research: completed AM power-weight compiler

[mountain]

Closes #146.

Result

Implements and evaluates a research-local, rank-one rational completed AM power–weight compiler with semantic carrier

$$\Phi_{\nu,w}=a^\nu e^{(w-\nu)v},\qquad M\Phi_{\nu,w}=w\Phi_{\nu,w},\qquad A\Phi_{\nu,w}=\nu\Phi_{\nu-1,w-1}.$$

Jets and matrices remain observer products; they are not identified with the carrier.

Frozen semantics

• native product, A, M, PBW, and finite affine law;
• paired power-weight / power-character canonicalization;
• exact readout coefficients in $\mathbb Q[\exp(\mathbb Q)]$;
• positive-cone observer-directed add, multiply, scale, shift, exp, and log1p;
• typed log-a and v-jordan resonances;
• compact deterministic certificate and full semantic replay;
• strict resource budgets and fail-closed taxonomy;
• no generic expansion/limit oracle in evaluator or replay.

Commit–reveal

• preimplementation commitment: d578b7ed9f5193cc5b0a0212c3edad025a3b2cee4cf3f90885dd95a412ce5597;
• pre-reveal evaluator freeze: a4cd0a7918b4d79ea049852bb829e0feb92d5e2f;
• hidden payload verified after source freeze;
• held-out exact readout: 1/3;
• held-out replay: verified;
• 29 dependency requests over 8 weights, 111 exact coefficient operations, 888-byte certificate.

Evidence

• frozen public gate: 12/12 passed and replayed;
• workstream tests: 10/10 passed;
• root test suite before reveal: 750 passed, 25 skipped, 0 failed;
• both pre-reveal and post-reveal manifests verify;
• same-information SymPy baseline agrees on every exact readout.

The weight-32 cancellation probe reads 2 distinct weights versus the baseline's declared 33-weight materialized window. P3 and the held-out use the same distinct-weight span as the baseline. Wall-time measurements are non-authoritative.

Disposition

EXPAND, within the frozen claim ceiling: exact observer-directed coefficient semantics with replay for this bounded AM fragment. No general performance, multivariable, general function-theory, surreal-runtime, or Public API claim.

Detailed interpretation: sonnet/am-power-weight-compiler/01-results.md.

[mountain]

Pre-reveal evaluator source freeze

The evaluator and replay source are now remotely frozen at commit

a4cd0a7918b4d79ea049852bb829e0feb92d5e2f

with tree

b8bece048e6f1803afbb328be72413fb08fee159.

Checks before reveal:

• frozen public corpus: 12/12 passed and replayed;
• workstream tests: 7/7 passed;
• root repository: 750 passed, 25 skipped, 0 failed;
• pre-reveal source manifest verifies;
• evaluator/replay oracle firewall clean.

The held-out nonce and payload have not yet been disclosed. The next commit may reveal and execute that already committed payload; frozen grammar, budgets, scoring, evaluator, and replay remain unchanged.

[mountain]

Commit–reveal gate complete

• revealed payload digest matches the preimplementation commitment;
• held-out exact readout: 1/3;
• replay: verified;
• held-out dependency slice: 29 requests over weights 0–7;
• held-out exact coefficient operations: 111;
• certificate: 888 bytes;
• public gate: 12/12;
• workstream tests: 10/10;
• disposition: EXPAND, subject to the documented claim ceiling.

The same-information SymPy baseline agrees on all exact values. Only the sparse weight-32 probe demonstrates a distinct-weight window advantage (2 versus 33); the nested and held-out probes use the same weight span. Timing remains non-authoritative.

The evaluator/replay files still match the pre-reveal source manifest byte for byte.