docs: restate compression claims to held-out truth; retire ≥3× Abrash gate

[madmax983]

Before / After

Before: the repo headlined a ≥3× compression gate (the "Abrash rule") as MET. The measured-results table led with in-sample figures (T_v0 3.72×, tier-2 3.87×/3.92×) and stated the gate was cleared, presenting compression as the shipping thesis.

After: we lead with the blind held-out (sealed) number, 1.67×, and state plainly that the ≥3× gate FAILED out-of-sample — the 3.7–3.9× figures held only on dev corpora that co-evolved with the primitives (an in-sample benchmark-fitting artifact). Corroborated by ~1.7× on the 1,145-shape generated distribution and ~1.03× vs fair terse Python. The thesis is restated around what actually generalizes, all measured out-of-sample:

• Reliability — 100% pass@5 agent writability on unseen tasks.
• Per-solution economics — CSPM 2.124× held-out (correct solutions per million tokens, MTL/Python), widening with difficulty.
• Verified confinement — a machine-checked core (5 proof roots) plus capability confinement.
• Cheap cold start — a 487-token preamble.

Compression is retained as a niche property (2–4× on loop/fold shapes, ≤1× on scans/builtins), not the headline. In-sample numbers are preserved throughout, explicitly relabeled in-sample.

How (files touched)

• README.md — measured-results table gains a held-out column and leads with it; Abrash-rule section states the original gate, that it failed out-of-sample and why, and the restated thesis.
• docs/mtl-spec.md — §1.1 gate verdict note, §10.6 parallel verdict, roadmap MEASURE row marked CONCLUDED, new §16 changelog entry, in-sample qualifiers on §5 admission rationale.
• bench/README.md — retirement banner leading with held-out 1.67×; headline verdicts relabeled in-sample.
• docs/RELEASE-NOTES-0.4.0.md — superseded editor's note atop the frozen v0.4.0 body (recorded numbers preserved).

This is a docs-only claims restatement — no code, primitive, glyph, spec_step, or Outcome change, and no spec-version bump.

Testing

cargo test --workspace — green, 0 failures (docs-only change touches no code; the largest suites are 97 and 24 passed, every result ok).

---

Generated by Claude Code