CLAUDE.md: the matching rule (P0) + the entry whose corrections produced it

[AdaWorldAPI]

Supersedes #1237 — same content, clean history. See that PR's closing note for why.

Two things, and the second is why the first exists.

1. CLAUDE.md — the matching rule (P0, operator-ruled)

Matching without reading is strictly prohibited. A shared surface feature — a width, a count, a phrase, a file path, a name — is never evidence that two things are the same thing. Before claiming A tells you about B, read B.

Five measured instances in one session:

#	matched on	transferred	what B actually was
1	6 B == 6 rails	helix residue → a Hexagon synapse surface	6 B is ONE 48-bit Signed360 sphere angle
2	48 bit == 48 bit	cam_pq's ranking result → HelixResidue	the measurement never ranged over HelixResidue
3	the words "multi-hop"	jc's synthetic-SPD EWA bound → graph recall	that pillar never traverses a graph
4	a neighbouring entry's citation	le-contract.md:160-170 cited for the LUT	:160-170 is CausalEdge64 residual carriers
5	the phrase "256² LUT"	128 KiB u16 → bgz-tensor's Fisher-z canon	that table is Vec<i8>, k*k + 8 → 64 KiB + 8 B

Instances 2, 4 and 5 occurred inside the correction of 1, 3 and 4 — so the rule binds the retraction as hard as the claim.

Placed beside the falsifiability rule, same shape: failure mode, measured instances, consequences as prohibitions. One consequence is counterintuitive enough to state outright — prefer the precise form, because precision is what makes a claim falsifiable. Instance 4 was invisible while the citation path was partial (citation_decay reported UNVERIFIABLE and did not fail); widening it to repo-relative is what let the gate check it and go red. A vaguer citation is not safer, only unfalsifiable.

2. EPIPHANIES.md — the entry that occasioned it

adaptive_order_probe.rs measures the clustered regime's best ordering at 99.90 % of words skipped. The skip's granularity is the 64-row word, and 65,536 rows is exactly 1024 words:

1023 / 1024 = 0.9990234…  →  99.90 %

Not near the ceiling — it is the ceiling: one live word out of 1024. The probe's own ramp says it twice: [4092, 3069, 2046, 1023, 0], 4092 = 4 × 1023, every delta −1023. So there is no tuning headroom above 99.90 % on that tile, and 256×256 is the skip's unit rather than the fixture's geometry.

Three further readings, each arithmetic on a number measured elsewhere:

• The carriers hand off at 10 bits of 16. A 16-bit tile coordinate against 64-row words splits 10/6; a prefix buys skip for exactly 10 bits and the last 6 are structurally invisible to a mask carrier.
• Both questions have an 8-byte form; what differs is the condition. How far compacts unconditionally (FamilyGamma, 8 B of parameters — plus atanh/tanh, so compute-versus-table, not a smaller answer). Which compacts only when the set is structured: a prefix range is Cmp::MatchU32 { pattern, care }, 8 bytes. §8a's clustered conjunct is an address prefix, so the entry's own regime is where both compact forms exist.
• 256×256 is 8 KiB, 64× inside D-GTM-0n's measured 512 KiB L2-residency bound; the tile that breaks it is exactly 2048×2048. Named as a coincidence, not as load-bearing — no tile-size sweep was run.

Fences the entry carries: not a claim that the skip is good in general (§8a's permissive and moderate rows stay 0.00 %); not a claim about the 12-byte facet, which expresses six points and never a set; not an identification of a facet rail with a cascade tier (both two bytes — shape identity, not object identity); and not a measurement.

Relation to #1235

Complementary, prior-art checked before writing. E-FUSING-FORFEITS-THE-SKIP-…-1 (landed there) establishes under what conditions the lever exists; this entry establishes what its ceiling is and why. That entry's ⊘ correcting my over-general "survivor-independence is structural" framing is accepted — gt_i32_to_mask_under exists at simd_masking_ops.rs:1614, verified.

Gates

• append_only_gate.py: 9 protected files, none shrank (EPIPHANIES.md 30463 → 30604).
• citation_decay.py --since origin/main: 0 new decay, 145 pre-existing backlog.
• supersession_index.py: regenerated after the board write, byte-identical.

No code touched, so no cargo gate applies.

🤖 Generated with Claude Code

https://claude.ai/code/session_011JRoXmYPrE8KubXVrDB71L

---

Generated by Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_45ec21c6-0583-42f4-88e0-81239a1cbe6b)

[AdaWorldAPI]

Closing. The entry's organizing frame is a graft.

adaptive_order_probe.rs runs over 65,536 rows and says nothing about a tile. "256×256" comes from D-GTM-0m's fixture — a different probe, different repo, different fixture — and I fused the two on the shared count of 64k. Any 65,536-row mask is 1024 words of 64, square or not.

What that takes with it: the "256×256 is the skip's unit" claim, the 10-bits-of-16 carrier handoff (it needs a 16-bit tile coordinate this probe does not have), the 8 KiB-vs-LUT comparison, and the D-GTM-0n tile-size pairing. That is the frame, not a detail.

What survived was one arithmetic identity — 1023/1024 = 99.90 % on a 1024-word mask — which I also never traced to the probe's own denominator.

It is the third instance in this PR of the same move the PR's own P0 rule prohibits. Not worth re-cutting.

---

Generated by Claude Code