fix(ladder): de-claim by exact artifact (sha256 + file) — D2 for #3715 by noahgift · Pull Request #4615 · paiml/aprender

noahgift · 2026-09-28T20:28:37Z

Operator ruling D2 (cop-inbox/handoff/ruling-3715-2010.md): "claims keyed by exact artifact (file sha / quant), per BIN-001 identity. It lands in the rc.2 PR set with red/green proof, then IQ2_XXS is withdrawn by data."

Stacked on #4603, which adds cells.declaimed. Its base is 0d/3715-b1-main because main does not have the de-claim feature yet.

Change

Before this PR, cells.declaimed was keyed (host, arch). gx10 UD-IQ2_XXS is qwen35, so an arch entry would have withdrawn the whole flagship family on that host.
An entry now names either an arch or one artifact (sha256 + file). Naming both, or giving a sha that is not 64 lowercase hex, is a FAIL.
The judge (model_ladder_cells.py), the producer and the receipt writer (model_ladder.sh) match a file entry by sha256 only. If the host holds a same-named file with different bytes, that file still owes every cell, and the judge FAILs with a message naming the stale entry.
The contract withdraws gx10 Qwen3.5-0.8B-UD-IQ2_XXS.gguf, sha256 a369165c…e54, until 0.70.1. I checked the sha with sha256sum on gx10. All four verbs hit the 600 s timeout on GB10 in the H4 run (apr 0.70.0 8d021f6). The file is withdrawn by that data, not waived.

Red / green

check	before	after
`green-cells-declaimed-file` under the pre-change judge (`MODEL_LADDER_CELLS_LIB`)	rc 1 (RED): an arch key cannot withdraw one file	rc 0
`check_model_ladder.sh --self-test`	n/a	156 cases, 0 bad
new judge mutants: declaim-file, declaim-by-name, declaim-widen, declaim-stale, declaim-sha-hex	n/a	all 5 killed
producer self-test, new mutant declaim-name	n/a	killed
real H4 inventories, gx10 + lambda	n/a	only gx10 IQ2_XXS is DECLAIMED (40 cells); gx10 still owes 200, lambda 240
receipt writer fixture: same file name, other bytes	n/a	stays in `inventory[]` (owed)
`pv validate contracts/model-capability-ladder-v1.yaml`	n/a	valid

Refs #3715 · Agent: aprender-89

🤖 Generated with Claude Code

…ne — D2 for #3715 cells.declaimed was keyed (host, arch). gx10's failing Qwen3.5-0.8B-UD-IQ2_XXS is qwen35, so an arch entry would withdraw the whole flagship family on that host. Operator ruling D2: claims keyed by exact artifact per BIN-001 identity. A de-claim now names either an arch or one artifact (sha256 + file), never both. The judge, the producer and the receipt writer all match a file entry by sha256 only. If the host holds a same-named file with different bytes, that file still owes every cell, and the judge FAILs with a message naming the stale entry. The contract withdraws gx10 IQ2_XXS a369165c… (0.70.1). All four verbs hit the 600 s timeout on GB10 in the H4 run. Every other gx10 Qwen3.5 file, and lambda, still owe their cells. Red/green: - The pre-change judge FAILs green-cells-declaimed-file. - check_model_ladder.sh --self-test: 156 cases, 0 bad. - 5 new judge mutants and 1 new producer mutant are killed. - On the real H4 inventory, only the 40 IQ2_XXS cells on gx10 are withdrawn. Refs #3715 Agent: aprender-89 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Two conflicts, and both sides are kept. - contracts/model-capability-ladder-v1.yaml: base's qwen3moe arch entry (D1) plus the D2 IQ2_XXS sha256+file entry. - scripts/check_model_ladder.sh: base's declaim-claimed cmutant plus the 5 D2 cmutants. declaim-host stays on the dkey() form. declaim_still_claimed (from base) looks up (host, arch) and never matches a sha256: key, so a file de-claim does not contradict its rung. After the merge: check_model_ladder.sh --self-test 157 cases, 0 bad; producer self-test all as expected; pv validate valid. Refs #3715 Agent: aprender-89 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…rum finding on #4615 Finding from the sonnet-5 lane at 11395f7. declaim_still_claimed (from base) looked up only (host, arch). A rung carries its artifact's sha256, so a rung naming exactly the de-claimed file would keep claiming it on that host in released pv, and no check would catch it. The lookup now goes through declaim_of (sha first, then arch). The arch message is unchanged. - New case red-cells-declaimed-file-still-claimed. - New cmutant declaim-claimed-file puts back the arch-only lookup; the new case kills it. - green-cells-declaimed-file narrows ctrl-q4km to hosts: [lambda], the same fix base applied to green-cells-declaimed. The real contract has no rung for the IQ2_XXS file, so it is unaffected. check_model_ladder.sh --self-test: 158 cases, 0 bad. Producer self-test ok. pv validate: valid. Refs #3715 Agent: aprender-89 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

noahgift · 2026-09-28T21:46:11Z

Head is now b957e7fd79. It merges base 2205cfd761 (merge commit, no rebase). Both conflicts are resolved by keeping both sides: the qwen3moe arch de-claim and the IQ2_XXS sha de-claim, plus both sets of mutants.

Quorum round 2, at 11395f7810: the sonnet-5 lane returned FAIL. The base's declaim_still_claimed ignored file de-claims, so a rung naming the de-claimed file would still claim it. Fixed in b957e7fd79 with a new red case red-cells-declaimed-file-still-claimed and a new cmutant declaim-claimed-file, which that case kills.

Quorum round 3, at b957e7fd79: PASS 3/3, degraded: same-family. The lanes were sonnet-5, a second sonnet-5 lane in the agy seat, and haiku-4-5. agy is not_measured: its headless plan mode auto-denied a command twice, so it produced no output.

check_model_ladder.sh --self-test: 158 cases, 0 bad. Producer self-test ok. pv validate: valid.

Agent: aprender-89

…w receipt so its patch-id binds the final diff Agent: aprender-89 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # .github/workflows/nightly.yml # docs/roadmaps/roadmap.yaml

…se-folded Independent pr-review (Sonnet 5) at 0ba69a1 found, measured: - an unquoted all-digit sha256 is a YAML int: it passed the 64-hex check via str() and then crashed SHA_KEY + int with TypeError. Now a FAIL row. New case red-cells-declaimed-file-sha-int + mutant declaim-sha-int (killed). - declaim_of / model_ladder.sh gone() compared inventory sha case-sensitively against the lowercase-only declaration; both now lower() the inventory side. Self-test: 159 cases, 0 bad (declaim-by-name/widen/sha-hex re-anchored). Refs #3715 Agent: aprender-89 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…und to patch-id 9e2c2357 Reviewer claude-sonnet-5 (author claude-opus-5-5). Verdict FINDINGS: 5 advisory, 0 blocking (measured 4, asserted 1). Consultations pmat/cuda/crux/mutation (148/148 killed)/agy (gemini-3.1-pro-high) all consulted. It supersedes the 0ba69a1 receipt (committed here as evidence). The receipt is unsigned; the ci pr-review-sign job signs it. Agent: aprender-89 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…rename release:ModelCell Re-ran the hotspot check that §4 requires before B2 and OBS-05, git only, against origin/main 316dee2. Trial merges used the old git merge-tree, with one positive and two negative controls. Every conflict in B2 (6 code files) and OBS-05 (ontology/shapes.rs) comes from one commit, 989cb01 (fold(0.70) #4655). Of the added lines in the ONT-10 PR heads, #4587/#4588/#4589/#4611 have 99-100% on main and #4615 has 93%; the not-on-main controls score 2.6% and 11.2%. So the "after ONT-10" condition is met in content. B2 duplicates nothing on main, but the class release:ModelCell now means two things. On main, v1 (since 989cb01) uses it for a non-generating model-command cell whose shape requires release:row. Seg 7's v2 uses it for a cell derived from kernel cells whose shape requires release:usesKernel. B2 renames its class before the PR. This is a static reading; no cargo was run (C277). Agent: la-0.71 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

noahgift · 2026-10-04T08:55:53Z

Closed: grows build-path Python, against the operator ruling of 2026-10-04. The work returns as a Rust port (the build-path Python port ticket). Branch kept.

noahgift and others added 3 commits September 28, 2026 22:26

noahgift and others added 2 commits September 29, 2026 00:16

merge origin/main (194294f) into D2 — cop 22:16Z, before the pr-revie…

0ba69a1

…w receipt so its patch-id binds the final diff Agent: aprender-89 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # .github/workflows/nightly.yml # docs/roadmaps/roadmap.yaml

This was referenced Sep 29, 2026

feat(ladder): class-keyed cells — an admitted helper host supplies a lead's owed cells (M3) #4629

Closed

feat(ladder): class-keyed cells — an admitted helper host supplies a lead's owed cells (M3) #4626

Closed

Base automatically changed from 0d/3715-b1-main to split/ont10-rest-0.70 September 29, 2026 20:03

noahgift closed this Oct 4, 2026

Navigation Menu

Repository navigation

fix(ladder): de-claim by exact artifact (sha256 + file) — D2 for #3715 - #4615

fix(ladder): de-claim by exact artifact (sha256 + file) — D2 for #3715#4615
noahgift wants to merge 6 commits into
split/ont10-rest-0.70from
89/3715-d2-file-key

noahgift commented Sep 28, 2026

Uh oh!

noahgift commented Sep 28, 2026

Uh oh!

noahgift commented Oct 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

noahgift commented Sep 28, 2026

Change

Red / green

Uh oh!

noahgift commented Sep 28, 2026

Uh oh!

noahgift commented Oct 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant