Skip to content

Gate honesty: checks that pass with nothing measured or judge the wrong commit #5307

Closed as not planned
0 of 1 issue completed
Closed as not planned

Description

@noahgift

Scope

Checks that pass without measuring anything: they read the wrong commit, count zero items as green, or accept a self-declared result.

Why it matters

A check that cannot fail is worse than no check, because it is trusted. These rows are known cases.

Done when

Each check named below fails on a planted bad input (a mutation proof is recorded) and judges the commit it claims to judge.

Parent epic: #3863.

Rows


Each row was a separate ticket, folded here under the 100-open-ticket cap (TRIAGE-100, 2026-10-14) and closed as not planned. Reopen a row's ticket when its work starts; an open PR for it reopens it.

Agent: aprender-clerk

Activity

  1. added this to the 0.71.1 milestone on Oct 14, 2026
  2. changed the title [-]MASTER S1: folded sub-tickets[/-] [+]Gate honesty: checks that pass with nothing measured or judge the wrong commit[/+] on Oct 14, 2026
  3. added
    bugSomething isn't working
    P1High priority
    on Oct 14, 2026
  4. modified the milestones: 0.71.1, 0.72.0 on Oct 14, 2026
  5. noahgift commented on Oct 14, 2026

    @noahgift
    ContributorAuthor

    Row state, measured against main @ 8c3b19c:

    🤖 Generated with Claude Code

  6. noahgift commented on Oct 14, 2026

    @noahgift
    ContributorAuthor

    Owner aprender-a7n. Plan per row, measured on main @ 8c3b19c (0.72 freeze 10-16 12:00Z, due 10-17).

    row state on main plan ETA
    #5034, #5048, #5064 landed via #5223 (main's tree matches each branch; see the earlier comment) evidence only; closing is the batch call done
    #5108 landed via #5112 (6303235). No live pkill/killall line is left in any *.sh, workflow or Makefile. The three remaining hits are prose in contracts/claude-code-parity-apr-v1.yaml evidence only. A guard against reintroduction would be a new gate, so it is not proposed here done
    #4652 branch a7n/4652-d6-expiring @ 07ff7e864d, main merged, selftest 56/0 push it with trailers and ask for a slot 10-15 18:00Z
    #3846 branch c1/3846-into-b2 is 186 behind re-cut on main and re-measure the unmeasured models 10-16 12:00Z
    #3844 ticket closed, fix state unknown check each of the three items on main. If any is still open, post a row and its mutation proof 10-16 12:00Z
    #5082 open list every gate that reads main's head at read time (yokoten survey). Fixes carry to 0.73 list by 10-16 12:00Z
    #5147 open the move is a workflow change and goes to a quorum first. For 0.72: measure each stage only stage timings 10-16 12:00Z; move 0.73
    #5265 open build a case table for the three escapes (added lines, filenames, commit trailer) against the push-path scan. The fix follows the table table 10-16 12:00Z

    In 0.72: #4652, #3846, plus the evidence for #5108 and the #3844 audit. #5082, #5147 and #5265 deliver their survey or table by the freeze and land the code in 0.73.

  7. noahgift commented on Oct 14, 2026

    @noahgift
    ContributorAuthor

    #3846 row, correction to the plan above: it is not a plain re-cut. Of the branch's 19 commits, only two are this row's: 087c8d7 and a34b910. 087c8d7 adds a new scripts/lib/model_ladder_universe.py (207 lines; it reads GGUF, APR and safetensors headers) and more inline python3 in scripts/model_ladder.sh. New Python is not accepted, so cherry-picking it would add exactly what is banned.

    New plan: port the universe scan so the header reads go through apr inspect --json (Rust, already in tree) plus jq. The fixtures (four case dirs: green-held-not-swept, red-held-not-swept-no-reason, red-held-not-swept-unreadable, red-held-unaccounted) carry over unchanged as the case table. That needs an apr build to measure against, so the row moves to 0.73. For 0.72: the port design on this issue by 10-16 12:00Z.

  8. noahgift commented on Oct 14, 2026

    @noahgift
    ContributorAuthor

    #3844 audit (row from the plan above): all three fixes are on main, nothing to post. The original commit e8c341c never reached main, but the same code landed via the v0.69.3 merge-back (#4338, 5fa13fe). Checked at main 8c3b19c:

    item where on main evidence
    1. coverage note scraped the Makefile banner's 95% scripts/dogfood.sh, cov_pct extractor anchored to lines covered (N%) banner-only input gives "" (fallback fires); TOTAL: 824853/939270 lines covered (87.82%) gives 87.82%
    2. gate() threw away its output; the note-picker matched crate names scripts/dogfood.sh gate(): writes $WORKLOG/<name>.log, [log: …] on red, anchored picker the 6-line case table from the issue, run against main's pattern: 6/6 as expected (thiserror, proc-macro-error-attr2, 0 fail no longer match; the three real diagnostics still do)
    3. red scalar not reconciled with the rows scripts/check_model_ladder.sh declared_red check --self-test: 153 cases, 0 bad, rc 0; red-receipt-red-not-recorded rc=1 (the mutant is caught); the #3842 cross-check passes

    Mutation proof for 3 is that fixture. 1 and 2 are pattern case tables, run as above.

  9. noahgift commented on Oct 14, 2026

    @noahgift
    ContributorAuthor

    Merged into #2876 (TRIAGE-100, 2026-10-14): same theme, so this master's scope and every checklist row now live on #2876, and this ticket is closed as not planned. Reopen it if the theme splits again.

    Agent: aprender-clerk

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P1High prioritybugSomething isn't working

    Type

    No type

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions