Skip to content

spec(crons): two Pareto acceleration cards (Refs #6488, #9065) - #11807

Merged
gHashTag merged 3 commits into
masterfrom
spec/pareto-cron-cards
Oct 23, 2026
Merged

gHashTag merged 3 commits into
masterfrom
spec/pareto-cron-cards

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

Nightly sieve gate + nightly differential oracle for the 28 blocked ml specs. Both are .t27 cron cards per specs/crons/README.md, 1 test each, 0 vacuous. Variant 1 (lab rebuild automation) needs a CI secret and is requested separately. Refs #6488, #9065.

, Refs #6488)

build.rs ran bare `git rev-parse --short HEAD`: git sizes the abbreviation by the
clone's object count, so a depth-1 clone (actions/checkout) bakes 7 characters and
a full clone 9, and two builds of one commit disagree (measured in #8148, builds
A/C/D). Pinning -c core.abbrev=9 on the command makes the identity
clone-independent, equal to the full clone's abbreviation. With the measured
RUSTFLAGS --remap-path-prefix=$CARGO_HOME=/cargo (builds F, G, Z,
specs/compiler/theory/reproducible.t27) both named C9 causes close.

Foreign budget: bootstrap/build.rs +7/-7, net 0 lines (existing file, no label).

Refs #8178 -- the two-lab rebuild and the sieve cell flip remain.
…#6488)

Both Railway labs (t27b 9b1d552cadaa and t27c 9d133cebfa89) build t27c at master 89d05ac
to one sha256, f6397142a55ee31506407278b4d7392d8c78a51d3037a784163b5a1cd5c7f67b, with the
build.rs abbreviation pin and --remap-path-prefix=$CARGO_HOME=/cargo. cell_t27c(C9) is TRUE,
fix_issue(T27C, C9) returns 0, and G8 (bit-identical builds) is PRESENT (1 -> 2 present,
8 -> 7 partial). reproducible.t27 records the two lab builds.

Numbers: specs/compiler/theory/sieve.t27 T738/T739/T741/T742; reproducible.t27 LAB_BUILDS.
…differential oracle (Refs #6488, #9065)

Pareto variants 2 and 3 of epic #6488: a nightly gate that runs the compiler sieve
tests (so a FALSE cell cannot regress silently) and a nightly differential run of
the 28 blocked ml specs against the t27c reference (so #9065 fix attempts are judged
in minutes). Both are .t27 cron cards per specs/crons/README.md; variant 1 (lab rebuild
automation) needs a CI secret and is requested, not committed.
@gHashTag
gHashTag enabled auto-merge (squash) October 23, 2026 12:47
@github-actions

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-10-23 12:48:54 UTC

Summary

Status Count
Total Open PRs 29
PRs with Failing Checks 25
PRs with All Checks Green 4
READY 2
FAILING 25
PENDING 0
NO CHECKS YET 0

These columns do not partition: 2 + 25 + 0 + 0 = 27, and there are 29 open PRs. A PR is being counted twice or not at all.

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=5aec047d42ec != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

@tri-claw

tri-claw Bot commented Oct 23, 2026

Copy link
Copy Markdown

Summary by t27-bees

Checked by the t27 compiler

t27c parse, parse-complete and typecheck on each changed .t27 file at 7766f88. Nothing was generated, built or run. A pass means the file is well formed, not that it is right.

  • specs/compiler/theory/reproducible.t27: parses and typechecks.
  • specs/compiler/theory/sieve.t27: parses and typechecks.
  • specs/crons/t27-differential-oracle.t27: parses and typechecks.
  • specs/crons/t27-sieve-gate.t27: parses and typechecks.

Read by the model

The PR adds two nightly cron cards (sieve gate and differential oracle for 28 blocked ML specs) and updates the compiler theory specs/seals to reflect C9 (bit-identical builds) now passing after #8178 fixes. The build.rs change pins git abbreviation length to 9 chars to ensure reproducible commit IDs across clone depths.

Worth a look

  • bootstrap/build.rs:114 — The git -c core.abbrev=9 rev-parse --short HEAD command assumes git supports -c flag and core.abbrev config. Older git versions (pre-2.0?) or minimal builds might not honor this, causing the bootstrap to bake a different-length abbrev than expected, breaking seal reproducibility.

  • specs/compiler/theory/sieve.t27:280 — cell_t27c now returns TRUE3 for C9, but the sieve's external candidate logic (tests like t737_no_external_candidate_is_removed) assumes t27c is removed by C1, C4, C5 only. If an external candidate somehow satisfies C9 but fails others, the filter counts in t739_which_filters_fire (assert firing == 3) could become stale if new candidates are added.

  • specs/compiler/theory/sieve.t27:608 — fix_issue returns 0 for C9 (issue closed), but t741_the_c9_measurement_moved_two_unknown_cells_and_nothing_else asserts fix_issue(T27C, C9) == 0. If t27c build is not reproducible: git short-sha length and CARGO_HOME path are baked into the binary (sieve C9) #8178 is reopened or a regression occurs, this test would need updating but the spec doesn't encode a reversion path.

  • specs/crons/t27-differential-oracle.t27:14 — SERVICE points to .github/workflows/differential-oracle.yml which must exist in the repo. If that workflow file is missing or renamed, the cron card becomes unactionable (no runtime validation in the spec itself).

  • specs/crons/t27-sieve-gate.t27:14 — Same risk: SERVICE references .github/workflows/sieve-gate.yml. If the workflow doesn't implement the sieve gate (running sieve tests + reproducible tests, failing on FALSE/vacuous), the cron card provides false confidence.


Read: 7 files, +150 -52, the whole diff. Head 7766f88. Model: openai-compatible/nvidia/nemotron-3-ultra-550b-a55b. Commands: @t27-bees review | summary | help | pause | resume. Free for every repository.

@gHashTag
gHashTag merged commit 72cc6fb into master Oct 23, 2026
25 of 37 checks passed
@gHashTag
gHashTag deleted the spec/pareto-cron-cards branch October 26, 2026 03:46
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants