Skip to content

Latest commit

 

History

History
130 lines (98 loc) · 7.87 KB

File metadata and controls

130 lines (98 loc) · 7.87 KB

Reference implementations — repos to watch, and what each is authoritative for

Why this file exists. A 249-line root-cause investigation of a failure this fleet spent a night re-deriving from CI logs had been sitting on this machine since 2026-07-20 — a month, not an evening. The standing rule "check just-akash before blaming a provider" existed, was in a memory index, and nobody opened its docs/.

And searching is not the remedy. Measured 2026-08-20: 304 repositories under ~/code, and 14,517 markdown files mention "exec". A keyword sweep returns a haystack. What is needed is a curated list of what a repo is authoritative for, plus a way to notice when it changes.

⇒ So each entry records a blob sha at the time of writing. tools/reference-check.py tells you which entries have moved since — those are the fixes to consider adopting. A register that cannot detect its own staleness is the thing this fleet has caught four times tonight.

How to use it

python3 tools/reference-check.py          # 0 all current · 1 something moved · 2 established nothing

A CHANGED entry is not an instruction. It says the upstream artifact moved; whether the change applies here is a judgement, and the tool deliberately does not make it.

AND NEITHER IS AN ENTRY ITSELF. A reference is authoritative for its own repo's findings, not for this one's. Within twenty minutes of the first entry being read here, a diagnosis from it — a synthetic zero means the lease was closed — was attached to a bucket in this repo where it did not hold: the provider's own lease-status reported 1/1 READY, 0.3–0.7 seconds before the failing exec. Same symptom, different cause.

⇒ ★ Port the METHOD and the INSTRUMENT; re-derive the DIAGNOSIS. The frame tracer, the liveness check and the marker-echo transfer directly. "It was a closed lease" is a finding about their provider on their day, and using it here without running the cheap direct check is the wrong-population error one repo over.

The register

repo artifact authoritative for blob at time of writing
just-akash docs/exec-reliability-investigation.md lease-shell exec reliability — distinguishing a transient SPDY/CRI stdout drop from a synthetic {"exit_code":0} against a closed lease. Method: upstream code trace + in-process frame instrumentation + live A/B on two providers + two-round quorum review. Refutes the "reorder" hypothesis in code. f4f3e9db392ac526cf204ba9ec7a71dd6139d545
just-akash just_akash/transport/lease_shell.py the exec transport itself — the implementation measured at exec pass=41 fail=0. Carries the stdin handling that a hypothesis here named as prime suspect and this repo refutes by running it. ceaed425d28310c8832f249e22bb76264a1431d5
just-akash just_akash/smoke_providers.py JUST_AKASH_TRACE_FRAMES — per-frame (code, len, rel_time) recording. ★ The instrument that separates a drop from a reorder from a synthetic zero. This fleet has no equivalent. d621760971eb7af1ed7f9bfde5e9307c6c4b86d3

Three entries, one repo, and that is the honest state of it. just-akash is the only sibling I have verified is authoritative for anything this fleet re-derives — it is referenced by four workflows in DigitalFrontier-infra and shares an org, so it is watchable on the forge as well as on disk.

Do not pad this table. An entry that nobody verified is worse than a short list: it manufactures the impression that the "has this been done?" question has been answered broadly when it has been answered once. Add an entry when you have read the artifact and can say what it is authoritative for in one sentence.

What this does NOT do

  • It does not tell you a change is relevant. It tells you the file moved.
  • It does not search. Searching 304 repos was measured and returns a haystack; this is a curated list precisely because search failed.
  • ⚠ It watches files, not repos. A fix landing in a file not listed here is invisible to it. That is the gap to close by adding entries, not by widening the tool.

⛔ Evidence that lives OUTSIDE this repository

A team starting from nForma-NEXT sees a specification and no evidence. Measured 2026-08-20: the night's findings live as 10 issues in Borduas-Holdings/Blazing-Back and 2 pages in Digital-Frontier-LDA/df-wiki, and nothing here pointed at any of them.

These entries carry no blob sha and tools/reference-check.py cannot verify them. An issue has no tree object to pin. ⇒ The checker's silence about this section is not a pass — it is out of scope, and saying so is the difference between a register that knows its own limits and one that reads as complete. Every entry below was verified by fetching it, once, on the date shown.

The six prompt-vs-practice answers — nForma-AI/nForma-NEXT#184

The audit itself is here. Six of seven roles read their committed prompt against a day of real work. DEV1 never opened it and could not be reached.

Friction reports — what each is authoritative FOR

issue role authoritative for
#1248 DEV1 five guards built, four defective
#1250 DEVOPS nine instruments, six defective
#1256 DEV3 seven wrong-answer instruments, six gave a CLEAN total — the clean-total signature
#1257 DX thirteen tools audited, thirteen defective
#1263 DEV4 "a recorded lesson does not fire at the moment of the mistake; only a mechanical check does" — and one retracted leg
#1268 DEV5 seven silent-failure workarounds; the git archive harness that tests the wrong tree
#1269 ARCHITECT eleven silent-zero workarounds; compaction that dropped authorship, not evidence
#1275 DEV2 nine workarounds; the shared-file identity collapse

Two findings that are not friction reports

issue what it establishes
#1264 C0 destroys the evidence its own failure would require — the deletion is inside the test under investigation, so the question is unanswerable retroactively
#1267 two conventions naming an enforcement path that does not exist

#1267 was re-verified on 2026-08-20 by a method that had just failed elsewhere. A sibling claim — that ci_guard_closing_keywords.py never existed — was retracted: it is 161 lines on an unmerged branch, and two agents had each searched one ref and concluded about the repository. So #1267 was re-checked across all refs:

iter_console_backends       git rev-list --all --objects | grep -c  ->  0    genuinely absent
ci_guard_closing_keywords   git rev-list --all --objects | grep -c  ->  6    exists, unmerged

That count is the discriminator, and it is cheap. "Never existed" and "exists on a ref you did not search" are different defects with different remedies — a wrong sentence versus an unmerged branch — and git ls-files, git grep and a working-tree scan cannot tell them apart.

The two wiki pages

Digital-Frontier-LDA/df-wikireading-an-instrument (an instrument returning a wrong value under correct calling) and dx-measurement-register (falsifiers, not success criteria; row 20 is falsified: four tools with real control suites, four green suites, four wrong answers on first contact with real data).

⛔ And 26 instruments that exist only on unmerged branches

Measured across 1,450 branches in Blazing-Back: 26 instrument-shaped files under scripts/ are unreachable from origin/main. Four are on a single branch each — audit_guard_populations.py, ci_skip_reason.py, ci_unfunded_relabel_audit.py, reach.py.

They are not lost — they are unfindable, which is this register's whole subject, in the one place where the remedy is a merge rather than a rewrite.