Giorgio Gilestro
6f8cef1ac5
Removed from main (all preserved on the dev branch): the arXiv build and
its sources, design documents (blueprint, results summary, review responses,
essay drafts), tasks/ and CLAUDE.md, the cover letter and reference tooling,
two unused manuscript figures, and every experiment that feeds no figure or
number in the paper: the collapse null, the sexual-vs-asexual lineage, the
NK speciation variant, the 0.5B single-seed LLM prototypes, the compose and
society experiments with their calibration and pilot runs, and their
configs, runners, tests, figure scripts and PBS jobs. Their result bundles
are moved to results/_archive/ (ignored) so the parquets stay on disk.
Also: plot_llm_speciation reads the s{seed}/ layout; the mating-breadth
plot writes under its bundle name; Makefile targets reduced to the kept
experiments; REPRODUCING.md and README point to dev for the rest.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Y64o8FKP7rCuXzC48pxpMm
47 lines
3.9 KiB
Markdown
47 lines
3.9 KiB
Markdown
# E12 — Model speciation: when two diverged models are too incompatible to merge
|
||
|
||
**Claim tested.** The sexual society (E7–E11) recombines complementary parents. E12 asks the limit:
|
||
**how far can two lineages diverge before recombination (model merging) stops working?** In biology the
|
||
answer is *reproductive isolation* via **Bateson–Dobzhansky–Muller incompatibilities** (BDMIs) — alleles
|
||
benign on their own lineage's background but deleterious *in combination*, which a recombinant inherits
|
||
untested. A merged model is a single recombinant (an F2-like *hybrid-breakdown / recombination-load*
|
||
object, not an F1), so the predicted signature as parental divergence grows is
|
||
**compatible → outbreeding depression → hybrid inviability**, arriving earlier the more epistatic the
|
||
capability landscape.
|
||
|
||
**Setup.** Pure seeded NumPy on the E7–E11 genotype machinery (bitwise-reproducible; no external
|
||
simulator, whose separate RNG would break that guarantee). Two landscapes:
|
||
- **BDM** (`configs/inheritance/fig5_speciation_bdm.yaml`, headline): an ancestor; two lineages each substitute a *disjoint*
|
||
set of loci (each parent adaptive, neither carrying an incompatibility); a fraction `ρ` of
|
||
cross-lineage locus pairs are BDMIs (penalty `s`), biting only when a hybrid inherits *both* derived
|
||
alleles. Sweep divergence `d` (substitutions) for several `ρ`; `L=20`, 15 reps.
|
||
- **NK** (`speciation_bdm_nk`, on the dev branch): parents are *local optima* (hill-climbed) on a Kauffman NK
|
||
landscape; sweep ruggedness `K`. The emergent version.
|
||
|
||
### Results
|
||
- **The three-regime collapse (BDM).** Parent fitness rises linearly with divergence; hybrid fitness
|
||
*tracks it while compatible, then peels off, peaks, and crashes*. At dense epistasis (`ρ=0.5`) hybrids
|
||
peak near `d≈8` and fall to **−1.0** by `d=20` (below the ancestor = inviable); at sparse epistasis
|
||
(`ρ=0.1`) there is mild outbreeding depression and **no** isolation.
|
||
- **The isolation cliff moves with epistasis density.** Reproductive-isolation rate (P hybrid inviable)
|
||
at `d=20`: `ρ=0.1`→0.00, `ρ=0.25`→0.03, `ρ=0.5`→**0.50** — the cliff arrives at lower divergence the
|
||
denser the epistasis.
|
||
- **The Orr–Turelli snowball.** The number of incompatibilities grows ~`(d/2)²` (≈48 at `d=20`, `ρ=0.5`
|
||
≈ `0.5·10²`), so hybrid fitness falls *super-linearly* — divergence is punished faster than it accrues.
|
||
- **The epistasis wedge (NK).** At `K=0` (additive) recombination is neutral (no isolation — and the two
|
||
parents can't even diverge, since there is one peak); as ruggedness rises, recombining two adapted
|
||
local-optimum parents flips from a gain to **outbreeding depression** (recombination gain 0 → −0.13;
|
||
OD rate 0 → 0.90 across `K=0→10`). *At matched divergence, mergeability is governed by epistasis* —
|
||
the axis no divergence-only merge predictor captures.
|
||
|
||
### Why it matters / positioning
|
||
The ML *phenomenon* that "specialization/divergence eventually breaks merging" is known empirically
|
||
(Pari et al. 2024; Zhou et al. 2026), and part of the apparent incompatibility is a permutation artefact
|
||
(Git Re-Basin). E12's contribution is the **predictive theory** those lack: the functional form
|
||
(compatible→OD→inviability), the **snowball** onset, and the **epistasis wedge** — merge failure as a
|
||
Dobzhansky–Muller phenomenon whose onset is set by divergence *and* epistasis, not divergence alone. The
|
||
design rule: *before merging, check divergence against the landscape's ruggedness; beyond the cliff,
|
||
route (allopatry), don't merge.* **Falsifier (not triggered):** no OD/isolation progression as `d` and
|
||
`ρ` grow — instead the full progression appears, and the additive control shows none. Real-weight
|
||
confirmation (merging at increasing divergence *with* permutation alignment, isolating the residual
|
||
epistatic incompatibility) is the flagged next step; here the analytic model is the anchor.
|