Repository navigation
dogfood: every leak 0.69.0 shipped with gets an OWNING pre-release gate — lessons of 2026-09-21 become checks, not prose #3768
Description
Operator, 2026-09-21: "update memory and dogfood for this", on the list of what 0.69.0 taught us.
The pre-release protocol (.claude/skills/dogfood/SKILL.md) gains a section mapping each leak to the gate that now owns it: the check, where it runs, and its row. A leak with no owning gate is a filed row, never a paragraph.
| # | leak (0.69.0) | owning gate | row |
|---|---|---|---|
| 1 | verbs/flags/models enumerated by hand (serve, code, --prompt --chat, *Q4_K* glob, VERBS) |
cells DERIVED from the binary's surface + header dtypes, SHACL-enforced; hand-list guard | #3745, #3763 |
| 2 | apr retyped the model's chat template (0.8B 2+2="2") | prompt = rendering of the GGUF's own template, byte-equal to llama.cpp --jinja, as a dogfood gate |
#3755 |
| 3 | self-parity blind to shared tokenizer/template | tokenizer parity vs pinned llama-tokenize (dogfood gate) + CRUX per derived verb | #3726, #3739 |
| 4 | JIT cache key omits a baked parameter (FP8, RMSNorm ε) | per-key content-hash guard, run in the pre-publish GPU suite on lambda + gx10 | #3727, #3759 |
| 5 | a validator can be the wrong side | FP32/float64 ground-truth reference; an oracle change needs a mutant proving it still rejects | #3751, #3714 |
| 6 | removing a refusal delivered garbage (0-layer serve, HTTP 200) | 0-layer load refusal on every serve route; empty/length-only answer is not Pass | #3571, #3720 |
| 7 | silent CPU fallback under a GPU request | any fallback line under a GPU request = RED cell, never Pass | #3738, #3596 |
| 8 | a flag with no observable effect (temperature greedy, safetensors never samples) | flag-effect oracle: every derived mode arg changes output on ≥1 cell or is typed a no-op | #3745 (S2.5), #3754, #3760 |
| 9 | a receipt names what it inferred, not what it measured (ladder sha = worktree HEAD, not the binary) |
the judge checks the receipt's binary --version sha equals the pinned binary |
new row (aprender-62) |
| 10 | rows green alone, RED together | guard_tree + ratchet + workspace check on the ASSEMBLED batch before its PR | process (fold) |
| 11 | release scripts GNU-only (BSD sed on mini) | the multiplatform host receipt runs the release scripts ON mini | #3756, #3731 |
done_when
- The section exists in
.claude/skills/dogfood/SKILL.md, one row per leak, each naming its gate and row. No row without an owning gate or a filed issue. - Rows 2, 4 and 8 are wired as dogfood gates (not only PR tests), and the dogfood's
[package.metadata.dogfood] gateslists them once their PRs land. - Row 9 is filed and owned.
Activity
noahgift commentedon Sep 23, 2026
Milestone moved 0.69.1 → 0.70.0. Reason: each 0.69.0 leak gets a pre-release gate (0.70 Fast Train). Source: the #3988 disposition table (#3988 (comment)), row MISLABELED. Approval: operator, 2026-09-23 (relayed by cop aprender-cf: "yes apply both"). Applied by aprender-cb. Nothing was closed.
noahgift commentedon Sep 25, 2026
Parked by APR-EPIC-001: over S1 budget (rank 96/163). Reopen = pull. Applied by aprender-d1: claude+agy both said close, apr shadow row recorded (handoff/close-quorum-2026-09-25.json).