of the best parent on any task*: on hard tasks routing beat averaging by 0.09 in every seed and screened offspring by 0.07 (Fig. 3C), whereas on tasks the 7B base already answered at ceiling the plain average matched them and nothing was lost." / "Blending dilutes whichever parent's skill is rarest, so routing and offspring screening pay only where the plain average falls short of the best parent, and were inert where it did not (Fig. 3B–C)."
+17. "...and this is the cost the obligate-merge arm of the six-generation population paid from generation 3 onward, when its partners began carrying opposite conventions for the same prompts (Fig. 4B)."
+18. Table cell: "Fig. 3B–C: merging beats blending whenever the weight-average scores well below the best parent, and blending suffices when it does not".
+19. Table cell: "The irreversibility is reproduced (Fig. S3); the mutational mechanism of the ratchet is not modelled, see (30)".
+20. "...*grounded evaluation*: an agent is scored partly against reality and partly against the population's own consensus (`g`·true-fitness + (1−g)·conformity). The consensus term stands for what a population does when it has no verifier, which is to learn from its own outputs, so `g = 0` is a population that rewards agreement with itself."
+21. "Drift is only the entry point, because population genetics is above all a theory of what keeps a population from decaying (immigration, recombination, selection, population structure) and of where each of those fails, and every one of them has a counterpart an operator of a model population can switch on: real data entering each generation, merging, verifier-anchored selection and the choice of which models merge with which."
+22. "...transposed from a single network to a population whose members inherit from one another, and each of them has a population-genetic answer with a number attached (how many real samples, how far the average sits below the best parent, how much the parents disagree on shared inputs)."
+23. "Fig. 1B draws the change of viewpoint the transfer rests on. Models are usually pictured as a society in space, contemporaries exchanging messages. The couplings that matter here (training on model output, merging, real data entering each generation) run between generations, and a society coupled in time is what population genetics describes."
+24. "Compared against the exact model, the trained networks depart in two ways, both consequences of the estimator bias measured above. The threshold softens: ..."
+26. "The sharpest test is whether isolation emerges with no conflicting signal anywhere, as a true Bateson–Dobzhansky–Muller incompatibility would: children were diverged..."
+27. "The arm without grounding fails by construction, since a rule that scores agreement will converge on agreement; what the ablation adds is that the other two removals fail in different ways, so recombination and diversity are not substitutes for grounding or for each other. Magnitudes depend on the mutation, restart and selection schemes, which were not varied."
+28. Keep one, in the Discussion: "Three of the framework's refinements failed (confidence weighting, the modifier reading of declines, selection turning speed into level), and the results are consistent with any account in which rare items are lost by sampling and conflicting conventions cannot share weights. The population-genetic reading earned its place by supplying the nulls and the overlap control, not by being the only mechanism left standing."
+29. "In the inheritance model (Fig. 5E–F) hybrid fitness stays at the parents' level while the lineages remain compatible and then falls below the ancestor, sooner the more incompatibilities the genomes carry. Orr showed the number of such incompatibilities grows with the square of divergence (43). Whether a growing number of conflicts produces a fall in performance in a trained network is the question the simulation cannot answer."
+30. "...the *Fisher–Muller effect* (37, 38). In an asexual population two useful variants that arise in different individuals can never meet in one descendant; the lineages carrying them compete, and one is lost. Recombination puts both into one offspring, which is why sexual populations adapt faster."
+31. "...so what the parent learned in its lifetime passes to the child, the inheritance of acquired characters that Lamarck proposed and biology rejected, and that a weight file makes trivial."
+32. "(a network is nonlinear in its weights, so averaging weights does not average outputs; but an update held by one of `N` parents is still scaled by `1/N` in the average (67), which is the dilution the proposition describes)".
+33. "The resampling step is the Wright–Fisher process, population genetics' canonical model of neutral evolution, in which each generation is a random sample of size `n` from the last. In this *inheritance model* the Wright–Fisher population is the sample a child is trained on and its individuals are the `n + m` samples; it is a model of a learner."
+34. "An item is the counterpart of an allele; a *capability* is what an item stands for." / "(allele frequency, in Table 1)" / cut or "(Methods)".

tasks/lessons.md

@@ -50,3 +50,120 @@ Rules: (1) never mutate `.venv` while a background job is running against it —
 throwaway venv elsewhere; (2) `uv sync` is *declarative* — always pass **every** extra the project
 needs, or it removes the ones you omit; (3) the repo now pins `.python-version` (3.14) so the
 interpreter can never drift silently.
+
+## Before a compute campaign, derive the design from the theory and read the analytic code beside the neural code (2026-09-07)
+The v1 `llm_society` campaign (4 seeds, ~13 L40S-h) returned a null that was *structurally guaranteed*:
+3 families over 8 agents made founders near-clones (E8's ρ=1 control: recombining clones buys nothing);
+2^3 competence states left no room for a child to be "new" (E8 needs L=12); sex was a linear blend at
+0.5B (E4/`llm_moe`: the dilution regime); parents were truncated before breeding (E11 selects on
+survival over parents+offspring — v1 threw away half the families at gen 1 with no operator to
+restore them); `n_test`=40 put every contrast inside one SE. GG caught the first fault by asking what
+the founders knew; the rest fell out of comparing `dynamic_society.py` with `society.py` line by line.
+Rules: (1) a Layer-2 instantiation of an analytic experiment must be checked *operator by operator*
+against the analytic code, not against its description; (2) every free parameter that the theory
+constrains (ρ, L, the operator regime, the observation floor, selection intensity) is set by a
+prediction or a calibration measurement, never by feel; (3) write the falsifiers with numbers and the
+power analysis *before* submission — the pre-registration is `tasks/prereg-llm-society-v2.md`.
+
+## Calibrate the inheritance channel before the population (2026-09-07, evening)
+Nine local GPU-hours of calibration found three ceilings a 96-GPU-hour campaign would have hidden:
+(1) with 3 skills over 8 agents, founders were near-clones (E8 ρ=1 → recombination buys nothing);
+(2) a fresh LoRA distilled from a one-skill parent's answers on nine families retains only 0.6–0.8 of
+the skill — interference from confident off-expertise answers, not the E2 observation floor, and
+removable by gating on the source's own confidence; (3) a two-skill child holds each skill at
+~0.85× of its parents at ANY training budget (rank 64 overfits) — the learning budget, not the
+sample budget, is the conserved quantity, and it caps how many skills one adapter can carry. GG's
+call was no-go at 0.5B rather than a campaign that could only test the ablations. Rules: (a) measure
+transmission fidelity of the inheritance channel for one skill, then two, before breeding populations;
+(b) when a gate fails, re-derive it from the data you already have (the conflict gate moved 0.35→0.41
+from the epistasis grid's own outcomes) rather than by feel, and record it as an amendment; (c) a
+pre-registration that ends in a no-go has done its job — write the ceiling up, don't route around it.
+
+## A plan for reader-facing prose must carry the dual-audience standard explicitly (2026-09-09)
+GG rejected the approved-in-substance manuscript-revision plan until it stated, as a first-class
+section, that every term from either field is defined at first use with an example from each world.
+The plan had the right content and structure but treated accessibility as a verification
+afterthought; GG's rule is that it is "paramount" and must be designed in, not checked for. Rule:
+before drafting any passage for a mixed readership, build the term table (term / one-clause
+definition / biology example / model example) *in the plan*, and put a two-reader pass in
+verification. The same applies to my status reports — GG twice said "I am lost with all these C3,
+E9, H1"; spell codenames out.
+
+## 2026-09-11 — check the figure panel inventory before flagging cross-references
+In the manuscript review I flagged Table 1's "Fig. 4C–E" and "Fig. 3C" as inconsistent with the text. They were
+correct: `make_figs.py` puts the E11 ablation in Fig. 4's bottom row and the predictive grid in Fig. 3C–D; the text
+simply failed to cite them. Rule: before calling a cross-reference wrong, read `paper/pnas/make_figs.py` and the
+captions in `build.py` for the panel inventory; the fix is usually a missing citation in the text, not a wrong table.
+
+## 2026-09-11 — never type a result number that a script has not printed
+Writing the stop3 README I filled the per-seed cells from memory of the mean and had three rows wrong
+until the loader's pivot table exposed it. Rule: every number in a README, SI table or manuscript is
+pasted from a stats-script printout produced in the same step; if the script has not printed it,
+print it first. This is the same rule the plan stated ("copied from the stats-script output, not
+typed") and I broke it within the hour.
+
+## Reference numbers hardcoded outside the renumber path (2026-09-11)
+`renumber_refs.py` rewrites main.md, si.md and build.py captions, but `make_figs.py` carries a literal "(refs. N, M)" in the fig1a grounding cell, which went stale after references were added. Rule: after any renumber, grep `refs\.` and `ref\.` across `paper/pnas/*.py` and fix by hand (or extend renumber_refs.py to cover make_figs.py). Also: fig text is rendered, so verify by `pdftotext figs/fig1a.pdf`, not by grepping the source alone.
+
+## 2026-09-12 — GG's manuscript comments: heralds, undefined terms, and phantom SI references
+Forty-five comments on the ODT. Three patterns. (1) **Herald sentences** ("Two boundaries follow.",
+"Modifier theory predicts its fate:", "X is the measurement no other arm produces") — GG: "Breaking
+down sentences like this is also a claudism." Now a HERALD block in the declaudify detector; run
+`--list herald` before handing over any draft. (2) **Every technical term defined at first use, with
+one word per concept**: item/capability/allele, mass, refit, verified real samples, Zipf source,
+practitioner, "which trained networks" — the dual-audience rule from 2026-09-09 applied to *my own*
+vocabulary, not only the biology. Rule: after drafting, grep each noun of art for its first
+occurrence and check a definition precedes it. (3) **Never cite the SI for something the SI does not
+contain.** The text cited "the SI separates three cases" and "the proof is Poisson thinning" and
+neither existed; and every "(SI)" pointer must name a figure or text number. Rule: before writing
+"(SI)" grep si.md for the claim; if absent, write it (SI Text S4) or drop the sentence.
+
+## 2026-09-12 (round 2) — the Discussion must discuss; novelty is flagged where the result is shown
+GG on the "What is borrowed and what is new" inventory: "The discussion should discuss, not list."
+And on the Limits paragraph: stating small-scale limits "is usually done by undergraduate students";
+either run the experiment or discuss only problems too big for the paper. Rules: (a) never write a
+Discussion paragraph that is a list of prior-art citations or of caveats; each Discussion paragraph
+argues one point; (b) attribute novelty at the point of the result, with the figure panel, and name
+the prior finding it explains or extends in the same sentence ("an observation reported by others
+and left unexplained (60)"); (c) "much/some/most of X was known" is a hedge that gives novelty away
+without saying what is new; replace with the specific thing prior work lacks. Also: after
+`renumber_refs.py --apply`, the fig1a literal (refs. 22, 33) went stale (Shumailov became 23); the lesson from 2026-09-11 held.
+
+## 2026-09-12 — a prediction written into the Discussion must be run before it is printed
+The revised Discussion predicted that differential reproduction would turn recombination's speed
+advantage into a level advantage. Three GPU-hours later it did not (parity, 3/3 seeds). Rules:
+(a) when a Discussion sentence forecasts the outcome of an experiment we can run in under a day,
+run it in the same revision; (b) check adapter/cache directories for seed- and base-specificity
+before any HPC array (speciation shared one dir across seeds; the specialist cache would have loaded
+Qwen adapters into SmolLM2); (c) the local smoke gate for a new base (termination, base accuracy in
+(0.05, 0.95), sample generations) cost 4 minutes and is worth running every time.
+
+## 2026-09-13 — a figure must be readable without its caption
+GG on the manuscript figures after reading the student guide: "too unclear, cryptic"; figures should
+"give some clear information without the need to read the legend". The house rule in make_figs.py
+("no per-panel headline titles; interpretation lives in the captions") was the wrong rule for this
+audience and is reversed. Rules: (a) every data panel carries a one-line headline stating its
+finding plus a grey line naming the system and its size; (b) legend entries say in words what is
+plotted ("accuracy on the model's weakest task family", not "worst_family"); (c) where the set-up is
+not obvious, a schematic panel explains it inside the figure; (d) bar comparisons carry the test
+(paired over seeds, stars, key printed under the legend). Layout lesson: headlines longer than the
+panel run into the neighbour; wrap at ~40 characters per line for a half-width panel, ~70 for full
+width, and render before trusting.
+
+## 2026-09-13 — figure layout rules I should apply without being told
+GG had to ask three times for things a careful eye catches: headlines running past their panel,
+a schematic strip narrower than the data panels beneath it and not flush with their left edge,
+and a large blank band between a strip and the next row. Rules, now encoded in make_figs.py:
+(a) any panel placed by hand (schematics) is positioned from the neighbouring data axes' geometry:
+left edge = the data panels' frame, right edge = the last panel's frame, bottom = a fixed 0.75 in
+above the headline below, height from the content's designed aspect (never let equal-aspect centre
+a too-wide axes); (b) text wraps to its own panel width (`headline()` measures the axes); (c) after
+every regeneration, render at ≥ 90 dpi and check four things before reporting: nothing crosses a
+panel boundary, nothing overlaps, blank bands are no larger than the row gaps, and left edges of
+stacked panels line up. Report only after that check passes.
+
+## 2026-09-13 — prose: no staccato fragments
+- GG flagged "These results say X. They do not say where. The inheritance model does, in closed form."
+  as a claudism. Breaking a thought into short declaratives is rarely necessary; join them (colon,
+  "because", "and", "but"). Clarity comes from stating the concrete object, not from short sentences.
+- After any rewrite pass, scan for sentences of ≤7 words introduced by the edit and rejoin them.

tasks/prereg-llm-compose-v3.md

@@ -0,0 +1,345 @@
+# Pre-registration — `llm_compose` v3: does a composed capability survive inheritance?
+
+**Status:** draft for GG review, 2026-09-07. Supersedes `prereg-llm-society-v2.md` (calibrated,
+no-go at 0.5B) and `workorder-llm-society.md` (v1, run, negative). Nothing runs until §4's gates pass
+and GG signs off §12.
+
+**The question, in one sentence.** Every model-merging paper merges *once*; this asks what happens to a
+composed capability when the models that carry it keep reproducing, and whether the population-genetic
+closed forms predict the trajectory.
+
+---
+
+## 0. Why the design changed, and what carried over
+
+GG's objection to the v2→7B plan (2026-09-07): *"there is no structural reason 7B would succeed if
+0.5B failed. Most likely we are simply using the wrong LoRA specialisations… look in the literature
+and see what kind of test people use as paradigmatic for LoRA."* Correct on both points. Three things
+came out of the literature check:
+
+1. **The paradigmatic test is binary skill composition on a held-out, out-of-domain target.**
+   [LoRA Soups](https://aclanthology.org/2025.coling-industry.55.pdf) (COLING 2025): Llama-2-7B,
+   rank 8, math (MetaMathQA) × code (Code Alpaca) → GSM8k-Hard with program-aided evaluation;
+   also manual × instruction-following → closed-book QA. [LoraHub](https://arxiv.org/abs/2307.13269)
+   (COLM 2024): many Flan modules → held-out BBH. [MergeBench](https://arxiv.org/pdf/2505.10833):
+   math, code, multilingual, safety, IF. Nobody uses procedurally generated puzzle families.
+
+2. **Our v1/v2 families were disjoint but *non-composable*.** Sorting a list and counting letters
+   combine into nothing, so fitness had to be the *average of nine separate objectives* — a
+   **capacity** test (can one r = 16 adapter hold six skills?), which v2's calibration answered: no,
+   ≈ 0.85× per skill for two, worse for more. E8's genotype model has loci contributing to **one**
+   fitness function. Composable skills restore that and need only **two** parents per child, so the
+   capacity ceiling never binds. This, not scale, was the fault.
+
+3. **The operator was wrong in a way with a clean algebraic diagnosis.** peft
+   `combination_type="linear"` computes ΔW = (α₁B₁+α₂B₂)(α₁A₁+α₂A₂)ᵀ, which carries **cross terms**
+   B₁A₂ᵀ and B₂A₁ᵀ — one parent's output projection driven by the other's input projection.
+   `combination_type="cat"` gives α₁B₁A₁ᵀ + α₂B₂A₂ᵀ: each parent's rank-r subspace intact, rank 2r.
+   **That is E4's union operator in the natural algebra of the medium**, and the cross terms are the
+   mechanism of blending dilution. Their GSM-Hard numbers: CAT 21.11 > TIES 15.77 > DARE 14.78 >
+   MoE-routing 13.5 > LoRAHub 4.1 (below the 5.91 base). Routing below CAT is our own `llm_moe_hpc`
+   reading — selection is capped at the best parent, composition is not.
+
+**Carried over from v2's calibration (not wasted — it calibrated the channel v3 uses).** The
+self-consumption inheritance channel: a child distilled from its source's own answers loses 20–40%
+per generation to interference from confident off-expertise answers; gating on the source's own
+confidence (τ = 0.85) restores single-skill retention to **0.87–0.93**, and v3's lineages carry
+**one skill each**, which is exactly the regime the gate was measured in. `k = 300` prompts/skill,
+3 epochs, r = 16, confidence gate τ = 0.85.
+
+**Prior art to cite rather than re-demonstrate.** Offspring capability neither parent had is
+established: [Akiba et al.](https://www.nature.com/articles/s42256-024-00975-8) (Nature Mach. Intell.,
+Japanese × math) and LoRA Soups' *super-linear improvement* (base 5.91 → +code 8.04 → +math 14.18 →
+CAT 21.11; ≥ 5% of solved problems solved by neither parent). Population-based LLM evolution with
+crossover/mutation/selection exists ([Zhang et al. 2025](https://arxiv.org/abs/2503.01155), 40 models),
+as does iterated merging ([M2N2](https://arxiv.org/html/2508.16204v1), EvoGM). **None of them
+iterates the *reproduction* loop**: they optimise a merge, they do not ask what a merged capability
+does over generations. That gap is the experiment.
+
+**Positioning (GG, 2026-09-07): the manuscript does not need repositioning.** The paper's claim is
+that population-genetic *rules describe and predict* the phenomenon — closed forms, thresholds,
+conservation laws — not that recombination or collapse were discovered here. §5's H3 is that claim
+made falsifiable at the language-model tier. One genuine overlap to co-cite:
+[Model Collapse as Cultural Evolution](https://arxiv.org/html/2605.23054) runs ten generations of
+self-training and finds rare-variants-lost-first plus quality-filtering-as-remedy — our E1 and E2 in
+spirit — but under *iterated learning* (Bayesian convergence to the prior), with no equilibrium
+closed form, no threshold, no recombination, and no population structure.
+
+---
+
+## 1. The design
+
+**Two lineages, one measurement.** Two single-skill LoRA lineages on a shared frozen base — a **math**
+lineage and a **code** lineage. Each generation, each lineage reproduces by self-consumption (a fresh
+LoRA distilled from its own confidence-gated answers on fresh prompts). Each generation, the *current*
+two parents are merged and the composed model is evaluated on the held-out composed task.
+
+The composed model is a **measurement, not a lineage** — re-formed each generation from whatever the
+parents currently are. This isolates the question ("does composition survive parental drift?") from a
+confound ("does the composed model itself drift?"). An optional third arm makes the composed model a
+lineage too (§1.4).
+
+**1.1 Task.** Math (MetaMathQA subset) × code (Code Alpaca subset) → **GSM8k-Hard**, program-aided:
+the model emits Python, the code is **executed in a sandboxed subprocess**, and the return value is
+compared to the reference answer. Execution is the verifier — reality's "no" — and returns Layer 2 to
+the blueprint's original §3.6 specification.
+
+**1.2 Operators.** `cat` (union; the campaign operator) and `linear` (blending; the H6 control), both
+at merge weights fixed a priori to (0.5, 0.5) — *not* tuned, because a tuned blend would confound the
+operator contrast with search. LoRA Soups' learned-CAT is a stronger operator than ours; we do not
+need it, and using the untuned version makes the comparison to `linear` clean.
+
+**1.3 Grounding.** The arm knob, in the *training mix* this time (E2's immigration, not E11's
+selection channel): the dry arm's children see only the parent's own answers; the grounded arm mixes a
+fraction **g = 0.10** of fresh verified real examples (correct answers from the held-out pool of the
+lineage's own dataset) into each child's training data. This is the first LLM-tier test of *immigration*
+in this project; Fig. 1A's note that training-mix grounding at LLM scale is established elsewhere
+stands, but here it is the manipulated variable, not a claim of novelty.
+
+**1.4 Arms.**
+
+| arm | parent reproduction | grounding | operator | tests |
+|---|---|---|---|---|
+| `dry` | self-consumption | g = 0 | cat | H2, H3, H5 |
+| `grounded` | self-consumption | g = 0.10 | cat | H4 |
+| `dry_linear` | self-consumption | g = 0 | linear | H6 |
+| `dry_composed` *(optional)* | the merged child becomes the next parent of both lineages | g = 0 | cat | does composition survive in a self-consuming *composed* lineage |
+
+**1.5 Generations and seeds.** G = 6 generations, seeds 1–3, fixed now.
+
+**1.6 Measured each generation.** Per lineage: own-skill accuracy (math on MATH-500 subset; code on
+HumanEval-subset) → **q_t**, the retained skill. Between lineages: **ρ_t**, the correlation of their
+behaviour, measured as in `llm_epistasis` (agreement rate on a shared probe pool, and LoRA-delta
+cosine as a geometric companion). Composed: GSM-Hard accuracy, plus the **surplus** (composed − best
+parent on the composed task) and the **union-exceedance** (fraction of composed-solved problems that
+neither parent solves — LoRA Soups' super-linear signature).
+
+---
+
+## 2. What the framework predicts, quantitatively
+
+E4's closed form for two parents, U(K=2, ρ, q) = ρq + (1−ρ)(1−(1−q)²), gives the expected coverage of
+a capability held by either parent. Composition on a two-skill task is the conjunction rather than the
+union, so the corresponding prediction for a task needing *both* skills is the product form
+
+  **Ĉ_t = c₀ · q_t^math · q_t^code · (1 − ρ_t)/(1 − ρ₀)**
+
+with c₀ fixed by generation 0 (one free scale parameter, fit once, never refit). Two consequences the
+merging literature has no reason to expect:
+
+- **Composition decays faster than either parent.** Ĉ depends on the *product* of both retentions and
+  on decorrelation. If each parent retains 0.9 per generation, the composed capability retains 0.81
+  before any ρ effect. Super-linear gain becomes super-linear loss.
+- **ρ rises under dry self-training**, because both lineages drift toward the same attractor — the
+  base model's prior. Rising ρ removes the complementarity composition depends on, so the surplus
+  collapses even where q is still respectable. This is the mechanism, and it is measurable.
+
+---
+
+## 3. Hypotheses, thresholds, falsifiers
+
+Primary outcome: **composition surplus** S_t = (composed GSM-Hard accuracy) − (best single parent on
+GSM-Hard), per arm per seed per generation. Secondary: union-exceedance, q_t per lineage, ρ_t, and the
+predicted Ĉ_t.
+
+| | Prediction (source) | Threshold | Falsified if |
+|---|---|---|---|
+| **H1** *(gate, not a claim)* | Generation 0 reproduces the literature: CAT composes | S₀ ≥ +0.05 and union-exceedance ≥ 0.03 and CAT > linear by ≥ 0.03, in ≥ 2 of 3 seeds | any of these fails → the setup does not reproduce a published effect; **stop and fix before iterating** |
+| **H2** | Composition decays faster than its parents (§2) | S_t declines monotonically (Spearman ρ ≤ −0.7 vs t) and the composed capability's fractional loss by G exceeds each parent's own fractional loss, in ≥ 2 of 3 seeds | S_t flat or rising, or composed decays no faster than parents |
+| **H3** | **The closed form predicts the trajectory** (E4/§2) — the paper's central claim, made falsifiable | Ĉ_t (one parameter, fit at t=0) predicts observed composed accuracy with mean absolute error ≤ 0.05 across t = 1…G, and beats a two-parameter exponential-decay baseline on AIC | MAE > 0.10, or the atheoretical baseline wins → the closed form describes nothing the data did not already say |
+| **H4** | Grounding arrests it (E2 immigration) | S_G(grounded) − S_G(dry) ≥ +0.08, paired per seed, 3/3 seeds positive | grounded ≤ dry, or difference < 0.03 |
+| **H5** | Rising ρ is the mechanism | ρ_t rises monotonically in `dry` (Spearman ≥ +0.7) and is flat-or-lower in `grounded`; partial correlation of S_t with ρ_t controlling for q_t is negative | ρ flat in dry, or S_t–ρ_t partial correlation ≈ 0 → decay is pure retention loss, not lost complementarity (report either way; it is a mechanism result, not a claim of failure) |
+| **H6** | Blending conserves (E4) | `dry_linear` shows S₀ ≤ +0.02 and union-exceedance ≤ 0.01 at every generation — the conservation law, in the operator the literature already shows is worse | linear matches cat at generation 0 → the cross-term account of dilution is wrong |
+
+Analysis: per-seed paired contrasts, mean ± 95% CI over 3 seeds, sign counts reported. H3 is
+pre-registered as a *prediction with a fixed functional form and one free parameter*; the fit is at
+t = 0 only and is never refit.
+
+---
+
+## 4. Calibration gates (before any campaign)
+
+| Gate | What | Pass criterion | Cost |
+|---|---|---|---|
+| **C1 base** | Smallest Qwen2.5-Instruct (0.5B → 1.5B → 3B → 7B) whose *generation-0* CAT composition on GSM-Hard lands in **[0.25, 0.70]** | first size in band wins; if 7B exceeds 0.70 the task is saturated and GSM-Hard is swapped for its large-number variant | ≤ 2 h, escalating |
+| **C2 verifier** | Execution sandbox: determinism (same code → same verdict ×3), isolation (no filesystem/network), timeout, and agreement with reference answers on 100 gold solutions | 100% determinism, ≥ 0.98 agreement, no escape | 1 h, no GPU |
+| **C3 specialists** | Math and code LoRAs each beat base on their *own* skill by ≥ 0.15 and are ≤ 0.4 on the *other* skill (genuine specialists, decorrelated) | both | 1 h |
+| **C4 replication** | H1 at generation 0 (above) | as H1 | 1 h |
+| **C5 transmission** | Single-skill retention through one gated self-consumption step, per lineage, as v2's C2b | ≥ 0.85 per lineage | 1 h |
+
+C1's escalation is the honest form of the scale question: the base is chosen by *task discriminability*,
+not by hope. If 0.5B or 1.5B lands in band, the campaign is cheap.
+
+---
+
+## 4a. Calibration record
+
+**C2 execution verifier — PASS (2026-09-07).** 100/100 agreement with GSM-Hard's own reference
+`solution()` functions at 32 ms/item; deterministic across three runs; every hazard contained
+(infinite loop → timeout, allocation → memory, write outside the jail → PermissionError, socket →
+PermissionError, subprocess → PermissionError, syntax error, recursion). Two bugs the gate caught:
+a file write initially escaped (fixed with a `sys.addaudithook` guard) and CPU-rlimit kills were
+misreported as errors rather than timeouts. A third surfaced only on CX3 — temp paths are symlinked
+there, so the jail check needed `realpath`, not `abspath`.
+
+**C1 base — the escalation axis is instruction-tuning, not size.** Zero-shot GSM-Hard, program-aided:
+
+| base | GSM-Hard | executable | GSM8K |
+|---|---|---|---|
+| Qwen2.5-1.5B-**Instruct** | 0.500 | 0.950 | — |
+| Qwen2.5-3B-**Instruct** | 0.417 | 0.617 | — |
+| Qwen2.5-3B (base) | 0.633 | 0.950 | 0.750 |
+| **Qwen2.5-1.5B (base)** | **0.067** | 0.117 | 0.117 |
+
+A base that already has the skills makes the specialists vacuous — the v2 disease in new clothes, and
+it would have been *worse* at 7B, which is the quantitative form of GG's objection to the 7B plan.
+Qwen2.5-1.5B base is within noise of LoRA Soups' Llama-2-7B starting point (0.059), so the escalation
+runs along instruction-tuning rather than parameter count. **Amendment:** C1's band applies to the
+*base* (≤ 0.15) as well as to the composed model ([0.25, 0.70]).
+
+**C3 specialists + C4 replication (`results/llm_compose_gate`) — H1 FAILS, with a clean diagnosis.**
+
+| | composed (GSM-Hard) | executable | own skill |
+|---|---|---|---|
+| math parent (MetaMathQA) | 0.073 | 0.153 | GSM8K 0.620 |
+| code parent (CodeAlpaca) | **0.427** | 0.953 | MBPP 0.075 |
+| cat merge, 0.5/0.5 | 0.407 | 0.827 | — |
+
+Surplus **−0.020** (needs ≥ +0.05) → the pre-registered gate fails and no campaign is submitted on
+this configuration. But **union-exceedance is 0.073**: the merge solves 7.3% of items that *neither*
+parent solves, so composition is occurring and is being cancelled by a format cost (executability
+0.953 → 0.827 when the non-code parent is blended in).
+
+**The skill pair is unbalanced for this base, and the measurement says so precisely.** In the
+published setup math-only (0.142) beats code-only (0.080); here the ordering is *inverted* — code-only
+0.427, math-only 0.073 — because Qwen2.5's pretraining already carries the maths, so **code/format is
+the scarce skill and maths is not**. E8's premise is that each parent supplies something the child
+could not otherwise get; that holds for the code parent and fails for the math parent.
+
+**Diagnostic before any redesign (running):** a merge-weight sweep (0.5/0.5 → 0.1/0.9) under both
+operators, reusing the cached founders. It separates two possibilities that the single 0.5/0.5 point
+cannot: *(i)* the surplus is positive somewhere in weight space and 0.5/0.5 was a strawman — in which
+case the correct experiment is E10's directed version (breed offspring across weights, select on the
+verifier), which is this project's own operator and was fixed to 0.5/0.5 only to keep the operator
+contrast clean; or *(ii)* no weighting yields a positive surplus, in which case the pair is simply
+wrong for this base and the fix is a target whose *maths* the base cannot do (competition-level MATH
+program-aided), not a different merge.
+
+**Merge-weight sweep (cat, GSM8k-Hard, founders reused) — 0.5/0.5 was a strawman, and there is an
+interior optimum.**
+
+| math/code | composed | surplus | union-exceedance | executable |
+|---|---|---|---|---|
+| 0.5/0.5 | 0.407 | −0.020 | 0.073 | 0.827 |
+| **0.3/0.7** | **0.453** | **+0.027** | **0.093** | 0.967 |
+| 0.2/0.8 | 0.433 | +0.007 | 0.080 | 0.967 |
+| 0.1/0.9 | 0.433 | +0.007 | 0.040 | 0.960 |
+
+The surplus is positive over a range and peaks at an *interior* weight — E9's "optimal recombination
+rate is intermediate", in real weights — and the executability cost of blending disappears once the
+non-code parent is down-weighted (0.827 → 0.967, above even the code parent's 0.953). It is also
+LoRA Soups' own result that *learned* CAT beats *static* CAT, arrived at independently. **Amendment
+(adopted):** merge weights are chosen per generation on a **disjoint validation split** of the
+composed target and reported on the test split — E10's directed recombination, which was fixed at
+0.5/0.5 in §1.2 only to keep the operator contrast clean. The `linear` control arm keeps the same
+treatment, so the operator contrast survives. `_split_pool` makes val/test disjointness structural
+rather than a property of seeds, which matters at MATH-500's pool size.
+
+**Still short of the gate: peak surplus +0.027 against a +0.05 threshold, and n = 150 gives
+SE ≈ 0.04.** The bar is high because the *best parent* is at 0.427, where LoRA Soups' was 0.142.
+So the second gate configuration (`configs/llm/compose_gate_math500.yaml`) moves the target to
+MATH-500 rather than moving the threshold.
+
+**Full gen-0 sweep, both operators (GSM8k-Hard, 150 items, best parent 0.427):**
+
+| operator | 0.5/0.5 | 0.3/0.7 | 0.2/0.8 | 0.1/0.9 | range |
+|---|---|---|---|---|---|
+| cat | 0.407 | **0.453** | 0.433 | 0.433 | 0.046 |
+| linear | 0.333 | 0.467 | **0.507** | 0.413 | 0.174 |
+
+**H6 as pre-registered is FALSIFIED, and what replaces it is more interesting.** The prediction was
+that blending "conserves" — no composition to lose. In fact linear blending produces the *largest*
+union-exceedance (0.133 vs cat's 0.093) and the highest composed accuracy of any configuration
+(0.507, surplus +0.080), *provided the weight is chosen*. What distinguishes the operators is
+**variance, not mean**: concatenation is nearly flat in the blend ratio (range 0.046) and never
+catastrophic, while blending swings by 0.174 — worst of all at equal weights (0.333, executability
+0.680, the cross-terms wrecking the code parent's format), best of all at 0.2/0.8. That is E9's
+structure in real weights: blind recombination → outbreeding depression; directed recombination →
+gain; the union operator is the conservative strategy. Revised H6 (recorded before the campaign):
+*the operator ordering is weight-dependent at generation 0; does it stay so across generations, or
+does one operator degrade faster?* — measured by the `dry_cat` arm.
+
+**Second target, MATH-500 (`results/llm_compose_gate_math500`, cat + selected weights): composed
+0.200, best parent 0.158, surplus +0.042, union-exceedance 0.100.** The balanced-pair prediction
+holds — a much weaker best parent (0.158 vs 0.427) leaves more headroom, and cat's surplus rises from
++0.027 to +0.042. The composition effect therefore reproduces on **two independent targets**, at the
+cost of one extra evaluation pass since the founders are shared.
+
+**Gate verdict: PASS on the amended configuration** (directed weight selection; GSM8k-Hard primary,
+MATH-500 as the generality check). Campaign launched 2026-09-07: seed 1 local, seeds 2-3 as CX3 array
+`4000472` (6 elements, 3 arms x 2 seeds).
+
+## 5. Cost
+
+Per generation-arm: 2 lineages × (2700 gated inheritance answers + train ~1000 × 3 epochs) +
+composed eval on 300 GSM-Hard items with execution. At **1.5B**: ≈ 25 min. At **7B**: ≈ 75 min.
+
+| base (from C1) | arms | seeds | G | total |
+|---|---|---|---|---|
+| 1.5B | 3 | 3 | 6 | **≈ 12 GPU-h** (local, overnight) |
+| 3B | 3 | 3 | 6 | ≈ 25 GPU-h (local or 4 CX3 jobs) |
+| 7B | 3 | 3 | 6 | ≈ 40 L40S-h (9 array elements, 6 h each) |
+
+The optional `dry_composed` arm adds a third. Every figure in §3 is drawn from one parquet;
+checkpoint/resume carries over from `society_v2.py`.
+
+---
+
+## 6. Anticipated failure modes
+
+- **Gen-0 does not compose (C4 fails).** Most likely cause is the base being too weak for
+  program-aided math at all. C1's escalation should prevent it; if it survives C1, stop — the
+  experiment has no signal to measure the decay of.
+- **Parents don't drift.** If gated self-consumption is *too* good, q stays ≈ 1 and there is nothing
+  to observe. Mitigation: the gate is a knob (v2 measured τ = 0.5 → 0.87 and ungated → 0.81); if q_6 >
+  0.9 in the dry arm at τ = 0.85, drop to ungated, which is the *more* faithful self-consumption
+  channel anyway. Pre-declared, not a post-hoc rescue.
+- **ρ unmeasurable.** Math and code lineages answer disjoint prompt types, so behavioural agreement may
+  be uninformative. Fallback: ρ from LoRA-delta cosine (already implemented in `epistasis.py`) and
+  from agreement on the *shared* GSM-Hard prompts.
+- **Execution verifier flakiness** — timeouts counted as failures, reported as a rate.
+- **GSM-Hard contamination** in an Instruct base: report the base's zero-shot GSM-Hard number; if it
+  is implausibly high, switch to the perturbed-number variant.
+
+---
+
+## 7. Engineering checklist
+
+- [ ] `src/llm/execute.py` — sandboxed subprocess execution verifier (timeout, no network, no
+      filesystem writes, deterministic), + tests.
+- [ ] `src/llm/datasets.py` — MetaMathQA / Code Alpaca / GSM8k-Hard loaders, fixed subsets, cached.
+- [ ] `src/llm/compose.py` — `kind: llm_compose`; two lineages, per-generation merge-and-measure,
+      `cat`/`linear`, grounding fraction, confidence gate, checkpoint/resume (port from `society_v2.py`).
+- [ ] q_t / ρ_t instrumentation (reuse `epistasis.generate_with_confidence`, `delta_geometry`).
+- [ ] `configs/llm/compose_calib_{c1..c5}.yaml`, `compose_s{1,2,3}.yaml`, `hpc/llm_compose.pbs`.
+- [ ] `figures/plot_llm_compose.py` (4 panels: S_t per arm; q_t per lineage; ρ_t; observed vs Ĉ_t) and
+      `figures/stats_llm_compose.py` (H2–H6) — **written before unblinding**, as in v2.
+- [ ] Smoke: 1.5B, G = 2, all arms, 50 eval items.
+
+## 8. Outcome → manuscript
+
+| Outcome | What changes |
+|---|---|
+| H1–H4 pass, H3 within tolerance | Fig. 1A's "open — the stated gap" cell is filled by a *different and better* experiment than the one specified: the closed form predicting a real LLM capability trajectory over generations. New figure; the recombination and grounding sections each gain their language-model rung. |
+| H2 + H4 pass, H3 fails | The signs transfer, the quantitative law does not — report as such; the paper's predictive claim stays anchored on the biological model, and the LLM tier is confirmatory (which is what Fig. 1A already says of the other rows). |
+| H1 fails at every base size | No campaign. The SI records the 0.5B transmission ceiling (v2) plus the failure to reproduce a published composition effect in our harness — an infrastructure result, honestly labelled. |
+
+## 9. Decisions for GG
+
+1. **Base escalation cap** — stop at 3B (cheap, local, likely enough) or allow 7B if C1 demands it?
+2. **Optional `dry_composed` arm** (+33% cost): does composition survive when the *merged* model is
+   itself the reproducing lineage? It is the closest thing to the original society question.
+3. **Skill pair** — math × code (best-anchored to the literature) or a second pair alongside
+   (manual × instruction-following) for generality at double the cost?
+4. Whether the v2 no-go and this redesign are worth a short **SI subsection on negative results**, or
+   stay in the repository record only.

tasks/prereg-llm-society-v2.md

@@ -0,0 +1,563 @@
+# Pre-registration — `llm_society` v2: the composed society at LLM scale
+
+**Status (2026-09-07, 21:30): calibrated; NO-GO at 0.5B (GG, §12a); 7B plan in §13 awaiting scope.**
+Supersedes the design in `workorder-llm-society.md` (v1). Calibration record and every amendment are
+in §4a; the campaign was not submitted. Read §1 (what v1 got wrong), §4a (what calibration found),
+§12a (the decision), §13 (what next).
+
+**Why a v2.** The v1 campaign (3 CX3 seeds landed 2026-09-07, `results/llm_society_campaign/`;
+seed 1 still running locally) did not reproduce E11: `no_grounding` degraded (0.575 → 0.436, worst
+arm in every seed), but `full` also declined (→ 0.506) and `no_sex` was flattest (0.558). The
+conformity−truth gap did not separate the arms. Read against the framework, v1 had three
+*structural* faults that the theory would have predicted, plus one power fault. All four are
+diagnosed in §1 and designed out in §3. The point of this document is to make the remaining
+predictions explicit *before* spending the compute, so the campaign can fail informatively.
+
+---
+
+## 0. The question and the claims it tests
+
+Does a finite population of LLM agents under the four composed operators — grounded evaluation,
+directed recombination, diversity-preserving selection, lossy inheritance — climb to capability that
+no founder had and hold it, while each ablation fails in its own way? This is E11 at the language-model
+tier: the paper's "open — the stated gap" cell (Fig. 1A).
+
+Claims exercised, and the analytic experiment each rests on:
+
+| Claim | Analytic source | LLM prediction (§5) |
+|---|---|---|
+| Recombination assembles a genotype no parent had (vertical) | E8 (Fisher–Muller, unbounded parents) | H1 |
+| Blending conserves the single-parent level; only union realises the gain | E4 (conservation law), `llm_moe` | H2 |
+| Ungrounded selection → self-consumption → confident, unfit consensus | E11 (conformity mechanism) | H3 |
+| Without recombination, capability is capped at the best founder | E8 control (ρ=1) + no mutation operator here | H4 |
+| Greedy selection collapses diversity faster; QD holds it | E5, E11 | H5 |
+| A capability survives inheritance only if observed often enough | E2 (per-item floor 1−e^{−mp}) | H6 |
+
+**Explicit non-goal.** Grounding here is E11's *selection-channel* grounding
+(`g·fitness + (1−g)·conformity`), not E2's *immigration into the training mix*. No verified answer
+ever enters any child's training data, in any arm. Fig. 1A already states that LLM-scale training-mix
+grounding is established in prior work and not re-run; this campaign does not change that. A
+negative here is evidence against the *selection* mechanism only.
+
+---
+
+## 1. What v1 got wrong, read through the framework
+
+| # | Fault | What the theory says | Evidence in v1 | Fix (§3) |
+|---|---|---|---|---|
+| F1 | **Near-clone founders.** 8 agents over 3 families → 3 lists-, 3 strings-, 2 arith-specialists differing only by task draws. | E8 control: recombining ρ=1 parents buys **nothing** (flat at 6 for any K). Pigeonhole: 4 parents from 3 families always contains a same-family pair. | `no_sex` ≥ `full`: merging near-clones is pure perturbation cost. | L = 12 disjoint families, **one founder per family**, ρ = 0 by construction; verified at gen 0. |
+| F2 | **Combinatorial space too small.** 3 skills → 2³ = 8 competence states; founders occupy 3 of them. | E8/E11 use L = 12 (4096 genotypes). The vertical claim needs room for a child to be *new*. | Best possible gain over a founder was tiny. | L = 12 → the best founder holds 1/12 of the space. |
+| F3 | **Blending operator in the dilution regime.** Sex = 2-parent *linear* LoRA merge at 0.5B. | E4 conservation law; `llm_moe` 0.5B: soup dilutes lists 0.43 → 0.26. Headroom law: dilution wherever there is room to lose. | `full` declined while `no_sex` held. | Reproduction by **union-preserving recombination** (confidence-routed union of parents' answers → distil). Linear merge kept as an explicit control arm (`sex_linear`) — H2. |
+| F4 | **Truncation before breeding.** Top-4 of 8 selected as parents; children bred only from them. | E11 selects on *survival over the pooled parents + offspring*, never on breeding eligibility. Truncating parents discards half the alleles at gen 1 with no mutation operator to restore them (E6: loss is permanent). | Half the families were unreachable after gen 1. | Survival selection over the pool (§3.4). Elitism becomes emergent, as in E11. |
+| F5 | **Underpowered evaluation.** `n_test` = 40 → SE 0.079 per measurement. | — | Every contrast except vs `no_grounding` sat inside one SE. | `n_test` = 240 (20/family) → SE 0.032 overall. |
+| F6 | **Weak grounding contrast.** g = 0.5 vs E11's 0.85; G = 10 vs 80. | The conformity gap in E11 needs the population to converge; g = 0.5 leaves conformity with half the vote even in `full`. | Gap flat in all arms. | g = 0.85; G = 12 (§7 explains why 12 suffices here). |
+| F7 | **Transmission floor never measured.** `n_inherit` = 600 over 3 families chosen by feel. | E2: an item survives only if it is *observed* enough in the inheritance sample — the per-item floor. Pilot v1 measured a ~25%/gen "distillation tax" and fixed it by doubling data, without asking where the floor was. | — | Calibration C2 measures the retention curve and sets `n_inherit` from it. |
+
+---
+
+## 2. Theoretical predictions → design constraints
+
+Each constraint below is derived, not chosen.
+
+**2.1 Decorrelation (E8, E4).** Union coverage of K parents is U = ρq + (1−ρ)(1−(1−q)^K); the
+gain over a single parent is proportional to (1−ρ). Founders must therefore be as decorrelated as the
+task space allows: one family each, no shared training items, and the gen-0 behavioural-distance
+matrix must show no pair below 0.5 disagreement (gate C1c).
+
+**2.2 Combinatorial headroom (E8).** With one family per founder, q = 1/L. The doubling bound for
+2-parent recombination gives ≥ ⌈log₂ L⌉ = 4 generations to *reach* full coverage under lossless
+inheritance; with per-generation retention r per family the plateau is set by r, not L. So L = 12
+gives headroom; G must exceed 4 by enough to see the plateau: G = 12.
+
+**2.3 Operator (E4, `llm_moe`, `llm_directed`).** At 0.5B on unsaturated families the framework
+predicts linear blending dilutes and union preserves. The society's reproduction operator must be
+union-preserving or the experiment re-measures a known result. The union is implemented in the
+*inheritance data*, not in weight space: for each inheritance prompt the child learns the answer of
+whichever parent is more confident (mean token log-probability of its own answer). This is E4's
+`max` operator applied per item, it is verifier-free (legal in the `no_grounding` arm), and it is
+directed sex in E10's sense — mate choice by complementarity plus per-item selection. Gate C4 checks
+that confidence tracks competence (the routing precondition); gate C3 checks the union child beats the
+linear child on a single 2-founder cross before any campaign money is spent.
+
+**2.4 Selection acts on survival (E11).** E11 pools N parents with n_off offspring and keeps the top
+N by `score + novelty·λ`. Reproducing that exactly gives: elitism for free (a strong parent survives
+by out-scoring its children), no gen-1 truncation, and a directly comparable selection intensity
+(keep 12 of 24 = top ½; E11 keeps 60 of 180 = top ⅓ — pre-noted as a difference).
+
+**2.5 Conformity must be decoupled from truth for H3 to be testable.** E11 initialises random
+genotypes, so its consensus is uninformative at gen 0. In the LLM, consensus is the modal answer
+over agents. With one expert per family, 11 of 12 agents answer any given family's prompt at roughly
+base level, so the modal answer ≈ the base model's answer, and conformity rewards *being base-like*.
+Prediction: consensus accuracy at gen 0 ≈ base overall (gate C5 measures it; it must be < 0.35, i.e.
+well below the best founder's own-family accuracy, otherwise conformity is a truth proxy and the
+`no_grounding` arm cannot fail by the predicted mechanism — see §6 F-alt).
+
+**2.6 The inheritance floor (E2).** A family survives distillation only if the child sees enough
+*correct* examples of it. Retention as a function of examples-per-family, r(k), is measurable
+(C2). Set k* = min k with r ≥ 0.85, `n_inherit` = L·k*. Consequence for a *diluted* skill at accuracy
+a: effective correct examples = a·k*, so skills are lost at the distillation step, not the merge step
+(H6, a mechanism prediction that distinguishes this account from "merging destroyed it").
+
+**2.7 Headroom, not size (the `llm_moe_hard` lesson).** Families must be unsaturated for the
+specialist (≤ 0.9) and non-trivial for the base (≥ 0.05, ≤ 0.4) at 0.5B. C1 rejects families outside
+that band. The three existing easy families straddle it (base lists 0.15 / strings 0.15 / arith 0.53;
+specialists 0.43 / 1.00 / 0.91) — strings saturates, arith's base is high. Both are candidates for
+replacement, decided by C1 not by preference.
+
+**2.8 No cross-family conflict (E12, `llm_epistasis`, `llm_speciation`).** Recombination helps on
+additive landscapes and hurts under functional conflict. Families must not share a prompt shape with
+different answer conventions (e.g. two families that both say "sort" meaning different orders). Gate
+C1b: pairwise confidence-weighted functional conflict between founders (the `llm_epistasis` measure)
+below the grid's no-conflict axis level.
+
+---
+
+## 3. Design
+
+**3.1 Base and agents.** Qwen2.5-0.5B-Instruct (frozen). Agent = LoRA r=16 α=32 on attention + MLP
+projections. N = 12.
+
+**3.2 Task space.** L = 12 disjoint procedurally generated families, exact-match verifier, chosen from
+the candidate list in §4 C1 by the calibration band. Fixed pools per family: test 20 (reporting only),
+val 10 (grounded selection signal), conformity 10 fresh per generation (prompts only). Inheritance pool:
+k* per family, fresh per generation.
+
+**3.3 Founders.** Agent i is a specialist on family i, trained on 600 × 3 epochs of that family only,
+seed `master·100 + i`. Gate C1c: gen-0 pairwise behavioural distance ≥ 0.5 for every pair.
+
+**3.4 One generation.**
+
+1. *Produce & score.* Every agent in the current population answers test, val, and conformity pools.
+   Consensus = modal normalised answer per conformity prompt. `fitness` = val accuracy; `conformity`
+   = agreement rate with consensus; `novelty` = mean behavioural distance to the rest.
+   `score = g·fitness + (1−g)·conformity`, g = 0.85 (0 in `no_grounding`).
+2. *Mate.* Form n_off = N pairs by complementarity (greedy maximum behavioural distance, each agent
+   used at most twice), restricted to the current population.
+3. *Recombine (sex ON).* For each pair, each parent answers the inheritance pool once (with mean
+   token log-prob). Union child data: per prompt, the answer of the more confident parent. In
+   `sex_linear`: 6 Dirichlet-weighted linear LoRA merges screened on the arm's own signal (val
+   fitness, or conformity in ungrounded arms), winner answers the pool. Sex OFF: the child's data is a
+   single parent's answers (parent chosen by score-proportional sampling).
+4. *Inherit.* Each child is a fresh LoRA trained from the base on its data, `epochs` from C2.
+5. *Survive.* Pool = N parents + N children. Keep the top N by `score + λ·novelty` (λ = 0.3; λ = 0 in
+   `no_diversity`). Scores for children computed after training, on the same pools.
+
+**3.5 Arms (5).**
+
+| arm | g | recombination | survival |
+|---|---|---|---|
+| `full` | 0.85 | union (confidence-routed) | QD (λ=0.3) |
+| `no_grounding` | 0 | union; routing unchanged (verifier-free) | QD on conformity |
+| `no_sex` | 0.85 | none (single-parent redistillation) | QD |
+| `no_diversity` | 0.85 | union | greedy (λ=0) |
+| `sex_linear` | 0.85 | 2-parent linear merge, 6 screened candidates | QD |
+
+`sex_linear` is the H2 control and is the first arm to drop if budget forces it (§9).
+
+**3.6 Generations and seeds.** G = 12. Seeds 1–4 (fixed before running; not re-drawn). Reporting
+uses the verifier in every arm; the verifier enters the *loop* only where g > 0.
+
+---
+
+## 4. Calibration phase (local 16 GB GPU, ≈ 3–4 h, before any submission)
+
+Each gate has a pre-set pass criterion. Fail → stop, fix, re-run the gate. No campaign until all pass.
+
+| Gate | What | Pass criterion | Cost |
+|---|---|---|---|
+| **C1a** family band | Base and specialist (600×3) accuracy on each of ~15 candidate families (100 test items each) | Keep families with base ∈ [0.05, 0.40] and specialist ∈ [0.60, 0.90]; need ≥ 12 | ~60 min |
+| **C1b** no conflict | Pairwise confidence-weighted functional conflict between the 12 founders (from `llm_epistasis`) | Every pair below the `compat` axis level of the epistasis grid | ~15 min |
+| **C1c** decorrelation | Gen-0 behavioural-distance matrix on 120 mixed prompts | min pairwise disagreement ≥ 0.5 | (with C1b) |
+| **C2** transmission floor | Distil a child from a founder's *own* answers with k ∈ {25, 50, 100, 200} examples of its family (rest of the pool mixed), 2 and 3 epochs; measure retained own-family accuracy ÷ founder accuracy | Choose k* = min k with retention ≥ 0.85 at the chosen epochs; if no k ≤ 200 passes, the design is infeasible at 0.5B — stop | ~45 min |
+| **C3** operator | One cross (two founders): union-distil child vs best-of-6 linear-merge-distil child; both families' accuracy | Union child ≥ 0.85 × each parent on that parent's family; union ≥ linear on the *minimum* of the two. If linear ≥ union, H2 is already falsified — record it and reconsider the operator before the campaign | ~20 min |
+| **C4** routing precondition | For each founder: mean log-prob on own-family answers vs off-family; AUC | AUC ≥ 0.7 for ≥ 10 of 12 founders | (with C2) |
+| **C5** consensus anchoring | Consensus accuracy over the 12 founders at gen 0, 120 prompts | < 0.35 (conformity is not a truth proxy) | ~5 min |
+
+C2 also fixes the cost model (§9) — `n_inherit` = 12·k*.
+
+---
+
+## 5. Pre-registered hypotheses, thresholds, falsifiers
+
+Primary outcome metric: **best-agent overall test accuracy** at generation G (deployed capability,
+elite included), reported with the **best newborn** (child trained that generation) alongside, so a
+"climb" carried by a surviving founder is visible as such. Secondary: per-family accuracy of the best
+agent (the competence genotype), behavioural diversity, consensus accuracy, conformity−truth gap.
+Reference level **B₀** = best founder overall at gen 0 (≈ (0.7 + 11·base)/12 ≈ 0.24 if base ≈ 0.2;
+measured, not assumed).
+
+| | Prediction (from) | Quantitative threshold | Falsified if |
+|---|---|---|---|
+| **H1** vertical climb | E8: union recombination of decorrelated one-family founders assembles multi-family agents; sigmoidal, most of the climb in gens 1–5, plateau set by r | `full` best-agent at G ≥ B₀ + 0.20 and best newborn at G ≥ B₀ + 0.15; best agent competent (≥ 0.6) on ≥ 6 of 12 families; in ≥ 3 of 4 seeds | best-agent gain < 0.10 in ≥ 2 seeds |
+| **H2** operator | E4 conservation law in the dilution regime | `full` − `sex_linear` ≥ 0.10 at G (paired, per seed); `sex_linear` best agent competent on ≤ 3 families | `sex_linear` ≥ `full` in ≥ 2 seeds. *Pre-stated regime caveat:* this ordering is predicted to **invert** at 7B on easy tasks (headroom law); a 7B follow-up would test that, not this. |
+| **H3** self-consumption | E11 + §2.5: conformity rewards base-likeness; ungrounded selection regresses the population to base and homogenises it | `no_grounding` best agent at G ≤ B₀ + 0.05; conformity−truth gap (`no_grounding` − `full`) ≥ 0.30 at G; consensus accuracy in `no_grounding` non-increasing | `no_grounding` ≥ `full` − 0.05 on best agent, **or** gap difference < 0.10 |
+| **H4** sex necessity | E8 ρ=1 control + no skill-acquisition operator without recombination | `no_sex` best agent at G ≤ B₀ + 0.05 in every seed (a *ceiling*, stronger than E11's ~1-point effect) | `no_sex` gains ≥ 0.10 over B₀ in any seed → an unmodelled acquisition route exists (base competence amplified by self-distillation); report it |
+| **H5** diversity | E5/E11: greedy converges earliest; QD holds | AUC of behavioural diversity `full` > `no_diversity` in ≥ 3 seeds; `no_diversity` diversity < 0.1 by gen ≤ 6. **Low power on best fitness pre-declared** (E11: 0.78 vs 0.74) | no ordering in diversity AUC |
+| **H6** where skills die | E2 floor: loss occurs at distillation when correct examples/family < k*·a | For families lost between t and t+1 in `full`, the *source* (union answer set) accuracy on that family at t is ≥ 0.6 in ≤ 20% of cases — i.e. skills that were competently supplied are retained; skills die because they arrived diluted | ≥ 40% of lost families were supplied at ≥ 0.6 → the distillation channel itself is lossy beyond the calibrated floor; revisit C2 |
+
+Analysis is per-seed paired contrasts (4 seeds), reported as mean ± 95% CI and sign count. No metric
+introduced after unblinding is called a result. All rows above are also plotted whether or not they
+pass.
+
+---
+
+## 6. Anticipated failure modes and how each is handled
+
+- **F-alt (conformity anchored to truth).** If C5 shows consensus accuracy ≥ 0.35, the `no_grounding`
+  arm may fail by drift rather than by confident-wrong consensus (v1 workorder falsifier 3). Then H3's
+  gap threshold is not testable; run anyway, report the observed signature, and say so.
+- **Terminal degeneration.** A source that emits < 8 usable answers: copy the parent unchanged (existing
+  sentinel). Count and report occurrences per arm.
+- **Family extinction is permanent (E6).** No mutation operator reintroduces a lost family. Pre-state:
+  the number of families alive in the population is itself a reported curve; `full` is predicted to
+  hold ≥ 10 of 12 to G, `no_diversity` fewer.
+- **Router failure (C4 fails).** Fall back to self-consistency routing (two samples, prefer the parent
+  whose answers agree); re-run C4. If still failing, the union operator has no verifier-free
+  implementation at this scale — record and consider 7B.
+- **Elite lock-in.** A founder that survives to G on score alone makes "best agent" flat. Best newborn
+  is co-primary for exactly this reason.
+- **Screening noise in `sex_linear`.** 6 candidates on 120 val items (SE 0.046) — adequate for choosing
+  among merges that differ by ≥ 0.1, which is the dilution scale.
+- **Queue / wall-time loss.** Per-generation checkpoint (rows flushed to parquet; adapters on disk)
+  and `--resume`; each PBS array element = one (seed, arm), ≤ 6 h.
+- **Environment drift.** Never `uv sync` on a machine with a running job (`tasks/lessons.md`).
+
+---
+
+## 7. Power
+
+SE of an overall accuracy at 240 items, p ≈ 0.5: 0.032. Per-family at 20 items: 0.11 (per-family
+readouts are descriptive only). Predicted effects: H1 ≥ 0.20, H2 ≥ 0.10, H3 ≥ 0.20 on best agent and
+≥ 0.30 on the gap, H4 a ceiling — all ≥ 3 SE. H5 on best fitness is predicted small and is not powered;
+its diversity readout is. Four seeds give a sign test at p = 1/16 one-sided for a 4/4 outcome; the
+primary analysis is the paired mean and CI, the sign count is descriptive.
+
+Why G = 12 suffices: the climb needs ⌈log₂ 12⌉ = 4 doublings; the E11 `no_grounding` crash occurred
+by generation 6 of 80; diversity collapse in E11's ablations by 10–15. Twelve generations covers
+every predicted transition with margin to see the plateau.
+
+---
+
+## 8. Analysis and figure (fixed now)
+
+Figure, E11 layout plus one panel: **(A)** best-agent and best-newborn overall accuracy per arm over
+generations, B₀ dashed; **(B)** behavioural diversity; **(C)** conformity−truth gap; **(D)** competence
+heat-map — families × generations for the `full` best agent, with `sex_linear` beside it. Mean ± 95% CI
+over seeds. Script `figures/plot_llm_society.py`, reading only the committed bundles.
+
+Statistics: per-seed paired contrasts at G for H1–H4; AUC contrast for H5; the supplied-vs-retained
+tabulation for H6. `figures/stats_llm_society.py`.
+
+---
+
+## 9. Compute and schedule
+
+Measured anchor: v1 seed (N=8, G=10, 4 arms, n_inherit=600, 3 epochs) ≈ 3.3 L40S-hours.
+
+Per generation-arm at v2 defaults (N=12, 24-agent pool eval on 480 prompts; 12 parents answering the
+inheritance pool once each with log-probs; 12 children trained at 12·k* × epochs):
+
+| k* from C2 | training | inheritance answers | eval | per gen-arm | per (seed, arm), G=12 | campaign, 4 seeds × 5 arms |
+|---|---|---|---|---|---|---|
+| 50 | ~6 min | ~3 min | ~4 min | ~13 min | ~2.6 h | **~52 L40S-h** |
+| 100 | ~12 min | ~5 min | ~4 min | ~21 min | ~4.2 h | **~84 L40S-h** |
+
+Submitted as a 20-element PBS array (seed × arm), 6 h wall-time each, checkpointed. Dropping
+`sex_linear` saves 20%. The local GPU runs one seed's `full` + `no_grounding` in parallel as a hedge.
+
+**7B.** ≈ 4× per operation → ~200–340 L40S-h for the full grid: not a first shot. Pre-registered role
+for 7B: a **1–2 seed confirmation of `full` vs `no_grounding` and of the H2 inversion**, run only if
+0.5B passes H1 and H3. On the *easy* families 7B is predicted to be in the composition regime, so
+`sex_linear` should catch up with `full` there — a positive prediction of the headroom law, not a
+replication.
+
+**Schedule.** Calibration day 1 (local). GG gate on C-results. Engineering (§10) days 1–2. Campaign
+submission day 2–3; wall-clock ≈ 1 day if the queue cooperates. Analysis + figure day 4.
+
+---
+
+## 10. Engineering checklist (before submission)
+
+- [ ] `tasks.py`: ~15 candidate families with verifier formats (int / int-list / lowercase word /
+      uppercase word); `FAMILIES` becomes config-driven.
+- [ ] `society.py`: survival-over-pool selection replacing parent truncation; complementarity pairing
+      over the whole population; confidence-routed union inheritance (needs `generate()` to return
+      mean token log-prob); `sex_linear` arm; per-generation parquet flush + `--resume`; log the
+      *source* per-family accuracy before distillation (for H6) and families-alive per generation.
+- [ ] Tests for the pure pieces (union routing, pooled survival, pairing constraint) — extend the 155.
+- [ ] `configs/llm/society_v2_calib_*.yaml`, `society_v2_s{1..4}.yaml`; `hpc/llm_society_v2.pbs` array.
+- [ ] `figures/plot_llm_society.py`, `figures/stats_llm_society.py` written **before** unblinding,
+      against the smoke bundle.
+- [ ] Smoke: N=4, L=4, G=2, all 5 arms, exit 0, figure renders.
+
+---
+
+## 11. Outcome → manuscript
+
+| Outcome | What changes in the paper |
+|---|---|
+| H1 ∧ H3 pass (H2, H4, H5 whatever they are) | Fig. 1A cell "open — the stated gap" → filled; new figure (§8) enters as the LLM tier of the society; Table S2 row; the Discussion's prescriptive claim gains its LLM instantiation. |
+| H1 fails, H2 passes | The vertical claim does not transfer at 0.5B but the operator law does: report as a bounded negative in SI with the H6 diagnosis; Fig. 1A cell becomes "tested at 0.5B: operator law holds, climb does not"; 7B confirmation becomes the open item. |
+| H3 fails (no self-consumption signature) with C5 passed | The selection-channel grounding mechanism does not transfer; state it, keep the E11 result as biological-model-only; nothing prescriptive at LLM scale. |
+| C-gates fail | No campaign. The gate result itself goes in the SI as the reason the tier was not run. |
+
+---
+
+## 12. Decisions (GG, 2026-09-07)
+
+1. **Scale:** 0.5B full grid; 7B only as the gated confirmation of §9. *Decided.*
+2. **`sex_linear` arm:** **dropped from the first campaign** — founders are cached and shared, so it
+   can be appended later at ~20% of the grid cost, and the operator law is already established by
+   `llm_moe` (Fig. 3B). The code path stays (`arm_settings("sex_linear")`); H2 is therefore
+   *deferred*, not tested, in this campaign. Four arms × four seeds = 16 array elements.
+3. **Family candidates:** no vetoes; calibration C1 decides membership. Seventeen candidates are
+   implemented in `src/llm/families.py` (the three originals + fourteen new). Word-order reversal and
+   run-length encoding were dropped at implementation because the verifier cannot score multi-word or
+   alphanumeric answers; `sortletters`, `caesar` and `charfreq` use random pseudo-words so a 600-item
+   training set cannot cover the test space.
+4. **Go/no-go gate after calibration** — GG reviews the C-table before anything is submitted.
+   *Standing.*
+
+## 4a. Calibration record and amendments
+
+**Stage A, pass 1 (2026-09-07, `results/llm_society_v2_calib_a`, founders 600×3).** Only **6 of 17**
+families in band: setops, numtheory, mixedtoken, vectors, digits, alphabet. Out of band: lists and
+binary under-trained (specialist 0.44); strings 0.91, roman 0.97, prime 1.00 above the specialist
+ceiling; arith base 0.53 (the base already knows it — violates the sole-expert premise); liststats and
+charfreq weak (0.38; charfreq also fails the routing AUC, 0.44); sortletters 0.20, caesar 0.01,
+progression 0.10 unlearnable at this budget. C1b max pairwise conflict 0.348 (just under the gate),
+every top pair involving a *failed* specialist answering confidently wrong; C1c min distance 0.65.
+Smoke passed on all four arms (figure + stats script exercised).
+
+**Amendments before pass 2 — recorded here because they change what §4 promised:**
+
+1. **Specialist upper bound 0.90 → 1.00.** The bound encoded the `llm_moe_hard` headroom lesson,
+   which concerns *fusion composing to a ceiling* so that soup matches routing. The society uses
+   union inheritance, and the quantities under test are transmission and assembly, for which a
+   founder at 0.97 is not a problem. The **base** bound (≤ 0.40) is kept strict: it protects the
+   sole-expert premise and C5's consensus decoupling. Consequence: strings, roman, prime become
+   eligible; arith stays out unless the count forces it (then flagged).
+2. **Founder budget 600×3 → 1200×3, uniform**, to recover the under-trained lists and binary.
+3. **Prompt spaces enlarged** for roman (1–999), binary (1–511), prime (≤ 400) so a 600-item training
+   set cannot cover the test space (pass 1: 300 / 200 / 210 unique of 600).
+4. **Three candidates added** — `wordlen`, `lettercount`, `sumeven` (counting and filtered sums;
+   verifier-safe; large prompt spaces). Three dropped without retraining: sortletters, caesar,
+   progression (specialist ≤ 0.20).
+
+None of these touches a hypothesis, a threshold, or the campaign design; they change which families
+are *eligible*. Pass 2 is `configs/llm/society_v2_calib_a2.yaml`.
+
+**Stage A, pass 2 (`results/llm_society_v2_calib_a2`, founders 1200×3).** Nine in band: strings 1.00,
+setops 0.85, numtheory 0.83, mixedtoken 0.74, digits 0.97, alphabet 0.97, prime 0.96, wordlen 0.92,
+lettercount 0.77 (specialist accuracies; all bases 0.08–0.30). Roman misses only the base *floor*
+(base 0.04, specialist 0.99). The borderline families did **not** converge with more training —
+vectors 0.61 → 0.53, lists 0.44 → 0.50, binary 0.44 → 0.22, charfreq 0.38 → 0.55, liststats
+0.38 → 0.47 — so they are noise-level at 0.5B, not under-trained; sumeven is unlearnable (0.06).
+C1c min distance 0.68 (pass). C4: 9 of 10 candidates have confidence AUC ≥ 0.7 (mixedtoken 0.62).
+
+**C1b fails among the ten.** Max pairwise conflict 0.439 (wordlen × lettercount); five pairs ≥ 0.35,
+all among the *counting* families (wordlen, lettercount, digits, strings, mixedtoken): two specialists
+that both confidently answer "How many …?" with different small integers is precisely E12's
+conflicting-convention hazard, and under confidence-routed union the wrong one can win the prompt.
+The largest subset with every pair < 0.35 has **seven** members. Dropping lettercount (the hub, 7
+conflicts) leaves nine with max 0.396.
+
+**Re-deriving the gate from the grid, rather than from a midpoint.** 0.35 was chosen as midway between
+the grid's no-conflict axis (0.20–0.26) and its conflict axis (0.46–0.52). The grid's own outcome data
+say where the break is: P(merge penalty > 0.02) is **0.29 for epi_conf < 0.35, 0.22 for
+[0.35, 0.41), and 0.77 for ≥ 0.41** (n = 17 / 9 / 13). Below 0.41 the measure does not predict a
+penalty; above it, it does. A gate at **0.41** is therefore the data-derived boundary, and the
+nine-family set passes it (max 0.396).
+
+**Options put to GG (2026-09-07 evening):**
+1. *L = 9, gate 0.41* — strings, setops, numtheory, mixedtoken, digits, alphabet, prime, wordlen,
+   roman. Two amendments: gate 0.35 → 0.41 (grid-derived, above), base floor waived for roman (the
+   floor screened for unlearnable tasks; roman's specialist at 0.99 settles that). H1's family
+   threshold scales to ≥ 5 of 9. *Recommended.*
+2. *L = 7, gate 0.35 as written* — alphabet, mixedtoken, numtheory, prime, roman, setops, strings.
+   Weaker combinatorics (128 states, ~2.8 doublings) but no gate amendment.
+3. *Prompt tags per family* to suppress off-family confidence, then re-calibrate everything (~1.5 h).
+   Removes the conflict by construction; also makes routing trivially lexical (the `llm_moe` rider).
+
+**GG decision (2026-09-07, 20:30): option 1.** L = 9: strings, setops, numtheory, mixedtoken, digits,
+alphabet, prime, wordlen, roman. C1b gate amended 0.35 → 0.41 (grid-derived); base floor waived for
+roman. Consequential edits to §3/§5/§7: N = 9 agents; `n_test` 27/family (243 overall, SE 0.032);
+`n_val` 13/family; `n_conf` 13/family; H1's family threshold ≥ 5 of 9; B₀ ≈ (spec + 8·base)/9.
+Stage B launched on the nine (`society_v2_calib_b.yaml`; probe families setops / alphabet / digits;
+cross setops × alphabet).
+
+**Stage B, C2 transmission (`results/llm_society_v2_calib_b`) — FAILS as pre-registered.** Retention of
+a founder's own family in a child distilled from the founder's own answers, k examples per family × 9
+families, 2 or 3 epochs:
+
+| k | setops | alphabet | digits |
+|---|---|---|---|
+| 25 | 0.31–0.38 | 0.30–0.34 | 0.59–0.69 |
+| 50 | 0.67–0.73 | 0.50–0.64 | ~0.71 |
+| 100 | 0.62–0.67 | 0.57–0.67 | — |
+| 150 | 0.61–0.73 | 0.73–0.81 | — |
+
+No k ≤ 150 reaches 0.85; the curve is flattening. This is **not** the E2 observation floor — the
+source *supplied* the family at 0.77–1.00 accuracy, so the items were observed. It is interference: a
+one-family founder's inheritance data is one competent family and eight families of confident
+garbage, and a fresh LoRA fits all nine. That is the mechanism behind pilot v1's "distillation tax",
+now measured at 20–40% per generation. Under the pre-registered rule the ungated design is infeasible
+at 0.5B.
+
+**Proposed amendment (measured before adoption, C2b — `society_v2_calib_c2b.yaml`):
+confidence-gated inheritance.** The child learns only the prompts its source is confident on
+(exp mean token log-prob ≥ τ). Verifier-free; identical in every arm; makes the child *agnostic*
+rather than *wrong* off-expertise — E8's founder model. C2b measures retention by τ, the Youden τ*
+separating own- from off-family confidence (calibration uses family labels; the campaign uses the
+fixed τ), and the off-family harm of ungated inheritance. Adoption requires GG's sign-off because it
+changes §3.4 step 4. Implemented as `conf_gate` in `society_v2.py` (default None = ungated).
+
+**C2b (`results/llm_society_v2_calib_b_transmission_conf`) — the gate passes at τ = 0.5, and the
+mechanism is two-part.** Source answers a 300-per-family pool (2700 prompts); child keeps prompts with
+source confidence ≥ τ; 3 epochs. Retention (child own-family ÷ founder):
+
+| | ungated | τ = 0.5 | τ = 0.7 | τ = 0.85 |
+|---|---|---|---|---|
+| setops | 0.69 | **0.93** | 0.93 | 0.94 |
+| alphabet | 0.86 | 0.85 | 0.88 | 0.82 |
+| digits | 0.89 | 0.83 | 0.76 | — |
+| mean | 0.81 | **0.87** | 0.86 | — |
+
+Two drivers, not one. (i) **Pool size**: at k = 300 alphabet and digits already retain ≥ 0.86
+ungated (they were 0.73–0.81 at k = 150). (ii) **Confidence gating** rescues the family pool size does
+not (setops 0.69 → 0.93) and is neutral-to-slightly-negative where the specialist is confident
+everywhere (digits: off-family confidence median 0.67 vs 0.45 for setops — the gate discards data
+without separating). Youden τ* ≈ 0.95–0.97 in all three (own-family confidence median 1.00), so
+τ = 0.5 is a *mild* gate keeping 50–75% of the pool. **Off-family harm: not confirmed.** Ungated children
+score at base off-family (0.19–0.27 vs base 0.20–0.22), not below it; gated children slightly above.
+Interference costs own-family retention; it does not push off-expertise competence below the prior.
+
+**Adopted for the campaign (pending GG go/no-go on the full table): `k_inherit` = 300,
+`conf_gate` = 0.5, `epochs` = 3.** Mean retention 0.87 ≥ 0.85 meets the pre-registered C2 criterion
+under the amended channel. Cost consequence (§9): per-generation inheritance answers 2700 prompts per
+parent and children train on ~1300–2000 kept examples → ≈ 30 min per generation-arm, ≈ 6 h per
+(seed, arm) at G = 12, **≈ 96 L40S-h** for 16 elements; PBS walltime raised to 8 h.
+
+**Stage B, C3 cross setops × alphabet (`results/llm_society_v2_calib_b_cross`) — operator half
+passes, retention half fails for C2's reason.** Parents 0.84 / 0.96. Union child 0.57 / 0.62 — holds
+*both* families, routed 53% of prompts to the alphabet parent. Linear-blend child (best of 6 screened)
+0.66 / 0.30 — keeps one family and loses the other: E4's dilution, in the operator the campaign
+dropped. Union ≥ linear on the minimum (0.57 vs 0.30) ✓. Union ≥ 0.85 × parent ✗ (0.68× / 0.65×) — the
+same transmission tax as C2. The cross is re-run under the confidence gate if C2b passes
+(`conf_gate` is now a cross-stage option; both children gated by their own source's confidence).
+
+**C3 re-run under the gate (`results/llm_society_v2_calib_b_cross_gated`, k = 300, τ = 0.5) — fails
+for a NEW reason.** Union child setops 0.46 / alphabet 0.73 (0.55× / 0.76× of parents) — both held,
+both diluted. Linear child 0.80 / 0.22 — one skill at 0.95×, the other lost. Read together with C2b
+(a *one*-skill child retains 0.87–0.93 under the same gate and budget): a single skill transmits; a
+**two-skill child dilutes each skill by 25–45% even from union-preserved data.** This is E4's
+conservation law relocated from the sample budget to the *learning* budget — a fixed adapter (r = 16)
+and fixed epochs split across skills. Under it H1 (≥ 5 families at ≥ 0.6 in one agent) is predicted to
+fail by construction, whatever the operators do. F8 in the fault ledger.
+
+**C3b (before deciding anything): does capacity or budget lift the two-skill child?** Three variants
+of the gated cross — 6 epochs at r = 16; r = 64 (α = 128) at 3 epochs; both. Pass criterion as C3:
+union child ≥ 0.85 × parent on *each* family. If one passes, the campaign adopts that child budget
+(cost re-estimated). If none passes, the vertical claim cannot be tested at 0.5B with self-distilled
+inheritance, and the honest options are a reduced campaign (H3–H5 only, which do not need multi-skill
+children) or 7B.
+
+**C3b results (`results/llm_society_v2_calib_c3b_*`; union child accuracy and ×parent):**
+
+| child budget | setops | alphabet | verdict |
+|---|---|---|---|
+| r16, 3 ep, τ 0.5 (C3 gated) | 0.46 (0.55×) | 0.73 (0.76×) | fail |
+| **r16, 6 ep, τ 0.5** | **0.73 (0.87×)** | **0.79 (0.82×)** | at the gate within noise (SE ≈ 0.06 on the ratio) |
+| r64, 3 ep, τ 0.5 | 0.72 (0.90×) | 0.58 (0.67×) | fail; r64 founders weaker (0.80/0.87) and less confident (35% routed to alphabet) |
+| r64, 6 ep, τ 0.5 | 0.35 (0.44×) | 0.50 (0.57×) | fail — overfits |
+
+**Budget, not capacity, is the lever; rank stays 16.** The kept count explains the residual: at τ = 0.5
+the union child kept 1941 of 2700 prompts, of which only ~600 are its two competent families — the
+mild gate passes the *max* of two parents' confidences, so ~70% of the child's data is confident
+garbage. C2b's own table had τ = 0.85 as the best mean retention (0.88) at a third of the data.
+**C3c** (last calibration run): the cross at τ = 0.85, 3 and 6 epochs.
+
+**Prediction update carried into the go/no-go, whatever C3c says.** A two-skill child retains ≈ 0.85×
+per skill at best; H1 as written (best agent ≥ B₀ + 0.20, ≥ 5 of 9 families at ≥ 0.6) needs five or
+six skills co-resident at ≈ 0.6 in one r = 16 adapter, which the calibration does not support at 0.5B.
+The realistic bar the calibration *does* support — **H1′: children holding 2–3 families beat every
+founder on overall accuracy (≥ B₀ + 0.05) and the best agent climbs monotonically for ≥ 3
+generations** — is recorded now, before the campaign, as the primary vertical readout, with H1 kept
+as the stretch criterion. H3, H4, H5 do not need multi-skill agents and are unchanged.
+
+**C3c results (`results/llm_society_v2_calib_c3c_*`, τ = 0.85):** 3 epochs → union child 0.72 / 0.79
+(0.86× / 0.82×), kept 1041 of 2700; 6 epochs → 0.69 / 0.75. The tight gate reproduces the 6-epoch
+mild-gate retention at half the training, and more budget beyond that buys nothing: a two-skill child
+plateaus at ≈ 0.85× / 0.8× of its parents. **Final child budget: τ = 0.85, 3 epochs, k = 300, r = 16.**
+Cost re-estimate (§9): inheritance answers (9 parents × 2700 prompts with log-probs) now dominate at
+≈ 11 min per generation-arm; training ≈ 9 min; evaluation ≈ 3 min → ≈ 25 min per generation-arm,
+≈ 5 h per (seed, arm), **≈ 80 L40S-h** for 16 elements. Walltime 8 h.
+
+**Stage B, C5 consensus (`results/llm_society_v2_calib_b_consensus`) — passes.** Consensus accuracy over
+the nine founders at gen 0 = 0.31 (< 0.35); corr(conformity, own accuracy) = +0.29; min pairwise
+distance 0.63. Conformity is not a truth proxy here, so H3 is testable by the predicted mechanism.
+
+## 12a. Go/no-go (GG, 2026-09-07, 21:30): **NO-GO at 0.5B; plan 7B.**
+
+Grounds: the calibration passed C1, C4, C5, and — under the amended inheritance channel — C2, but C3's
+retention half exposed a ceiling no budget moves: a two-skill child holds each skill at ≈ 0.85× / 0.8×
+of its parents, and the vertical claim needs five or six skills co-resident in one r = 16 adapter. At
+0.5B the society experiment could test H3–H5 but not the claim the paper's stated gap is about.
+Nothing is submitted. What today produced is a **measured transmission ceiling for self-distilled LoRA
+inheritance at 0.5B** — three mechanistically distinct limits (near-clone founders; interference from
+confident off-expertise answers, 20–40%/generation, removable by a confidence gate; the multi-skill
+learning-budget plateau) — and it goes in the SI as the reason the 0.5B tier was not run (§11, row 4).
+
+## 13. The 7B plan (for GG review; nothing runs without a go)
+
+**What changes at 7B, and why it is not a re-run.** The premise "one founder is the sole expert on its
+family" requires a base that *cannot* do the family. Qwen2.5-7B-Instruct already scores 0.99 on easy
+arith, 0.69 on easy strings, and will be high on roman / binary / setops / digits; the base bound
+(≤ 0.40) will exclude most of the current nine. The hard variants exist for three families only (7B
+base: lists 0.34, strings 0.67, arith 0.48). **Phase 0 is therefore task design**: ≥ 9 disjoint
+families with 7B base ≤ 0.4 and specialist ≥ 0.8 — multi-step, cipher, and compositional variants of
+the current generators — plus the C1 band and conflict gate re-run at 7B. This is a day of work before
+any GPU time, and it cannot be done on the 16 GB local card (7B training needs the L40S), so every
+calibration step goes through the CX3 queue (183 queued at last check).
+
+**Cost anchors (L40S, from `llm_merge_hpc` / `llm_moe_hpc` / `llm_hard`):** 7B generation ≈ 10
+prompts/s (0.5B ≈ 40); 7B LoRA SFT ≈ 15 example-passes/s (0.5B ≈ 60). Per generation-arm at N = 9,
+k = 300, τ = 0.85, 3 epochs: inheritance answers 9 × 2700 / 10 ≈ 40 min; evaluation 18 × 373 / 10
+≈ 11 min; training 9 × (≈ 1000 × 3 / 15) ≈ 31 min → **≈ 80 min per generation-arm**.
+
+| scope | elements | G | per element | total L40S-h | walltime |
+|---|---|---|---|---|---|
+| Phase 1 — calibration A + B at 7B | 2 jobs | — | ~2 h each | **~4** | 4 h |
+| Headline: `full` vs `no_grounding`, 3 seeds | 6 | 10 | ~13 h | **~80** | 2 × 8 h with resume, or one 16 h |
+| H3 + H4: 3 arms, 3 seeds | 9 | 10 | ~13 h | **~120** | as above |
+| Full grid: 4 arms, 4 seeds | 16 | 12 | ~16 h | **~260** | 3 × 8 h with resume, or one 24 h |
+
+Checkpoint/resume already makes multi-requeue elements safe. A vLLM generation path would cut the
+dominant 40-minute term by 5–10× but adds a dependency and a second code path; noted, not proposed.
+
+**Gates carried over unchanged:** C1 band (base ≤ 0.40 strict, specialist ≥ 0.60; upper bound 1.0), C1b
+conflict < 0.41, C1c distance ≥ 0.5, C2/C2b retention ≥ 0.85 (gated channel), C3 union ≥ 0.85× per
+family on a two-founder cross, C4 AUC ≥ 0.7 for ≥ 8 of 9, C5 consensus < 0.35. **The 7B-specific
+prediction that decides whether to proceed past Phase 1:** with the larger adapter margin at 7B, the
+two-skill cross should clear 0.85× on *both* families at 3 epochs. If it does not, the multi-skill
+plateau is not a 0.5B artefact and the vertical claim should be pursued with a different inheritance
+channel (e.g. inheriting *weights*, not answers — which is what `llm_merge_hpc` already showed
+composes at 7B) rather than with more scale.
+
+**Hypotheses:** H1 restored as written (≥ 5 of 9 families at ≥ 0.6, ≥ B₀ + 0.20) — that is the point of
+going to 7B; H1′ kept as the fallback readout; H3–H6 unchanged; H2 deferred.
+
+**Decisions for GG before Phase 0 starts:** (i) scope row from the table; (ii) whether Phase 0 task
+design is worth the day, given the alternative in the prediction paragraph above; (iii) whether the
+SI text for the 0.5B ceiling (§11 row 4) is drafted now or after 7B.
+
+## 14. Build log
+
+- 2026-09-07 — `families.py` (17 candidates, all self-verifying and deterministic), `society_ops.py`
+  (pooled survival, capped complementary mating, confidence-routed union, score-proportional single
+  parent), `society_v2.py` (`kind: llm_society_v2`; per-generation checkpoint + resume; founder lock
+  for concurrent arm-jobs; H6 source diagnostics; families-alive), `calibrate.py`
+  (`kind: llm_society_calib`, stages families / transmission / cross / consensus), configs
+  (`society_v2_calib_a/b`, `society_v2_smoke`), `hpc/llm_society_v2.pbs` (16-element seed × arm
+  array), `figures/plot_llm_society.py` (the §8 layout, written before unblinding), 9 new pure tests
+  (164 green). v1 code path untouched and still green. Smoke → calibration A launched locally.

tasks/prereg-llm-society-v4.md

@@ -0,0 +1,631 @@
+# Pre-registration — `llm_curriculum` v4: does a society accumulate more than its members?
+
+**Status:** draft for GG review, 2026-09-08. Supersedes `prereg-llm-compose-v3.md` (run; H1 passed,
+H2–H5 null and uninterpretable). Nothing runs until §5's gates pass and GG signs off §8.
+
+---
+
+## 0. Why v4: v3 measured the wrong thing
+
+v3 had a **fixed skill set**. Two founders were trained once and every later generation was a lossy
+copy, so the experiment's ceiling was its own generation 0 and no outcome could have shown capability
+climbing. It answered "do ancestral skills degrade?", a retention question. The paper's claim (C3) is
+that capability **climbs** — each specialty re-earned and exceeded. GG, 2026-09-08: *"are models
+learning NEW skills at EACH generation? or are we just seeing if the ancestral skills degrade?
+because that was not the problem being addressed… we need to ground this into continual learning."*
+
+The specific technical fault: v3 trained **a fresh LoRA from the base each generation**, so knowledge
+survived only through the data channel. That is Weismannian — nothing acquired is inherited as
+structure. v4's children **start from their parent's adapter**, which is the actual Lamarckian
+channel and the precondition for accumulation.
+
+**Pattern across v2 → v3 → v4, recorded so it stops recurring:** each design was checked against the
+*mechanism* (drift, immigration, recombination) and never against the *claim*. §5's gate G0 exists
+solely to check the claim is reachable before any compute is spent.
+
+---
+
+## 1. Design
+
+**Curriculum.** Nine task families from v2's calibrated set (`llm_society_v2_calib_a2`: base ≤ 0.40,
+specialist ≥ 0.60, pairwise conflict < 0.41 — strings, setops, numtheory, mixedtoken, digits,
+alphabet, prime, wordlen, roman). Three lineages, nine generations. **Each lineage sees all nine
+families in a different order** (a cyclic Latin square), so at generation *t* every lineage has met
+*t* families but **different ones**. Complementarity is maximal early and decays to zero by
+generation 9 — a shape the analysis can test, not just a condition it assumes.
+
+**One generation, per lineage:**
+1. **Acquire** — the environment presents the next family; train on `n_new` verified real examples.
+2. **Inherit** — training starts from the *parent's adapter*, not the base (Lamarckian transmission).
+3. **Maintain** — old families are kept alive by `n_replay` real examples (grounding = immigration),
+   or by self-generated answers (dry), or not at all, depending on arm.
+4. **Recombine** — merge with the arm's partner (contemporary, ancestor, or nobody), weights chosen
+   on a held-out validation split (directed recombination, carried over from v3).
+
+**Two external-information channels, deliberately separated** — v3 conflated them. *Acquisition* is a
+capability the population never had (novel allele; moves the frontier). *Replay* is re-supply of a
+capability already present (immigration proper; fights loss, never advances). They fight different
+diseases and must be separate factors.
+
+## 2. Arms
+
+| arm | recombines with | old skills maintained by | isolates |
+|---|---|---|---|
+| `isolated` | nobody | real replay | asexual continual learning — the drift/forgetting baseline |
+| `society` | a decorrelated contemporary | real replay | the treatment |
+| `society_dry` | contemporary | self-generated only | replay's contribution (E2 immigration) |
+| `seed_bank` | **its own ancestor at t−3** | real replay | see below — this is not a throwaway control |
+
+**The seed-bank arm is a substantive comparison, not a null.** My first reading was that an ancestor
+is your own lineage and therefore highly correlated (ρ→1), so E8 predicts it buys nothing. Working
+through the curriculum shows that is wrong: **your t−3 ancestor knows exactly the families you learned
+three generations ago and have since been forgetting.** It carries *temporal* complementarity where a
+contemporary carries *spatial* complementarity. So the arms pose a real question with predictions
+pulling opposite ways:
+
+- E8 (decorrelation is the fuel): the contemporary is more decorrelated → larger union → `society` wins.
+- E12 (merge compatibility): the ancestor is same-lineage, so no Bateson–Dobzhansky–Muller
+  incompatibilities have had time to accumulate → it merges *more safely* → `seed_bank` wins.
+
+Which dominates is not obvious from the framework, and the answer is directly translational: *when a
+model forgets, is it better recovered from a peer who knows something else, or from your own earlier
+checkpoint?* Nobody has posed that as a population-genetic question.
+
+**Dropped (GG, 2026-09-08):** a `no_acquisition` arm — v3 already is that experiment.
+
+## 3. The mono-generational baselines, and why they are the falsifier
+
+The difficulty GG named is pairing a multigenerational design against single-shot SoTA. Rather than
+work around it, it becomes the spine: **does structure across time beat the same compute spent all at
+once?** Three references, all at matched total training examples:
+
+| baseline | what it is | role |
+|---|---|---|
+| `sequential` | one model, all nine families in sequence | the standard continual-learning baseline |
+| `single_shot_merge` | three specialists trained from base in parallel (three families each), merged **once** at the end | actual SoTA — LoRA Soups / TIES / model soup |
+| `joint` | one model trained on all nine jointly | the conventional ceiling |
+
+**Budget accounting (must be equal, and is checked in the artifact).** A lineage trains
+9 generations × (`n_new` + `n_replay`) examples; three lineages give 27 × (`n_new` + `n_replay`).
+Each single-shot specialist gets 9 × (`n_new` + `n_replay`) so three of them match exactly;
+`sequential` and `joint` receive the same total. Compute per arm is recorded in the manifest.
+
+**If `society` does not beat `single_shot_merge` at matched budget, iterating buys nothing and the
+multigenerational framing is decoration.** That is the claim worth staking, and the merging
+literature has never tested it because every paper in it merges once.
+
+## 4. Hypotheses
+
+Primary outcome: **cumulative capability** — accuracy of the population's best model on *all nine
+families*, at each generation. Secondary: per-family forgetting curves, union-exceedance at each
+merge, ρ between partners, and the complementarity decay predicted by the Latin-square design.
+
+| | prediction (source) | threshold | falsified if |
+|---|---|---|---|
+| **H1** | Capability **climbs**: `society` at generation 9 exceeds its own generation 1 by ≥ 0.15 | ≥ 0.15 in ≥ 2 of 3 seeds | flat or declining → the design still cannot show accumulation (this is G0 restated as a result) |
+| **H2** | Fisher–Muller (E7): `society` > `isolated` on cumulative capability at generation 9 | ≥ +0.08, 3/3 seeds positive | recombination adds nothing over isolated continual learning |
+| **H3** | **The multigenerational claim**: `society` > `single_shot_merge` at matched budget | ≥ +0.05 | iterating buys nothing; the framing is decorative and the paper should say so |
+| **H4** | Immigration (E2): `society` > `society_dry` | ≥ +0.08 | self-generated replay suffices; grounding is not load-bearing here |
+| **H5** | Spatial vs temporal complementarity: `society` ≠ `seed_bank`, direction **not** pre-committed (§2 gives arguments both ways) | report with CI either way | — (this is a measurement, not a gated prediction) |
+| **H6** | Complementarity decays by construction, so the `society` − `isolated` gap is **largest at intermediate generations** and shrinks by generation 9 | peak gap at 3 ≤ t ≤ 6 | a monotone gap → the advantage is not coming from complementarity, and the mechanism story is wrong |
+
+H6 is the design's internal check: the Latin square makes complementarity a *known* function of
+generation, so the framework predicts the shape of the advantage, not just its sign.
+
+## 5. Gates — G0 is the one that would have caught v2 and v3
+
+| gate | what | pass criterion |
+|---|---|---|
+| **G0 — can capability climb at all?** | One lineage, 3 generations, 3 families, `isolated` settings. Measure cumulative accuracy over families seen. | Generation 3 exceeds generation 1 by ≥ 0.10. **If capability cannot accumulate in the simplest arm, no outcome of the full design is interpretable — stop.** |
+| **G1 — does inheritance transmit?** | Adapter-continued training on family 2 starting from the family-1 adapter | family-2 accuracy ≥ 0.6 × a from-scratch specialist's |
+| **G2 — does forgetting occur?** | Same, measuring family-1 accuracy after learning family 2 without replay | family-1 accuracy drops ≥ 0.15 (if nothing is forgotten, replay and recombination have nothing to fix) |
+| **G3 — does recombination combine?** | Merge two lineages holding disjoint families at generation 3 | merged model ≥ 0.8 × each parent's accuracy on that parent's own families |
+| **G4 — budget parity** | Recorded example counts across all arms and baselines | equal to within 2% |
+
+G1 and G2 must **both** pass: transmission without forgetting means nothing decays; forgetting
+without transmission means nothing accumulates. The experiment needs the tension.
+
+## 6. Cost
+
+Base: Qwen2.5-1.5B (base, not Instruct — v3's C1 measured that instruction-tuned checkpoints already
+hold these skills). Per generation-lineage: train (`n_new` + `n_replay` ≈ 400 examples × 3 epochs)
+plus evaluation on nine families × 60 items. ≈ 6 min. Society arms: 3 lineages × 9 generations ×
+4 arms ≈ 108 generation-lineages ≈ **11 GPU-h**; baselines ≈ 2 GPU-h; three seeds ≈ **40 GPU-h**
+total. Seed 1 local overnight, seeds 2–3 as a CX3 array — the same split that worked last night.
+
+## 7. Engineering
+
+Reused unchanged from v3: execution/verifier infrastructure, LoRA training, merge operators,
+directed weight selection on a disjoint validation split, checkpoint/resume, the figure and stats
+scaffolding. Reused from v2: the nine calibrated families and their difficulty/conflict measurements.
+
+New: adapter-continued training (inherit the parent's weights rather than a fresh LoRA); the
+Latin-square curriculum scheduler; replay buffers per lineage; the ancestor registry for `seed_bank`;
+budget accounting in the manifest; the three mono-generational baselines.
+
+## 8a. Gate record (2026-09-08) and the v5 curriculum
+
+**G0 passes.** Three families, three lineages: `all_families` (mean over every family in the
+curriculum — the accumulation metric) climbs 0.739 → 0.950 (isolated) and 0.611 → 0.939 (society).
+The design can show accumulation, which v3 structurally could not. A labelling hazard caught on the
+first line: `retention_seen` (mean over families *taught*) starts near 1 and can only fall — watching
+it would have recreated v3's error. Both are recorded; the primary is `all_families`.
+
+**G2 fails on the v2 families, twice.** Three-family gate with replay: nothing forgotten (every family
+only rises). Nine-family single-lineage probe with replay **off** (`results/llm_curriculum_g2`): mean
+drop across families learned before the last is only **+0.074**, and it is carried by one family —
+`mixedtoken` 0.80 → 0.15 (+0.65), oscillating violently throughout (0.18, 0.20, 0.27, 0.80, 0.60,
+0.37, 0.42, 0.20, 0.15) — while two families *improve* through positive transfer (numtheory −0.12,
+alphabet −0.13) and the rest are within ±0.08. Forgetting here is a single pairwise-interference event
+between confusable counting families (the cluster v2's conflict measure flagged at 0.44), not a general
+pressure a population could smooth.
+
+**G3 is negative on the v2 families.** `society − isolated` = −0.128, −0.022, −0.011 across the three
+gate generations. Merging costs and cannot pay, because a partner can only contribute what the recipient
+lacks and nothing was lacking.
+
+**The base reference quantifies the format confound.** Qwen2.5-1.5B base on all nine v2 families:
+**0.094** (strings 0.02, setops 0.02, alphabet 0.00, wordlen 0.00, prime 0.02, digits 0.07, roman 0.10,
+mixedtoken 0.12, numtheory 0.52). Training on **one** family lifts the nine-family mean to **0.417**.
+Most of the apparent accumulation is a one-off format acquisition shared by all nine families.
+
+**One root, three faults.** The v2 families were built for a *specialisation* experiment and calibrated
+for low mutual conflict so merging would be safe. That makes them (i) format-homogeneous — one family
+teaches the convention for all; (ii) too compatible — no interference, hence no forgetting; (iii) too
+easy — the base is unformatted, not incapable. They are not a curriculum, and no arrangement of arms
+fixes that. The same lesson as v3's task pairing: the paradigmatic continual-learning benchmarks use
+naturally heterogeneous tasks *because* those interfere, differ in format, and exceed a small base.
+
+**v5 curriculum (`src/llm/curriculum_data.py`, 2026-09-08).** Eleven candidates from public datasets,
+each with a disjoint train/test split and its own verifier — gsm8k (number), mbpp (code, executed),
+boolq (yes/no), mnli (3-way label), sst2 (sentiment word), csqa (A–E), arc (A–D), winogrande (1/2),
+squad (extractive span, normalised EM over aliases), nq_open (short text, aliases), hellaswag (A–D).
+All eleven self-verify 40/40 and reject garbage 0/40. Answer shapes span five forms (code, number,
+word, letter, phrase) against v2's one. **Selection rule, fixed before running:** the C1 band
+(base ≤ 0.40, specialist ≥ 0.60) from `curriculum_v5_calib` (one specialist per candidate, evaluated on
+every family — the full transfer matrix), then a single-lineage zero-replay probe over the chosen set
+with **mean forgetting ≥ 0.15 and not carried by a single family** (max single-family share of the
+total drop ≤ 50%). Nine survivors form the curriculum; if fewer than nine pass, L and G shrink to match
+and the Latin square is recomputed.
+
+**v5 stage A (`results/llm_curriculum_v5_calib`, founders at 300 × 3, Qwen2.5-1.5B base).** Base on
+all eleven: **0.011** (mbpp 0.08, gsm8k 0.02, the rest 0.00) — a non-instruct base on real tasks, as
+expected; the C1 base floor is moot. Specialist × family matrix (own-family on the diagonal):
+
+| specialist | own | mean off-family | character |
+|---|---|---|---|
+| mnli | **0.82** | 0.46 | permissive — lifts most others |
+| arc | **0.77** | 0.21 | |
+| hellaswag | **0.72** | 0.27 | |
+| squad | **0.68** | 0.34 | permissive |
+| boolq | **0.65** | 0.04 | **destructive** — zeroes others |
+| csqa | 0.55 | 0.36 | permissive |
+| sst2 | 0.38 | 0.23 | |
+| winogrande | 0.38 | **0.00** | **destructive** — 0.00 on all ten others |
+| mbpp | 0.18 | 0.08 | destructive |
+| nq_open | 0.18 | 0.36 | |
+| gsm8k | 0.10 | 0.42 | permissive (0.92 on sst2, 0.83 on arc; 0.10 on gsm8k itself) |
+
+Five pass C1 at this budget. Mean off-family transfer 0.25 against own-family 0.49 — half, where v2's one
+family lifted all nine to near-own level. **Two kinds of specialist, which is what a curriculum needs:**
+*format-permissive* ones (mnli, csqa, squad, gsm8k) teach general instruction-following and lift other
+families — the residual format-transfer effect, now bounded and measurable; *format-destructive* ones
+(winogrande, boolq, mbpp) learn one narrow output form and erase the rest. The destructive group is the
+forgetting mechanism made visible: a lineage that meets winogrande loses what it held, and replay or a
+partner who did not just learn winogrande is what can restore it — E8 with something to act on.
+
+The six failures were under-trained, not unlearnable: every specialist that has worked in this project
+(v2, v3) had 1200 examples; these had 300 (gsm8k reached 0.54–0.62 at 1200 in v3). Pass 2
+(`curriculum_v5_calib_b`, the six at 1200 × 3) decides six families or nine. **Budget consequence,
+pre-noted:** if 1200 is what a family needs, `n_new` in the campaign rises accordingly and §6's cost
+scales by ~4× on the training term.
+
+**v5 stage A pass 2 (`results/llm_curriculum_v5_calib_b`, the six failures at 1200 × 3) — the
+"under-trained" hypothesis is refuted.** winogrande 0.38 → 0.57, sst2 0.38 → 0.52, mbpp 0.18 → 0.22,
+gsm8k 0.10 → **0.07**, nq_open 0.18 → **0.07**, csqa 0.55 → **0.18** (chance on 5-way; off-family
+transfer collapsed to 0.00). Three got worse with four times the data — a training instability of the
+fresh-adapter learning rate on these tasks, not a data shortage. gsm8k is a *data-source* issue: v3's
+0.54–0.62 came from MetaMathQA's augmented chain-of-thought, not raw GSM8K. None of the six clears 0.60.
+
+**Selection (2026-09-08): six families.** mnli 0.82, arc 0.77, hellaswag 0.72, squad 0.68, boolq 0.65
+pass C1; **winogrande 0.57** is the sixth. **Amendment:** the C1 specialist floor is relaxed 0.60 → 0.55
+for one family so that F is divisible by L = 3 (the pre-registered shrink rule needs F ∈ {3, 6, 9}).
+winogrande is also the right sixth on the merits: it is the most format-destructive specialist in the
+matrix (0.00 on every other family), i.e. the strongest forgetting pressure available — the mechanism the
+design exists to test. Curriculum: **L = 3, F = 6, G = 6**; complementarity 1.0 at t = 2, 0 at t = 6.
+Founders at **300** examples (the budget that passed; 1200 destabilised). Campaign `n_new` = 300.
+
+**Stage B running:** `curriculum_v5_g2` — one lineage, zero replay, the six in sequence. Pass criterion
+unchanged: mean drop ≥ 0.15 across families learned before the last, no single family > 50% of the total.
+
+**v5 stage B — G2 on the six (`results/llm_curriculum_v5_g2`, one lineage, zero replay).** Order
+mnli → arc → hellaswag → squad → boolq → winogrande. Drops (learned → final): mnli +0.27, arc +0.03,
+hellaswag −0.05, squad **+0.57**, boolq +0.05. **Mean +0.173 — magnitude gate (≥ 0.15) PASSES.**
+All-families 0.575 → 0.725 (gen 3) → 0.625 (gen 5): acquisition then loss as the destructive families
+arrive — the tension the arms need. **Concentration criterion (≤ 50% in one family) MISSES at 62%**
+(squad). Recorded as a marginal miss, with the reasons it does not reproduce the v4 failure the
+criterion was written against: two families forgotten (not one pair), the destruction lands exactly
+where the transfer matrix predicted (winogrande's option format erases span and 3-way label; letter
+formats survive), and the probe tested one order where the campaign's Latin square gives each lineage
+a different one — so different families are forgotten in different lineages, which is the
+complementarity recombination acts on. **Recommendation: go**, pending GG.
+
+## 8b. The v5 campaign result, and the two-kinds-of-variation measurement (2026-09-08)
+
+**Campaign (3 seeds, 6 families, 3 lineages, 6 generations; `results/llm_curriculum_v5/`).** Best model
+per arm at the final generation, mean over seeds: `sequential` (one model, no population) 0.802 ·
+`isolated` (population, never merges) 0.796 · `joint` (multi-task ceiling) 0.748 · `seed_bank`
+(merges with its own ancestor at t−3) 0.663 · `single_shot_merge` 0.549 (0.125 / 0.758 / 0.764 — the
+huge variance is which families landed in which allopatric split) · `society_dry` 0.307 · `society`
+(merges with a contemporary) 0.269. Budget parity within 6%.
+
+Every pre-registered hypothesis fails, consistently across all three seeds: society − isolated
+**−0.527 ± 0.092** (3/3 negative); society − single-shot −0.280 ± 0.363 (unresolved, huge variance);
+society − society_dry −0.038 ± 0.073 (replay policy irrelevant once the merge channel dominates).
+The one uncommitted contrast resolves decisively: **seed_bank − society = +0.394 ± 0.089, 3/3
+positive** — merging with your own past beats merging with a peer, so the compatibility argument
+beats the decorrelation argument.
+
+**Mechanism, identified and isolated.** Two of the six families (boolq, winogrande) are answer-format
+destroyers — the calibration matrix measured them at 0.00–0.04 mean off-family. A lineage that learns
+one propagates it through the merge into partners that never trained on it; because merged offspring
+continue the lineage, the damage compounds (society lineage 0, gen 2→3: five families fall together
+while boolq alone rises). An ancestor cannot transmit a family the lineage never met, which is exactly
+why the seed-bank arm holds.
+
+**Scope limits, stated plainly.** (i) Merging was **obligate** — no veto, no option to keep the parent
+unchanged. (ii) There is **no selection between lineages**: all three persist regardless of fitness,
+so the design has transmission, acquisition, gene flow and immigration but no differential
+reproduction. It is a gene-flow experiment, not a natural-selection one, and is therefore not a test
+of the composed-society claim. (iii) The operator was linear averaging, which the merging literature
+ranks below concatenation — but concatenation doubles adapter rank per merge, so iterated merging
+faces a capacity constraint single-shot merging never meets (16 → 1024 over six generations). That
+constraint is itself a finding about iteration.
+
+**Two kinds of variation are opposite in sign (`/tmp/paralleldiv.py`, 2026-09-08).** Three adapters on
+the *same* family differing only in seed and data draw: accuracies 0.762 / 0.800 / **0.312** (one run
+simply failed — training instability); pairwise output disagreement 0.237 between the two good ones;
+weight cosine **+0.006** (near-orthogonal); either-right 0.887 vs both-right 0.675. **Merging the two
+good ones gives 0.887 — +0.087 over the better parent, landing exactly on the either-right ceiling.**
+Same base, same linear operator, same scale, same evaluation as the collapsing arms.
+
+So the framework's single decorrelation parameter conflates two quantities that behave oppositely:
+- decorrelation in **what parents know** → risk (−0.53 measured);
+- decorrelation in **how parents encode the same knowledge** → benefit (+0.087, at the ceiling).
+
+Equal weights are best for same-skill merging (0.887 at 0.5/0.5 vs 0.863 at 0.3/0.7), the reverse of
+skill composition where asymmetric weights won decisively (v3: 0.507 at 0.2/0.8 vs 0.333 at 0.5/0.5).
+The optimal merge weight is a signal of which regime the merge is in.
+
+**Design consequence for v6 (GG, 2026-09-08 — selection is needed to generalise the hypothesis).**
+Three arms — never merge · complementary partners (different skills) · **parallel partners (same
+skills, different seed/data draw)** — with selection added in two places: a **veto** ("keep the parent
+unchanged" is always a candidate offspring) and **population selection** (score parents and offspring
+together, keep the best; the failed 0.312 run is exactly what it should discard). Roughly doubles
+evaluation cost per generation: ≈ 10 GPU-h per seed, ≈ 30 for three.
+
+## 8c. Mechanism probes, 2026-09-08 — two of my explanations retracted
+
+Four cheap probes run after the v5 campaign, chasing why cross-lineage merging collapsed. They
+retract two explanations I had given and leave a third standing.
+
+**Probe 1 — same-skill variation (`/tmp/paralleldiv.py`).** Three adapters, one family (arc), same
+data distribution, differing only in seed and draw. Accuracies 0.762 / 0.800 / **0.312** (one run
+simply failed). Pairwise output disagreement 0.237; weight cosine **+0.006** (near-orthogonal);
+either-right 0.887 vs both-right 0.675. **Merging the two good ones: 0.887 — +0.087 over the better
+parent, exactly at the either-right ceiling.** Equal weights beat 0.3/0.7 (0.887 vs 0.863), the
+reverse of skill composition, where asymmetric weights won.
+
+**Probe 2 — signal/noise and the inbred-lines cross (`/tmp/inbred.py`).** Across-seed decomposition:
+signal power 17.8 vs noise 34.9 for arc, 19.0 vs 37.1 for boolq — and correcting for the K=3 sample
+mean's own noise puts the true signal near 6.2, i.e. **~85% of a LoRA's weight change is
+run-specific and arbitrary.** That is why raw weight distance measures mostly noise and predicted
+nothing in Fig. 3C-D. Variance-weighted overlap surfaces ~4x more of the real overlap
+(0.0015 -> 0.0066) but different skills stay near-orthogonal even in signal directions.
+Denoising before crossing helps modestly and on both skills at once: raw mix 0.825 -> denoised mix
+**0.850** (+0.025), while the denoised singles are no better alone (arc 0.825 vs 0.838). The
+inbred-line signature: averaging within a line does not improve the line, it makes it cleaner to cross.
+
+**RETRACTION 1 — "a destructive skill propagates through the merge and destroys partners" is wrong.**
+A single merge of clean single-skill adapters is excellent: arc alone 0.838 (boolq 0.000), boolq alone
+0.800 (arc 0.300), **merged 50/50 = arc 0.863 / boolq 0.787, mean 0.825 vs 0.550 for the best parent.**
+Merging is *protective* — it stops either adapter dominating the output format. Consistent with LoRA
+Soups rather than contradicting it.
+
+**Probe 3 — iterated merging (`/tmp/decay.py`), five chained merges, three weight schemes.**
+arc retention: convex [0.5,0.5] **1.02** · selfish [0.8,0.4] **1.02** · additive [1.0,1.0] **0.52**
+(arc 0.867 -> 0.450, incoming skills 0.033). **RETRACTION 2 — geometric signal dilution is not the
+mechanism.** Convex merging loses nothing over five rounds even though the first adapter's coefficient
+falls to 1/32. The scheme that *preserves* signal coefficients is the only one that collapses, because
+the accumulated change grows without bound and leaves the region where the base still functions. The
+operative constraint on iterated merging is **bounding total drift from the base**, not preserving signal.
+
+**Probe 4 — the scaling dose-response (`/tmp/scale.py`), prompted by GG asking the obvious control:
+does dividing the change vector by 30 retain the skill?** Base (no adapter) **0.000**; scale 1 0.838;
+1/2 0.875; 1/4 0.850; 1/8 0.863; **1/16 0.450; 1/32 0.000; 1/30 0.000.**
+
+**This forces a reinterpretation of Probe 3.** At the coefficient arc actually held after five merges
+(1/32) the adapter alone delivers *nothing*. So the 0.883 measured in the chain was never arc's
+residual. What propagates through a merge is the **answer format**, supplied by whichever partner has
+enough weight to carry it: arc needs "a single letter", and it collapsed at round 4 (partner squad,
+free-text spans, arc 0.567) and recovered at round 5 (partner hellaswag, single letter A-D, arc 0.883).
+The chain measured *format compatibility with the dominant partner*, not skill retention.
+
+**What survives, and what it implies.** (i) A sharp **effectiveness threshold at ~1/8**: an adapter
+works at full strength down to an eighth and collapses below it, so useful merge depth is ~3 rounds at
+convex weights, not 5. (ii) The additive collapse stands (it did not depend on the misreading).
+(iii) The curriculum result becomes coherent for the first time: the two families that destroyed
+everything, boolq and winogrande, are the ones with the most idiosyncratic output formats. If format
+is what propagates, a partner carrying a dominant format overwrites the ability to answer anything
+else. **That is a claim about output conventions, not weight geometry** — and it is consistent with
+Fig. 3C-D, where functional conflict predicted merge damage (rho 0.45) and weight geometry did not (0.03).
+
+**Still untested, and now the leading candidate for the curriculum collapse:** continued training *on
+top of* merged weights. Probe 3 chained merges without ever training between them and lost nothing;
+the curriculum merges then trains, every generation. Test: repeat the chain with a fine-tune on the
+next family after each merge, and see whether that alone reproduces the collapse.
+
+## 8d. Per-skill scaling thresholds, and the bespoke-weights negative (2026-09-08)
+
+**Dose-response per skill (`/tmp/thresholds.py`), accuracy vs adapter scale, base = 0.000 on all six:**
+
+| skill | 1 | 1/2 | 1/4 | 1/8 | 1/16 | 1/32 | own optimum |
+|---|---|---|---|---|---|---|---|
+| arc | 0.87 | 0.88 | 0.88 | **0.92** | 0.42 | 0.00 | 1/8 |
+| squad | 0.72 | 0.73 | **0.78** | 0.65 | 0.10 | 0.00 | 1/4 |
+| hellaswag | 0.75 | **0.82** | 0.80 | 0.63 | 0.07 | 0.00 | 1/2 |
+| boolq | 0.78 | 0.78 | 0.78 | **0.37** | 0.00 | 0.00 | 1–1/4 (flat) |
+| winogrande | 0.55 | 0.55 | 0.55 | 0.37 | 0.30 | 0.00 | 1–1/4 (flat) |
+| mnli | 0.40 | 0.40 | **0.68** | 0.48 | 0.48 | 0.18 | 1/4 |
+
+Three facts. (i) **Thresholds are skill-specific** (GG predicted this): boolq dies at 1/8 where arc,
+squad and hellaswag are still at full strength, so merge depth in a population is set by the *weakest*
+skill. (ii) **Cliffs are sharp** — full effectiveness right up to the edge, then near-total loss in one
+halving; there is no graceful degradation to trade against. (iii) **Four of six skills are BETTER
+scaled down** — mnli 0.40 -> 0.68 at 1/4, hellaswag 0.75 -> 0.82 at 1/2, squad 0.72 -> 0.78 at 1/4,
+arc 0.87 -> 0.92 at 1/8. These adapters are over-trained at full strength; attenuation recovers
+accuracy with no retraining.
+
+**Denoising does NOT move the threshold.** arc raw 0.87/0.88/0.88/0.92/0.42/0.00 vs denoised
+0.87/0.83/0.88/0.87/0.35/0.00; boolq raw 0.78/0.78/0.78/0.37/0.00 vs denoised 0.78/0.78/0.77/0.38/0.03.
+Identical cliffs. So the limit is **signal magnitude**, not signal-to-noise: averaging leaves signal at
+full strength and only removes noise, and therefore cannot buy merge depth. Denoising remains worth
+doing for cross-skill merge quality (+0.025, §8c) and for knowing what a line contains — not for depth.
+
+**Bespoke per-skill merge weights — tested and NEGATIVE (`/tmp/bespoke.py`).** All six skills merged
+into one model:
+
+| scheme | sum | mean |
+|---|---|---|
+| six separate adapters, full strength | — | 0.678 |
+| six separate, each at its own optimum | — | **0.755** |
+| **uniform convex (1/6)** | 1.00 | **0.708** |
+| uniform 0.25 | 1.50 | 0.686 |
+| bespoke: cliff (lowest viable per skill) | 1.12 | 0.689 |
+| bespoke: optimum (best-accuracy per skill) | 1.62 | 0.686 |
+| additive (1.0 each) | 6.00 | 0.156 |
+
+All three non-uniform schemes cluster at 0.686–0.689, *below* plain equal weighting. **Why the
+inference failed:** solo dose-response curves do not transfer to the multi-way case. A skill's
+effective strength in a merge is set by its coefficient *relative to the other five* — six output
+formats compete for one model — so raising one skill's absolute weight starves the others. Clearest
+in squad: 0.65 at uniform 1/6, but 0.47–0.55 whenever given a larger absolute weight alongside others.
+The curves are sound; the inference from them to merge weights was not.
+
+**The two results worth keeping.** (a) **One merged model beats six separate specialists on their own
+tasks** — 0.708 vs 0.678 — with mnli the clearest case (0.40 alone, 0.67 merged, because merging
+dilutes it to near its optimum and undoes the over-training). The merge is doing compression plus
+incidental regularisation, which is a more honest description than "combining capabilities".
+(b) **The largest free win needs no merging at all**: attenuating each specialist to its own optimum
+takes the separate-models baseline from 0.678 to **0.755**, the best number in the table — one scalar
+per adapter, no retraining. Merging then costs ~5 points and saves five models: a real engineering
+trade, honestly stated.
+
+## 8e. The last candidate eliminated — merge-then-train is the best procedure tested (2026-09-08)
+
+**Test (`/tmp/trainmerge.py`).** Two chains, identical partners and order. Control: merge only.
+Test: merge, then continue-training on the partner's family (300) plus replay across everything seen
+(150 split) — i.e. the v5 sequence. The control reproduced the earlier chain **exactly at all five
+rounds**, so the comparison is clean.
+
+| round | partner | merge-only arc | merge+train arc | merge-only partner | merge+train partner |
+|---|---|---|---|---|---|
+| 1 | boolq | 0.883 | **0.933** | 0.733 | **0.850** |
+| 2 | winogrande | 0.900 | **0.933** | 0.550 | **0.717** |
+| 3 | mnli | 0.850 | **0.917** | 0.467 | **0.867** |
+| 4 | squad | 0.567 | **0.883** | 0.733 | **0.767** |
+| 5 | hellaswag | 0.883 | 0.833 | 0.800 | **0.850** |
+
+**Training on merged weights is not the mechanism — it is a substantial improvement.** Better on the
+tracked skill in four of five rounds, better on the incoming skill in all five, and it absorbs the
+round-4 format shock that dropped the control to 0.567. Merge-then-train holds the old skill near its
+ceiling *and* acquires the new one far better than merging alone (mnli 0.867 vs 0.467).
+
+**All three proposed mechanisms for the v5 collapse are now refuted**, each by direct test:
+a destructive skill propagating through merges (§8c — a single merge is protective); geometric signal
+dilution (§8c — five convex merges lose nothing); training on merged weights (here — it helps).
+The one structural difference left is that these chains merge *clean single-skill* adapters, whereas
+v5 merged accumulating lineages carrying up to six skills each in rank 16 — i.e. capacity. **Recorded
+as unexplained rather than attributed:** three mechanisms have been proposed and refuted, and a fourth
+guess would not have earned its place.
+
+**Consequence for the plan (§8b/§8c).** The veto arm was scheduled to make the v5 negative
+interpretable. With three mechanisms refuted and merge-then-train shown sound in isolation, the v5
+result is not reportable whatever a veto arm shows — recommendation is to stop, not to run it.
+
+## 8f. The veto arm (GG overruled my recommendation to skip it — correctly) 2026-09-08
+
+**Change:** identical to v5's `society` arm except that "keep the parent unchanged" is scored on the
+same validation split as the merge candidates, and wins if no weighting beats it. One bit of
+selection. Everything else — operator, weight grid, curriculum, replay, seeds — unchanged.
+
+**Seed 1 result: a single veto converts total collapse into a healthy trajectory.**
+
+| generation | 0 | 1 | 2 | 3 | 4 | 5 |
+|---|---|---|---|---|---|---|
+| veto (declinable) | **0.719** | **0.756** | 0.744 | **0.781** | 0.781 | 0.783 |
+| society (obligate) | 0.689 | 0.719 | 0.700 | 0.597 | 0.439 | **0.211** |
+| isolated (never merges) | 0.625 | 0.700 | **0.728** | 0.753 | **0.789** | **0.814** |
+
+**The veto decisions are structured, and this is the substantive finding.** Merges declined, per
+generation (of 3): 1, 1, 1, **3, 3, 3** — 67% overall, and from generation 3 onward *every* lineage
+declines *every* merge, unanimously. Median validation gain when accepted: +0.042. That timing
+tracks complementarity, which is 1.00 through generation 1, 0.80 at 2, then 0.67 and falling: the
+population discovers on its own that recombination has stopped paying and stops — H6's predicted
+*shape* of the advantage, reached from the opposite direction.
+
+**GG's caveat, and it is the right reading (2026-09-08):** once merges are always declined the arm is
+*literally* the isolated arm, so the comparison at the end is between "merged early, then stopped" and
+"never merged". Early merging gives a large lead (+0.094 at generation 0) but **isolated overtakes at
+generation 4 and finishes higher (0.814 vs 0.783)** — the early merges leave a residual cost that
+never-merging avoids. So the claim is not "the veto fixes recombination". It is: *recombination pays
+only while partners differ, a population can detect when that stops, and even then it ends slightly
+behind never having merged.*
+
+**Consequence for reportability.** This makes the v5 negative interpretable and no longer vulnerable
+to "you forced merging, of course it broke": obligate recombination collapses (0.211), one bit of
+selection rescues it (0.783), and never recombining is still marginally best (0.814). Risk, remedy,
+and honest limit — the language-model rung beside Fig. 5A.
+
+**My error, recorded:** I recommended skipping this experiment on the grounds that v5 was unreportable
+whatever it showed. That judged the experiment by whether it would rescue a conclusion I had already
+written off, rather than by what it would measure. The per-generation veto rate is information neither
+other arm could produce, and it is the most interesting thing in the arm.
+
+**Replication:** seeds 2-3 submitted to CX3 as array `4007703` (`hpc/llm_veto.pbs`), pairing against
+the existing v5 isolated/society/seed_bank runs for those seeds.
+
+**Control worth considering if the seeds hold:** a *forced* stop at generation 3, to separate "the
+veto's timing is smart" from "any early merging then stopping does this". The veto's stopping point
+coincides with complementarity falling below 0.8, which is principled rather than arbitrary — but that
+is an observation, not a test.
+
+## 8g. The vocabulary substrate: contamination screen and the prior-art problem (2026-09-09)
+
+GG's proposal: replace task families with *content* — teach 100 words of a language the model does
+not speak, one word-set per modifier, so thousands of words yield hundreds of modifiers and the
+generation count rises tenfold. Retention is trivially measurable ("what does X mean in English?").
+
+**Contamination screen — the first measurement was invalid.** Generating an answer and string-matching
+it gave Italian 0.092, French 0.050, Basque 0.050, Welsh 0.017, Zulu 0.008, pseudo-words 0.000. French
+tying Basque is impossible if the quantity measured were knowledge, so the probe was measuring whether
+a *base* model obeys "answer with one English word" — the same instruction-following floor that gives
+0.011 on the task families. Re-run as an 8-way forced choice over candidate translations scored by
+likelihood (domain-conditional PMI, chance 0.125), which needs no instruction-following:
+
+| language | generated | forced choice | verdict |
+|---|---|---|---|
+| French | 0.050 | 0.950 | fully known |
+| Italian | 0.092 | 0.908 | fully known |
+| Welsh | 0.017 | 0.508 | half known |
+| Basque | 0.050 | 0.483 | half known |
+| **Zulu** | 0.008 | **0.142** | at chance — genuinely unknown |
+| pseudo-words | 0.000 | 0.158 | floor (cycling 20 nouns inflates this slightly) |
+
+Only Zulu is clean among natural languages; pseudo-words are clean by construction and unlimited in
+supply. Probe: `/tmp/contam2.py`.
+
+**Prior art makes the experiment-as-framed a reproduction.** WikiBigEdit (arXiv:2503.05683) runs
+506K factual QA pairs across 8 sequential timesteps. Locate-then-edit methods (ROME, MEMIT) collapse
+within the first few hundred updates, but their **LoRA + merging** baseline — a fresh adapter per
+timestep, interpolated into the accumulated adapter at weight 0.25 — is stable across the whole
+benchmark and beats every dedicated editing method past ~100K updates. That is our isolated-plus-
+attenuated-merge arm, run three orders of magnitude further, and it does not collapse. Separately,
+arXiv:2506.14126 finds that over-training experts harms merging via late-stage memorisation, which is
+the published version of our §8d observation that attenuating four of six adapters was free gain —
+cite it, do not claim it.
+
+**Consequence for the diagnosis.** If a single lineage accumulates 500K disjoint facts by
+fresh-adapter-plus-interpolation without collapsing, capacity is not what stopped v5 at generation
+3-4 with six families. The difference between the two settings is that WikiBigEdit's content is
+homogeneous QA in one output format, whereas v5's families conflict at the output (label vs span vs
+number vs code). The remaining candidate is interference between competing output formats.
+
+In the population-genetic frame the two are distinct: a new word-set is a **new locus**, and adding
+loci is cheap; two families demanding different output formats for the same input shape are
+**competing alleles at one locus**, and that is what collapses. A pure vocabulary curriculum is all
+loci and no allelic competition, so it would run to a hundred generations and confirm only that
+capacity is ample — removing precisely the variable that produced the phenomenon.
+
+**The collision sweep I proposed here is also occupied — do not run it either.** *In Praise of
+Stubbornness* (arXiv:2502.04390) sweeps exactly this: non-contradictory updates integrate safely, while
+contradictory ones destroy up to 80% of unrelated knowledge with as few as 10-100 facts, consistently
+across model scales, and the authors conclude explicitly that the cause is conflict rather than
+capacity. *Interference and Retention in Continual Learning* (arXiv:2607.09202) supplies the theory:
+disjoint task supports make forgetting structurally eliminable, conflicting overlap imposes an
+unavoidable distortion floor. Both the measurement and its formalisation exist.
+
+**What this buys us anyway: the v5 collapse now has a cause.** Three of our own explanations were
+retracted (§8c, §8d) and capacity was the standing suspect. Between 2502.04390 and 2607.09202 the
+mechanism is settled and it is allelic conflict, not locus exhaustion — the six families conflict at
+the output, and contradictory updates corrupt disproportionately and non-locally. WikiBigEdit is the
+mirror control: homogeneous single-format content accumulates to 506K facts without collapsing. The
+collapse we could not explain is a known, characterised, independently replicated phenomenon.
+
+**The gap that survives.** Every merging paper in the landscape still merges *once* — GENOME (the ACL
+2026 population-evolution paper) evolves a population toward a single target task, with no collapse,
+forgetting or grounding analysis. The iterated reproduction loop is still unoccupied, we have run it,
+and v5's negative answer (obligate recombination collapses by generation 3-4; veto-gated recombination
+merely matches isolation) is now interpretable rather than mysterious. No further LLM compute is
+required to state it.
+
+## 8. Decisions for GG
+
+1. **Three lineages × nine generations × nine families** (complementarity maximal at t=3, zero at
+   t=9), or fewer families and more generations per family?
+2. **Replay budget** — fixed `n_replay` split across all families seen so far (so per-family replay
+   thins as the curriculum grows, which is realistic and makes forgetting a live pressure), or fixed
+   per-family (constant protection, more compute)?
+3. Whether `single_shot_merge` gets the directed weight selection the society arms use, or plain
+   uniform soup as published. I would give it the *same* selection, so the comparison isolates
+   iteration rather than handing the society a free operator advantage.
+
+## 8h. Two controls for the declinable merge, pre-registered before running (2026-09-11)
+
+Both were identified in the 2026-09-11 manuscript review as the weakest hedges in the six-generation
+population section. Code: `merge_until` and `orders` config keys in `src/llm/curriculum.py`;
+configs `curriculum_v5_stop3.yaml`, `curriculum_v5_decor.yaml`; stats `figures/stats_llm_curriculum.py`.
+
+**Control 1 — forced stop at generation 3 (`llm_curriculum_v5_stop3`).** The v5 `society` arm with
+recombination switched off from generation 3 (`merge_until: 3`, `allow_veto: false`). Rationale: in
+the seed-1 veto run lineages declined 1/3 of merges at generations 0–2 and 3/3 at 3–5, so this is the
+matched fixed schedule. Readout: best-lineage all-family accuracy at generation 5, paired per seed
+(3 seeds) against veto, isolated and society.
+- stop3 ≈ veto (within ±0.03 in every seed): the veto's outcome is explained by *when* it stopped;
+  the paper keeps "the population found the schedule by itself" and drops any claim that per-decision
+  evaluation adds value beyond timing.
+- stop3 < veto in every seed: the early declines avoided specific harmful merges; the modifier reading
+  strengthens.
+- stop3 ≈ society (collapsed): three obligate merges already carry the format destroyers; stopping is
+  not enough, screening is required.
+
+**Control 2 — decorrelated curriculum (`llm_curriculum_v5_decor`).** Same six families, G = 6, but
+every lineage starts with mnli, then diverges maximally, then converges, so partner complementarity by
+generation is 0.00, 0.67, 0.70, 0.58, 0.33, 0.00 (Latin square: 1.00, 1.00, 0.80, 0.67, 0.33, 0.00).
+Arms: `isolated` and `society` with `allow_veto: true`. Primary readout, pooled over both curricula
+(2 × 6 generations × 3 seeds = 36 points of mean `veto_used`): partial Spearman correlation of the
+fraction declined with complementarity, controlling for generation (rank-regress both on generation,
+correlate residuals), seed-clustered bootstrap CI.
+- Modifier hypothesis: partial ρ(declined, complementarity | generation) < 0 with CI excluding 0.
+- Adapter-age hypothesis: that partial ρ ≈ 0 while partial ρ(declined, generation | complementarity) > 0.
+- Known ambiguity, stated in advance: at generation 0 of the new curriculum all lineages hold the
+  *same* family from different training draws, and §8b measured that such same-skill merges gain
+  +0.087 (encoding decorrelation). A low decline rate at generation 0 therefore does not test
+  complementarity; the primary test is the pooled partial correlation, not that point.
+- Secondary, descriptive: whether the new curriculum's veto arm finishes level with its own isolated
+  arm, as in the Latin square (0.792 vs 0.796).
+
+Seed 1 of each control runs locally (batch 24 / train batch 2, as the v5 seed-1 runs); seeds 2–3 on
+CX3 (`hpc/llm_curriculum_controls.pbs`, batch 48 / train batch 4, as the v5 seeds 2–3 runs).
+
+### §8h outcome (2026-09-11, 3 seeds each; `figures/stats_llm_curriculum.py`)
+
+- **Control 1, forced stop:** stop3 0.793 vs veto 0.792 vs isolated 0.796 (per-seed veto − stop3:
+  −0.008, −0.006, +0.011). First branch: the veto's outcome is explained by *when* it stopped.
+- **Control 2, decorrelated curriculum:** partial ρ(declined, complementarity | generation) = −0.067,
+  CI (−0.211, +0.088); partial ρ(declined, generation | complementarity) = +0.31. Adapter-age branch:
+  declines track generation, not complementarity. The Latin-square ρ = −0.57 was carried by
+  generation. Decor veto 0.790 = decor isolated 0.790.
+- Manuscript consequence: the recombination-modifier / reduction-principle reading of Fig. 4B is
+  withdrawn; the declinable merge remains the mechanism that avoided the obligate-merge collapse at no
+  cost against never merging, and the forced-stop control shows a fixed schedule does the same.

tasks/todo.md

@@ -412,3 +412,248 @@ offspring screened on the arm's own signal), QD selection. `src/llm/society.py`
 155 green), `kind: llm_society`, configs `society_smoke.yaml` / `society.yaml`. Stages: smoke
 (local, ~15 min) → pilot full vs no_grounding (GG gate) → 4-arm × 3-seed CX3 campaign → figure +
 manuscript fold-in.
+
+**2026-09-07 — v1 `llm_society` campaign landed (4 seeds) and is NEGATIVE; v2 pre-registered.**
+Best-agent overall at gen 9, 3-seed means: `no_sex` 0.558 ≥ `no_diversity` 0.539 ≥ `full` 0.506 ≫
+`no_grounding` 0.436 (worst arm in every seed from gen 2). Conformity−truth gap does not separate the
+arms. Read through the framework the null was structurally guaranteed (near-clone founders over 3
+families; 2³ competence states; linear blending at 0.5B = the dilution regime; parents truncated
+before breeding, unlike E11's survival-over-pool; `n_test`=40 → SE 0.079) — details and fixes in
+`tasks/prereg-llm-society-v2.md` §1, lesson in `tasks/lessons.md`. Nothing enters the manuscript;
+Fig. 1A's "stated gap" stands. **v2** (`kind: llm_society_v2`): L=12 families / one founder each,
+confidence-routed union inheritance, pooled survival, checkpoint+resume, 240 test items, g=0.85,
+G=12; six numerical hypotheses H1–H6; calibration gates C1–C5 must pass before submission (GG
+reviews). GG decisions: 0.5B; `sex_linear` dropped (H2 deferred); no family vetoes.
+- [x] families / operators / v2 loop / calibration runner / configs / PBS array / figure script / 9 tests (164 green)
+- [x] smoke (4 arms, figure + stats script) → calibration A pass 1 (6/17 in band) → pass 2 (9 in band; C1b needed the gate re-derived 0.35→0.41 from the grid) → GG chose L=9
+- [x] calibration B: C2 FAILED as pre-registered (retention ≤0.81 at k≤150; interference, not the observation floor) → C2b: k=300 + confidence gate τ=0.5 gives mean retention 0.87 (PASS); C3 operator half passes (union holds both families, linear loses one), retention half re-run gated; C5 passes (consensus 0.31)
+- [x] campaign configs set: L=9, k_inherit=300, conf_gate=0.5, epochs=3, n_test 27/family, g=0.85, G=12; PBS 16 elements × 8 h
+- [x] gated cross (C3): two-skill child plateaus at ~0.85×/0.8× of parents at any budget (3 vs 6 epochs; r64 hurts); tight gate τ=0.85 gives the 6-epoch retention at 3 epochs
+- [x] **GG go/no-go (21:30): NO-GO at 0.5B** — the vertical claim needs 5–6 co-resident skills the r=16 adapter cannot hold; today = a measured transmission ceiling (SI material). 7B plan drafted: prereg §13
+- [ ] GG: 7B scope (headline ~80 / H3+H4 ~120 / full ~260 L40S-h), Phase-0 task design go, SI text timing
+- [ ] SI: the 0.5B calibration ceiling as the reason the tier was not run (prereg §11 row 4) — three limits, numbers from results/llm_society_v2_calib_*
+- [ ] REPRODUCING.md: rows for the v1 campaign (4 seeds), v2 smoke, and the 9 calibration bundles
+- [ ] stage code on CX3, `qsub hpc/llm_society_v2.pbs` (16 elements); local hedge = `society_v2_s1.yaml`
+- [ ] `figures/stats_llm_society.py` (per-seed paired contrasts H1/H3/H4, AUC for H5, supplied-vs-retained for H6)
+- [ ] fold the outcome per prereg §11
+
+**2026-09-08 — v3 `llm_compose` run (3 seeds): H1 PASS, H2–H5 null; design could not show the claim.**
+Composition at gen 0 is real and replicated (surplus +0.087/+0.033/+0.093; union-exceedance ~0.12; also
+on MATH-500). Decay hypotheses uninterpretable: the lineages barely drifted (q_math 0.54 → 0.50–0.57)
+and, more fundamentally, a fixed skill set has its ceiling at gen 0 — GG: "are models learning NEW
+skills at EACH generation? … that was not the problem being addressed." v3 was Weismannian (fresh LoRA
+each generation) and retention-only. Two of my errors: C3 unchecked (code specialist 0.075 on MBPP →
+q_code noise), and three premature reads of a single-seed trajectory.
+**v4 `llm_curriculum` — continual learning in a population** (`prereg-llm-society-v4.md`): Lamarckian
+channel (`continue_lora_training`), Latin-square curriculum (complementarity 1.0 → 0.0 by construction,
+H6 predicts the *shape*), arms isolated/society/society_dry/seed_bank (GG's ancestor-merge idea —
+temporal vs spatial complementarity, direction uncommitted), single-shot SoTA baselines at matched budget
+as the falsifier. GG decisions: 3×9×9, replay fixed-total, baselines get the same directed selection.
+- [x] G0 PASS (accumulation 0.74 → 0.95); G2 FAIL ×2 (v2 families don't interfere; one pair at +0.65);
+  G3 negative (merging costs −0.01…−0.13 with nothing to repair); base = 0.094 → one family lifts all to 0.417
+- [x] **v5 curriculum**: 11 real-dataset families, 5 answer shapes, per-family verifiers, disjoint splits
+  (`curriculum_data.py`, +6 tests, 56 green); selection rule fixed in prereg §8a
+- [ ] stage A calibration running (`curriculum-v5-calib`): base + 11 specialists × 11 families
+- [ ] stage B: zero-replay forgetting probe on the survivors (mean drop ≥ 0.15, not single-family)
+- [ ] GG go/no-go → seed 1 local + CX3 array (seeds 2–3), then baselines
+
+**2026-09-08 — v5 curriculum campaign done (3 seeds); all hypotheses fail; the useful finding is a
+split in the theory.** Real-dataset curriculum (6 families, 5 answer formats, per-family verifiers)
+replaced the procedural set. Results: not-merging wins (0.80), merging-with-own-ancestor middling
+(0.66), merging-with-a-peer collapses (0.27), single-shot merging unstable (0.125–0.764). Cause:
+two families are answer-format destroyers that propagate through merges and compound because
+offspring continue the lineage. Scope limits: merging was obligate (no veto) and there is NO
+selection between lineages — a gene-flow experiment, not a selection one.
+**Key new measurement:** same-skill adapters (seed/data draw only) are near-orthogonal in weight
+space (cos +0.006), disagree on 24% of prompts, and **merging them beats the best parent by +0.087,
+exactly at the either-right ceiling**. So decorrelation-in-what-you-know is harmful while
+decorrelation-in-how-you-encode-it is beneficial — the framework's single rho conflates them.
+- [ ] v6: three arms (no-merge · complementary · parallel) + veto + population selection (~30 GPU-h)
+- [ ] decide how the E9-risk result and the two-variations split enter the manuscript (beside Fig. 5A)
+
+**2026-09-08 (later) — four mechanism probes; two of my explanations retracted.** (1) Same-skill
+adapters: 85% of a LoRA's change is run-specific noise; merging two beats the better parent by +0.087,
+at the either-right ceiling. (2) Denoising before crossing adds +0.025 on both skills at once
+(inbred-lines signature). (3) A single merge of clean adapters is PROTECTIVE (0.825 vs 0.550 best
+parent) — retracts "destructive skill propagates through merges". (4) Five chained convex merges lose
+nothing, while signal-preserving additive weights collapse (1.02 vs 0.52 retention) — retracts
+"geometric signal dilution"; the real constraint is bounding drift from the base. (5) Scaling probe
+(GG's control): base 0.000, and the adapter works down to 1/8 then dies — 1/16 = 0.450, 1/32 = 0.000.
+So the chain's apparent retention was ANSWER FORMAT supplied by the dominant partner, not the skill.
+Consistent with Fig. 3C-D: functional conflict predicts merge damage, weight geometry does not.
+- [ ] test the remaining candidate: continued training ON TOP of merged weights (chain + fine-tune each round)
+- [ ] if confirmed, the finding is about output conventions propagating through merges — reframe accordingly
+
+**2026-09-08 (evening) — scaling thresholds measured; bespoke weights tested and NEGATIVE.**
+Per-skill dose-response: cliffs are sharp and skill-specific (boolq dies at 1/8, arc survives to 1/8
+at its BEST score 0.92); 4 of 6 adapters are over-trained and improve when scaled down (mnli
+0.40->0.68 at 1/4). Denoising does NOT move the cliff -> the limit is signal MAGNITUDE, not
+signal-to-noise, so denoising buys quality (+0.025) but not merge depth. Bespoke per-skill weights
+(cliff and optimum variants) both LOSE to plain uniform 1/6 (0.686-0.689 vs 0.708): solo curves don't
+transfer because effective strength is relative, not absolute.
+KEEP: (a) one merged model beats six separate specialists on their own tasks (0.708 vs 0.678);
+(b) attenuating each specialist to its own optimum gives 0.755 with no merging and no retraining.
+- [ ] still untested: continued training ON TOP of merged weights (the last candidate for the v5 collapse)
+- [ ] decide whether the compression trade (0.708 merged vs 0.755 separate) is a paper result or an appendix note
+
+**2026-09-08 (late) — last candidate eliminated; v5 collapse recorded as UNEXPLAINED.**
+merge-then-train beats merge-only on the tracked skill in 4/5 rounds and on the incoming skill in 5/5
+(mnli 0.867 vs 0.467); it even absorbs the round-4 format shock. So training-on-merged-weights is not
+the mechanism — it is the best procedure tested. All three proposed explanations for the v5 collapse
+are now refuted by direct test. Remaining structural difference: v5 merged multi-skill accumulating
+lineages (rank 16, up to 6 skills), these chains merge clean single-skill adapters -> capacity is the
+suspect, but NOT claimed: three guesses have been wrong, a fourth is not earned.
+- [x] veto arm: recommend NOT running — v5 is unreportable regardless (awaiting GG)
+- [ ] GG decision: close the LLM-society file for this paper; keep engineering findings separate
+
+**2026-09-08 (late) — VETO ARM: one bit of selection converts collapse into a healthy trajectory.**
+Seed 1: veto 0.783 vs obligate-merge society 0.211 vs isolated 0.814. Veto rate 67%, and structured:
+1/3 declined at generations 0-2, then 3/3 at generations 3-5 — the population stops merging exactly as
+complementarity falls (1.00 -> 0.80 -> 0.67). GG's caveat is right: once all merges are declined the
+arm IS isolated, and isolated overtakes at gen 4 and finishes higher. Honest claim: recombination pays
+only while partners differ, the population detects when that ends, and still finishes slightly behind
+never merging. Makes the v5 negative reportable (risk + remedy + limit) beside Fig. 5A.
+I had recommended skipping this experiment; that was wrong — I judged it by whether it would rescue a
+written-off conclusion rather than by what it would measure.
+- [ ] CX3 array 4007703 (seeds 2-3) -> confirm the veto rate pattern and the isolated crossover
+- [ ] optional control: forced stop at gen 3, to test whether the veto's TIMING matters
+
+## Manuscript revision — multigenerational LLM population + new literature (2026-09-09)
+
+Plan: `~/.claude/plans/we-are-going-to-cheerful-fog.md` (approved by GG 2026-09-09). Dual-audience
+writing standard is paramount: every term defined at first use with an example from each field.
+
+- [x] Pre-write checks: chance-corrected competence count (claim dropped — single adapters unlock ~4
+      families via shared formats at gen 0; report retention_seen flat ≈0.78 and no first-family erosion
+      instead); Spearman veto-rate vs complementarity ρ=−0.57, p=0.013, n=18; pop-gen citations verified
+- [x] Fig. 6 → five panels (D trajectory, E veto rate vs complementarity); caption; REPRODUCING.md rows
+- [x] main.md: Abstract, Significance, Table 1 row, new Results subsection, society/speciation pointers,
+      Discussion (design rules, CL, borrowed/new, limits, creative diversity, outlook), Methods
+- [x] si.md: S3 text, Table S1/S2 rows, M2/M5/M6 additions, SI figures list; fixed two stale SI
+      citation numbers (41→44, 43→46 pre-renumbering) and one leftover "honest"
+- [x] References: +8 (73–80 appended, then renumbered to first-appearance order by
+      `paper/pnas/renumber_refs.py`; 80 refs, 0 orphans, recheck = 0 renumbered)
+- [x] Verification: fig6 rendered+inspected twice (legend fix); PDFs build (main 24 pp, SI 11 pp; no
+      unresolved FIG markers); gap/meta-language grep clean; two-reader pass (added "verifier",
+      "frozen", validation glosses); `make test` 196 passed
+- [x] **Compression pass (GG directive 2026-09-09).** 7,318 → 6,764 total, of which 6,520 is running
+      prose and 244 is the Table 1 grid (PNAS counts tables separately). −554 words with no content
+      removed: sentence-level density throughout, one genuine de-duplication (the MNIST collapse
+      figure was stated twice, in the biological-model section and again under Grounding — kept the
+      Grounding statement, which carries the 2× estimator-bias comparison), and two detail blocks
+      moved to where they belong (predictive-test per-seed ρ ranges → new Table S2 row; Methods
+      pointer to SI Methods). PDF 24 → 23 pp. Every number, citation, hedge, and gloss retained.
+      Further cuts would need structural calls: moving the blending-inheritance Proposition to SI
+      (~130 words, but it is a flagship claim) or trimming review-calibrated hedges — left for GG.
+- [x] **Fig. 1A updated (GG, 2026-09-09).** The composed-society × language-model cell was rendering
+      "open — the stated gap"; it now carries the result ("6 generations × 3 lineages: obligate merging
+      collapses, a declinable merge tracks partner complementarity") with tag Fig. 6D–E, and the
+      biological-model cell's tag narrowed to Fig. 6A–C. Tier header corrected to "Qwen 0.5B, 1.5B &
+      7B; exact-match and execution verifiers". Dead `OPEN` rendering branch removed. Caption in
+      build.py no longer ends on the gap clause. Repo-wide grep for gap language now clean.
+- [x] **Zotero library built (GG, 2026-09-10).** All 80 references resolved to authoritative metadata
+      via doi.org content negotiation: 77 from DOI (53 printed in the manuscript, 22 found by
+      title-matched Crossref search, 2 hand-verified — Brinkmann *Machine culture*, Schwarz *Progress &
+      Compress*), 3 hand-written because they predate DOIs (Jenkin 1867, Fisher 1930, Templeton 1986).
+      Artifacts in `paper/pnas/refs/`; generator `paper/pnas/build_zotero_library.py`.
+      **Not yet in Zotero** — the app is closed and its library lives in ownCloud; direct writes to
+      `zotero.sqlite` are unsafe, so import is one step in the Zotero UI (see refs/README.md).
+- [ ] Optional: sync long-form `paper/the-evolution-of-sex-for-ai.md` L797 ("LLM society is unbuilt")
+
+## Manuscript round 4 — research-paper restructure (GG feedback 2026-09-10)
+
+Plan: `~/.claude/plans/we-are-going-to-cheerful-fog.md`. Diagnosis: mean sentence 49 w vs GG's own
+31 w, 50% of sentences over 40 w, em-dashes 11.4/1k vs his 0.57 — long sentences in short paragraphs,
+the inverse of his rhythm. That is the measurable cause of "too cryptic".
+
+- [x] Phase 1 — Results restructured to question+design / result / implication; seven descriptive
+      section titles; grounding leads with the novel per-item floor and cites the g≈0.05 threshold as
+      corroboration of published values; Proposition lifted into its own block; Recombination split by
+      experiment; novelty of Fisher–Muller-in-LoRA conceded in place
+- [x] Phase 2 — Main figures 7 → 5. Old Fig. 4 (E4/E8) and Fig. 5 (E9/E10/E14) dissolved; E9/E10/E14
+      to SI as established results with no real-model counterpart. Panels reordered so the real-model
+      result leads and the inheritance model follows as reference (Fig. 2A/B, 4A–B before 4C–E,
+      5A–D before 5E–F). Fig. 1A column relabelled "Inheritance model (reference)"; tags repointed.
+      "biological model" → "inheritance model" throughout.
+- [x] Phase 3 — Prose to the measured fingerprint: mean sentence 49.0 → 31.4 w (GG's own 31.2),
+      >40-word sentences 50% → 22.6% (his 20.8), em-dashes 11.4 → 3.42/1k (his 0.57), semicolons
+      13.6 → 8.6, colons 13.6 → 8.4, antithesis 1.77 → 1.81/1k after re-cutting the ones the rewrite
+      introduced. 21 pp (from 23).
+- [x] Phase 4 — Discussion rebalanced: the 476-word (68 w/sentence) continual-learning block and the
+      242-word (80 w/sentence) borrowed/new block broken into paragraphs of 5–6 sentences.
+- [ ] Remaining: two-reader accessibility pass over the rewritten sections; `Fig. 2` cross-reference
+      in the inheritance-model section may want to be `Fig. 2A`; consider whether the Significance
+      statement and Abstract need to match the new section titles.
+
+## Manuscript review pass (2026-09-11)
+
+Review of `paper/pnas/main.md` (novelty, accessibility, calibration, cheap experiments); corrections applied:
+- [x] Abstract rewritten (one idea per sentence, jargon removed, 250 words); own-ancestor result added, mating-breadth hypothesis dropped
+- [x] Own-ancestor (seed-bank) merge given its own paragraph, Table 1 row, and design rule
+- [x] Emergent null (merge rescues forgetting specialists) and the overlap control (delta-cosine +0.60 → +0.03) promoted from asides to findings
+- [x] "Five specific results" recut to four; grounding floor named a corollary, ablation named a demonstration (conformity builds grounding in)
+- [x] Latin-square collinearity of complementarity and generation stated explicitly in Results
+- [x] Two SI-only design rules marked as inheritance-model predictions; 7B Fisher–Muller marked single run
+- [x] Terms defined at first use: forward KL, BDM, TIES, linear-mode-connectivity barrier, low-rank factor space, oracle parent potential
+- [x] 70-word speciation sentence split; Fig. 5 E–F, Fig. 3 C–D, Fig. 4C–E cross-refs added; stale "Fig. 6D–E" in SI Table S1 → Fig. 4A–B
+- [x] Author email fixed; PDF rebuilt (22 pp)
+- [ ] Cheap experiments proposed, none run: forced-stop-at-gen-3 control; non-Latin-square curriculum breaking the complementarity/generation confound; seeds 2–3 for the single 7B runs; pre-merge disagreement vs realised penalty on the existing population checkpoints; withholding curriculum; stylistic-diversity readout on saved generations; E11 with alternative selection schemes
+- [x] Compression/accessibility pass (2026-09-11): main-text prose 6,902 → 6,117 words (−11%); em-dashes 15 → 0; antithesis 0.33/1k; all 81 citations, 5 figure markers and every headline number verified present by script; PDF 22 → 21 pp. Pre-pass copy kept in session scratchpad only.
+
+## Experiments 1–3 from the manuscript review (2026-09-11) — plan `~/.claude/plans/atomic-rolling-sprout.md`
+
+- [x] `merge_until` (forced stop) and `orders` (custom curriculum) keys in `src/llm/curriculum.py`; manifest records them; +3 tests (127 green)
+- [x] configs `curriculum_v5_stop3.yaml`, `curriculum_v5_decor.yaml` (complementarity 0.00/0.67/0.70/0.58/0.33/0.00 verified); prereg §8h written before running
+- [x] PBS: `hpc/llm_curriculum_controls.pbs` (seeds 2–3 × {stop3, decor}), `hpc/llm_7b_seeds.pbs` (seeds 2–3, merge → moe_hard → directed_hard)
+- [x] 7B seed-1 bundles moved to `results/llm_*_hpc/s1/`; `load_seed_bundles` in `_figlib`; fig3 B, `plot_llm_{merge,moe,directed,seeds}.py` seed-aware (no more `.iloc[0]`)
+- [x] `figures/stats_llm_curriculum.py` (shared loader, now used by `make_figs._load_curriculum`; contrasts; partial-correlation test) and `figures/stats_llm_7b_seeds.py`; both reproduce the published numbers on existing bundles
+- [x] **Experiment 1 decided (3 seeds):** forced stop 0.793 vs veto 0.792 vs isolated 0.796 vs society 0.269; veto − stop3 = −0.008/−0.006/+0.011 (all within the pre-registered ±0.03). Reading: the declinable merge's outcome is explained by *when* it stopped; the "evaluation adds value beyond timing" reading is dropped. Fig. 4A carries the dashed control; `results/llm_curriculum_v5_stop3/README.md`
+- [x] **Experiment 2 decided (3 seeds):** partial ρ(declined, complementarity | generation) = −0.07 (CI −0.21…+0.09); partial ρ with generation = +0.31. Declines track generation, not complementarity; the modifier/reduction-principle reading is withdrawn. Decor veto 0.790 = decor isolated 0.790. `results/llm_curriculum_v5_decor/README.md`; Fig. 4B now shows both curricula
+- [x] **Experiment 3 done (7B, seeds 1–3, 33 min/seed on one L40S):** merge − best specialist +0.066 ± 0.036 (3/3); routing − soup +0.094 ± 0.015 (3/3); directed − soup +0.073 ± 0.031 (3/3). Not replicated: 'soup below best specialist on hard' (1/3; mean +0.001) — sentence softened in main text and caption. Fig. 3B now mean ± CI; READMEs carry per-seed tables
+- [ ] GG: `ssh -fN hpc`; then rsync code, `qsub hpc/llm_curriculum_controls.pbs` and `qsub hpc/llm_7b_seeds.pbs`
+- [ ] after data: fig4 (stop3 line; decor decline curve), captions in `build.py`, main/SI/REPRODUCING/READMEs/CLAUDE.md numbers from the stats scripts only
+- Discovered: the venv carried paths from before the repo moved into `LLMs/` (stale shebangs; `uv run pytest` could not spawn). `pytest` re-installed; other console scripts still stale — `uv sync --all-extras --reinstall` would fix all. Hardening candidate: specialist cache key lacks the base model (fails loudly, not silently).
+
+## Venue + novelty audit (2026-09-11)
+Target: Nature Machine Intelligence first; PLOS Comput Biol as the venue reaching both ML and pop-gen readers. All PNAS wording removed from `paper/pnas/` sources (SI Appendix → Supplementary Information; build/tex comments). Directory name `paper/pnas/` kept (Makefile/REPRODUCING paths); Significance statement kept pending GG decision.
+Literature audit (three WebSearch sweeps) found claims that need rewording/citations before submission:
+- [x] "Every merging study merges once" is false → narrow to "no study combines per-generation skill acquisition with repeated, optional merging across lineages". Cite iterated-merging work: model kinship 2410.12613 (stagnation by gen 2, inbreeding analogy), GENOME 2503.01155, M2N2, TIME 2412.06712, MagMax, ACMap 2412.18219 (early-stop precedent), K-Merge 2510.13537 (similarity-gated merge), SFA/"Soup to go" 2501.05559 + IMM 2503.02103 (ancestor-averaging precedent)
+- [x] Predictor section: "functional > weight geometry" is already shown by Cao 2603.09463 (must-cite), Zhu 2608.09490, Zhou 2601.22285 (gradient > cosine). Reframe novelty as held-out predictive design + the overlap control (cosine = shared-data artefact; not found anywhere)
+- [x] Speciation: credit permutation+rescaling decomposition to Git Re-Basin + REPAIR 2211.08403; cite ZipIt 2305.03053, Sharma non-local 2410.12766 for residual barriers; Git Re-Basin §5.4 already merges complementary-class parents. Keep as new: conflicting-label manipulation, three-arm contrast, emergent null (against Pari 2411.02207 / Horoi / Kozodoi)
+- [x] Grounding: must cite Alemohammad 2307.01850 (fresh-data loop fixed point), Bertrand 2310.00429 (stability theorem in real fraction), Dohmatob 2402.07043 + 2410.04840 (counter-claim: any synthetic fraction caps performance — reconcile with H_eq<H*), Kazdan 2410.16713 (cardinality not proportion — supports Pred. 4), Suresh 2412.17646 (per-item no-immigration law), Garg 2509.22341 / He 2502.18049 (fresh-data optimal ratio ≈0.62 under MSE — explain the different objective); Shumailov's 10%-retention datum
+- [x] Blending proposition: present as lemma (linearity + Poisson thinning); cite Yuan 2601.13572 (signal dilution), Malinin 2020 ensemble-distribution distillation, BTM/BTX, Bulmer 2004 for Jenkin/Fisher; Fisher–Muller-for-merging framing appears to be ours
+All five applied to main.md (2026-09-11): 19 references added (now 100), renumbered by first appearance, PDFs rebuilt. Not yet done: regenerate `paper/pnas/refs/` exports (Zotero/RIS/CSL) for the new entries; confirm Bertrand's λ convention and Alemohammad's fixed-point statement against the full texts before submission.
+
+
+## Manuscript review pass (2026-09-12)
+- [x] Act on the 45 comments in `paper/pnas/main_with_comments.odt` (clarity, nomenclature, heralds).
+- [x] Number Supplementary Figures S1–S13 (`paper/pnas/si_figures.py`, `build.py`, `si.tex` counter) and cite them from the main text.
+- [x] SI Text S4: proof of the blending-inheritance proposition (regime corrected to `n·p ≪ 1`).
+- [x] Clarity pass on the final Results section (predictive test), unprompted per GG's note.
+- [ ] **Discovered:** SI figure PDFs still carry codename suptitles ("E2 —", "grounding —") and teacher/pupil axis labels (Fig. S8); regenerate with manuscript vocabulary before submission (`figures/plot_*.py` title lines or a `--paper` flag).
+- [ ] **Discovered:** Fig. S4 caption quotes `g* = 0.048` as "95% of H*" while the main text says "95% of the source's diversity"; both are the same quantity, but Fig. 2B's caption should use identical wording.
+
+- [x] Round 2 (14 comments): novelty attribution in the grounding section, budget defined, six dataset references (renumbered), Discussion restructured (three theories of heredity; recombination bought speed not level; open problems only).
+- [ ] **Decision (GG):** experiments that would let the dropped "Limits" stand as results, not caveats: (i) seeds 2–3 for the LLM speciation tier (Fig. 5C–D is single-seed; ~1 h L40S); (ii) a curriculum that decouples adapter age from conflict arrival (conflicting families first vs last); (iii) a second base lineage (SmolLM2/Llama) for one LLM experiment; (iv) the six-generation population with culling (differential reproduction).
+
+## Four experiments from the dropped Limits (2026-09-12; plan ~/.claude/plans/cozy-nibbling-crayon.md)
+- [x] Code: seed-specific speciation adapter root; `cull_step`/`inherit_slot` + `cull:` in curriculum; manifest key; 5 pure tests (204 green).
+- [x] Configs: curriculum_v5_{early,late,early_obl,late_obl,cull}, merge_seeds_smol, moe_hard_seeds_smol.
+- [x] PBS: llm_speciation_seeds (2), llm_curriculum_timing (12), llm_cull (3) submitted 2026-09-12 20:0x (jobs 4035393-5); llm_smol pending the local smoke gate.
+- [x] Analysis code: stats_llm_curriculum (RELABEL, conflict-timing test, cull contrasts), stats_llm_speciation_seeds, stats_llm_smol, plot_curriculum_timing, plot_curriculum_cull, plot_llm_smol; fig5 C-D multi-seed; seed-1 speciation moved to s1/.
+- [ ] Local SmolLM2 smoke gate → submit hpc/llm_smol.pbs.
+- [x] Speciation seeds 2–3 fetched; README, Fig. 5C–D (CI bands), caption, Table S2, REPRODUCING updated.
+- [x] Timing (12 elements) and SmolLM2 bundles fetched; READMEs, Table S2, M5, M2, REPRODUCING, S14 + S16, main-text paragraphs written.
+- [x] Culling: 3 seeds fetched; README, S15, Results paragraph, Discussion rewritten (prediction withdrawn), Abstract, Table S2, M2, M5.
+- [x] SI figures S14-S16; Results/SI text; Table S2 rows; REPRODUCING.md; Fig. 5 caption; Discussion rewritten.
+- [ ] **Discovered:** a curriculum in which some skills are obtainable only by merging (not delivered to every lineage) is the experiment that would separate the LLM population from the inheritance-model society; not run.
+- [x] Student-level figure guide: `paper/pnas/figure_legends_for_students.md` (+ `build_lay_legends.py`, built by `make paper`); 21 legends, glossary.
+- [x] Figures made self-explanatory (2026-09-13): headlines on every data panel; Fig. 3 gains a schematic panel A (models compared), paired-t brackets on B/C, grouped predictors in E; Fig. 4 gains an explainer strip A; clearer legends in Figs. 2 and 5; all five captions rewritten at the midway register; panel letters renumbered in text, SI, figure map and student guide.
+- [x] SI figures S1–S16 lettered (shared `letter_axes` helper in `_figlib`, called in every plot script); captions re-lettered.
+- [x] SI figures brought to the main-figure standard: suptitles and codenames removed from all 16 plot scripts, panels lettered, captions rewritten in the main-figure format, appendix legends re-lettered.
+
+## 2026-09-13 — clarity pass on main text
+- [x] Fig. S1/S2 mis-citation fixed; ratchet paragraph split and explained; model section moved under Results
+- [x] Stationary-diversity paragraph rewritten around the closed form (count-not-fraction; island model / F_ST; one-migrant rule; Souly et al. poisoning as ref 62)
+- [x] Full clarity audit (36 items, tasks/clarity-audit-2026-09-13.md) applied in all three tiers; PDFs rebuilt; citation order verified
+- [ ] GG read-through of the rewritten passages