DIVERGE-PL1: Branch-Local Verified Plasticity Result

Status: oracle-typed mechanics gate passed. This is not a raw-language, Shohin, or continuation-pretraining result.

Result

The source-separated board contains 2,048 train episodes, 256 development episodes, and five unopened confirmation seeds of 256 episodes each. Every episode uses a new permutation of eight opaque symbols over eight noncommuting transforms in Z/97Z, twelve acquisition attempts with eight complete branches each, and sixteen context-free transfer programs at depths 12--20. Alias and program overlap is zero across every split pair.

On 1,280 confirmation episodes and 20,480 transfer programs:

PL1 transfer rates across the five fixed seeds are 85.205%, 85.107%, 88.818%, 83.984%, and 89.648%. The weakest seed remains above the frozen 80% floor. The paired 95% bootstrap lower bound against the strongest baseline, TRANSIENT_GRAD, is 81.948 points.

The assessor-only context-free transfer probe rises from 7/20,480 after attempt 1 to 17,726/20,480 after attempt 12. Text, demonstrations, branch transcripts, and verifier messages are absent at transfer; only the 64-scalar session policy matrix remains.

Causal controls

• plastic reset: 1/20,480;
• shuffled branch receipts: 485/20,480;
• wrong-branch receipts: 3/20,480, with 122,296 mismatched credits rejected;
• unrelated episode-state transplant: 3/20,480;
• no eligibility localization: 70/20,480;
• poison changed behavior in 1,280/1,280, then rollback restored exact pre-poison hashes and outputs in 1,280/1,280;
• protected-owner mutation injection failed closed in 5/5 seeds; and
• normal PL1 runs had zero rejected credits or protected-owner hash changes.

Removing the write cap and score clip is exactly tied with PL1 at 17,726/20,480. Homeostasis is therefore not part of the causal performance claim; it remains only a safety envelope.

Every frozen condition passes. The decisive factor is not generic mutable state: verified first-error receipts plus branch-local eligibility allow the policy to accumulate correct partial assignments across failed complete hypotheses. Whole-branch retention, scalar fast weights, and full-branch transient gradients cannot localize that credit.

Claim boundary and decision

PL1 v0 has zero learned parameters and uses REFERENT_ORACLE, an exact typed executor, and a verifier that identifies the first invalid transition. It proves a branch-local policy-update mechanism under deterministic feedback; it does not prove that a language model can compile the symbols, produce the certificates, or solve open-domain problems.

SRP1 failed, so raw-language PL1 remains blocked. Preserve this mechanism and return to the semantic source/referent interface. Only after a structurally different compiler qualifies may one natural PL1 integration run. Do not run PL1 rank, width, branch-count, attempt-count, seed, budget, or duration variants.

Receipts

• implementation/preregistration commit: 33e6a7b;
• data report SHA-256: 738b60f8c837fa479a9bb1cf82b11ff06441280a50f490f2d3155e82e24f3058;
• evaluation SHA-256: 0cc2a24267c38396e166dbce7e540aaba87c5acf2474ede26be0939cfe9fd5fb.