Phase 3: LLM-tier speciation + multi-seed firm-up of the recombination claims

llm_speciation (new kind; src/llm/speciation.py): E13 in LLM weights.
LoRA children share the frozen base's coordinates, so merge failure is
functional by construction. CONFLICT (ambiguous sort prompts learned
under opposite conventions — the BDM structure): function-specific
hybrid breakdown — merged coherence 0.02-0.08 falls below BOTH parents
(~0.2) on the conflicted function; and in the de-confounded `add` design
(private budget fixed, conflict added on top; 3 seeds after a
single-seed pilot showed one anomalous point) the merge's private-family
accuracy shows NO trend with conflict — the damage is surgical, not
global. DURATION (over-trained disjoint specialists, 1->12 epochs): the
merge improves (0.84->0.94) and stays above the best parent — the MLP
"no emergent isolation" null generalises; relevant to the
expert-training-duration report (2607.11997), with the epistasis
prediction left to the decisive experiment.

Multi-seed firm-up (seeds threaded into specialist caches; `seeds:` list
support in the runner; fixed test sets): all three recombination claims
hold with CIs — merges beat every specialist (5 seeds, ties
0.647±0.027 > best spec 0.592±0.009; worst-family 0.28 vs <=0.16); union
0.274±0.026 > fusion 0.174±0.102 on hard (3 seeds); directed 0.221±0.026
> soup. NEW finding: fusion is seed-FRAGILE where headroom exists
(CI ±0.10) while routing/directed selection are stable (±0.026) — the
union/selection operators win on reliability, not just mean.

Figures (llm_speciation 3-panel; llm_seeds 3-panel with 95% CI), READMEs,
+1 convention test (150 green), make llm-speciation / llm-seeds targets.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BkRLcc18rwT2Lysu6PbG7v

configs/llm/directed_hard_seeds.yaml

@@ -0,0 +1,22 @@
+experiment: llm_directed_hard_seeds
+kind: llm_directed
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# Multi-seed 0.5B directed sex on the HARD benchmark (offspring selection pays off exactly where the
+# default soup is suboptimal). Fixed test/val sets; training + Dirichlet-offspring seed varies; reuses
+# the per-seed hard specialists trained by llm_moe_hard_seeds when present.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+hard: true
+families: [lists, strings, arith]
+n_train: 400
+n_test: 80
+n_val: 60
+n_candidates: 16
+concentration: 0.5
+epochs: 3
+lora: {r: 16, alpha: 32}
+
+output: {dir: results/llm_directed_hard_seeds}

configs/llm/merge_seeds.yaml

@@ -0,0 +1,20 @@
+experiment: llm_merge_seeds
+kind: llm_merge
+seed: 1
+seeds: [1, 2, 3, 4, 5]
+n_replicates: 1
+
+# Multi-seed firm-up of the 0.5B merge experiment (PNAS work order Phase 3: removes the "one seed"
+# objection). Same protocol as configs/llm/merge.yaml; the test sets stay FIXED (seed 1000+i per
+# family) while the training seed varies, so across-seed variance is training variance only.
+# Specialists cache per-seed (spec_<family>_s<seed>).
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+families: [lists, strings, arith]
+n_train: 600
+n_test: 100
+epochs: 3
+lora: {r: 16, alpha: 32}
+merges: [soup, ties]
+
+output: {dir: results/llm_merge_seeds}

configs/llm/moe_hard_seeds.yaml

@@ -0,0 +1,21 @@
+experiment: llm_moe_hard_seeds
+kind: llm_moe
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# Multi-seed 0.5B union-vs-fusion on the HARD benchmark (the headroom regime where the ordering
+# matters). Fixed test sets; training seed varies; hard specialists cache per-seed
+# (spec_<family>_hard_s<seed>). Companion to the single-seed llm_moe_hard and the 7B HPC runs.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+hard: true
+families: [lists, strings, arith]
+n_train: 400
+n_test: 80
+n_route: 32
+epochs: 3
+lora: {r: 16, alpha: 32}
+operators: [soup, ties, moe_oracle, moe_learned]
+
+output: {dir: results/llm_moe_hard_seeds}

configs/llm/speciation.yaml

@@ -0,0 +1,32 @@
+experiment: llm_speciation
+kind: llm_speciation
+seed: 1
+n_replicates: 1
+
+# LLM-tier model speciation (E13 in language-model weights; PNAS work order Phase 3). Two LoRA
+# children from the same frozen base — so there is NO permutation/rescaling ambiguity by construction:
+# any merge failure is functional incompatibility, isolated architecturally. Two sweeps:
+#   conflict_fracs — the IMPOSED cliff: each child has a private disjoint family (A: strings,
+#     B: arith) plus a shared set of AMBIGUOUS sort prompts ("Sort the list [...]", no direction)
+#     answered ascending by A and descending by B (each convention harmless alone, contradictory
+#     jointly — the BDM structure). Prediction: the MERGED model's private-family competence degrades
+#     and its convention coherence (max of asc/desc grading) collapses as conflict grows, while each
+#     PARENT stays fine — hybrid breakdown in verifier units, echoing the MLP cliff.
+#   durations — the EMERGENT null: pure disjoint specialists over-trained (epochs swept), zero shared
+#     data. Arbitrates the MLP tier's null (no emergent isolation; the merge rescued specialists at
+#     every divergence) against the empirical report that averaging prefers under-trained experts
+#     (arXiv:2607.11997). Pre-registered readings: merged quality falls with duration while parents'
+#     own-family quality holds -> emergent incompatibility at the LLM tier; otherwise the null
+#     generalises. Either outcome is reportable; do not tune toward one.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+family_a: strings
+family_b: arith
+n_train: 400
+n_test: 100
+epochs: 3
+lora: {r: 16, alpha: 32}
+conflict_fracs: [0.0, 0.25, 0.5, 0.75, 1.0]
+durations: [1, 3, 6, 12]
+
+output: {dir: results/llm_speciation}

configs/llm/speciation_add.yaml

@@ -0,0 +1,30 @@
+experiment: llm_speciation_add
+kind: llm_speciation
+seed: 1
+seeds: [1, 2, 3]
+n_replicates: 1
+
+# The de-confounded conflict sweep (companion to configs/llm/speciation.yaml). The "replace" design
+# holds TOTAL training fixed, so the merge's private-family decline with conflict_frac is confounded
+# with shrinking private data (coherence is the clean readout there). Here conflict_mode: add holds
+# each child's PRIVATE training fixed at n_train and ADDS conflict data on top, so any decline in the
+# MERGE's private-family accuracy relative to its parents is interference from the conflicting
+# convention, not a data-budget artefact. Pre-registered readings: merged private accuracy tracks the
+# parents at every frac -> conflict damage is localised to the conflicted function (function-specific
+# hybrid breakdown); merged private accuracy falls below the parents as frac rises -> the conflict
+# corrupts shared circuitry beyond the conflicted function (global hybrid breakdown). Run at 3 seeds:
+# the single-seed pilot showed one anomalous grid point (frac=0.75, a bad parent-B run), so per-seed
+# replication is required before reading the curve.
+
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+family_a: strings
+family_b: arith
+n_train: 400
+n_test: 100
+epochs: 3
+lora: {r: 16, alpha: 32}
+conflict_mode: add
+conflict_fracs: [0.0, 0.25, 0.5, 0.75, 1.0]
+durations: []
+
+output: {dir: results/llm_speciation_add}