Skip to content

Harden V5 evaluation and terminal confirmation - #3537

Open
joshlf wants to merge 1 commit into
Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3from
Gbwo6xhidw7eogojb7aokfi2gthtzr6wq
Open

Harden V5 evaluation and terminal confirmation#3537
joshlf wants to merge 1 commit into
Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3from
Gbwo6xhidw7eogojb7aokfi2gthtzr6wq

Conversation

@joshlf

@joshlf joshlf commented Aug 3, 2026

Copy link
Copy Markdown
Member

Update the evaluator-only testing plan in response to limitations exposed by
the V3 and V4 scoring rounds. Separate a direct proof decision from the
downstream atoms that depend on it, so one missing root premise remains visible
without artificially counting the same reasoning error many times.

Require metamorphic holdouts that preserve a semantic obligation while varying
surface syntax, terminology, order, and irrelevant structure. This tests the
general proof method rather than recall of frozen fixtures. Canonicalize report
selection, validation, retry handling, and output caps; preregister objective
machine-materializable gates and retain operational attempts separately from
the scored corpus.

Define a finite sequence of diagnostic and terminal rounds with an explicit
maximum round Nmax. Each revision must trace to a general failure class, and
the final VN decision must use frozen unseen tasks and absolute gates. Stop only
when the preregistered terminal criteria pass or Nmax is reached and the
remaining limitations are reported; do not keep adapting tests until a desired
answer appears.

This commit changes the evaluation protocol, not the V5 runtime instructions,
and records no V5 evaluation outcome.

Agent-Authored-By: AI agent acting on Josh Liebow-Feeser's behalf


This PR is on branch codex/unsafe-rust-stack.

Latest Update: v22 — Compare vs v21

📚 Full Patch History

Links show the diff between the row version and the column version.

Version v21 v20 v19 v18 v17 v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v22 v21 v20 v19 v18 v17 v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v21 v20 v19 v18 v17 v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v20 v19 v18 v17 v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v19 v18 v17 v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v18 v17 v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v17 v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v16 v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v15 v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v14 v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v13 v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v12 v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v11 v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v10 v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v9 v8 v7 v6 v5 v4 v3 v2 v1 Base
v8 v7 v6 v5 v4 v3 v2 v1 Base
v7 v6 v5 v4 v3 v2 v1 Base
v6 v5 v4 v3 v2 v1 Base
v5 v4 v3 v2 v1 Base
v4 v3 v2 v1 Base
v3 v2 v1 Base
v2 v1 Base
v1 Base
⬇️ Download this PR

Branch

git fetch origin refs/heads/Gbwo6xhidw7eogojb7aokfi2gthtzr6wq && git checkout -b pr-Gbwo6xhidw7eogojb7aokfi2gthtzr6wq FETCH_HEAD

Checkout

git fetch origin refs/heads/Gbwo6xhidw7eogojb7aokfi2gthtzr6wq && git checkout FETCH_HEAD

Cherry Pick

git fetch origin refs/heads/Gbwo6xhidw7eogojb7aokfi2gthtzr6wq && git cherry-pick FETCH_HEAD

Pull

git pull origin refs/heads/Gbwo6xhidw7eogojb7aokfi2gthtzr6wq

Stacked PRs enabled by GHerrit.

@codecov-commenter

codecov-commenter commented Aug 29, 2026

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.85%. Comparing base (18f87cf) to head (46e28aa).

Additional details and impacted files
@@                        Coverage Diff                         @@
##           Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3    #3537   +/-   ##
==================================================================
  Coverage                              91.85%   91.85%           
==================================================================
  Files                                     20       20           
  Lines                                   6093     6093           
==================================================================
  Hits                                    5597     5597           
  Misses                                   496      496           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:
  • ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 546936e to 8d7339d Compare August 29, 2026 11:11
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from ddfd561 to 3253c71 Compare August 29, 2026 11:11
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 8d7339d to 7887aaf Compare September 1, 2026 22:15
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 3253c71 to 11da6ed Compare September 1, 2026 22:15
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 4cde1af to 5d5644a Compare September 2, 2026 21:51
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 06ab818 to 30053ce Compare September 2, 2026 21:51
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 5d5644a to 51b9b60 Compare September 2, 2026 22:28
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 30053ce to 02185c3 Compare September 2, 2026 22:28
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 51b9b60 to 5663f05 Compare September 2, 2026 23:04
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 02185c3 to 51a01e7 Compare September 2, 2026 23:04
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 5663f05 to ec78a58 Compare September 3, 2026 00:07
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 51a01e7 to aee096e Compare September 3, 2026 00:07
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from ec78a58 to b99ddae Compare September 3, 2026 01:00
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from aee096e to 5089878 Compare September 3, 2026 01:00
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from b99ddae to bd8bf74 Compare September 3, 2026 01:33
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch 2 times, most recently from 4e5de45 to 235283d Compare September 3, 2026 02:26
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch 2 times, most recently from 26e4b89 to f52021f Compare September 3, 2026 03:03
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch 2 times, most recently from 1a0d48b to 1ea2125 Compare September 3, 2026 04:06
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from f52021f to 0f13ff8 Compare September 3, 2026 04:06
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 1ea2125 to 1b47f32 Compare September 3, 2026 04:42
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch 2 times, most recently from 24450e1 to 68cc8b3 Compare September 3, 2026 05:14
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 1b47f32 to da1c836 Compare September 3, 2026 05:14
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 68cc8b3 to fd3c9d2 Compare September 3, 2026 05:47
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch 2 times, most recently from 01bff0d to 46e28aa Compare September 3, 2026 06:17
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from fd3c9d2 to 18f87cf Compare September 3, 2026 06:17
Update the evaluator-only testing plan in response to limitations exposed by
the V3 and V4 scoring rounds. Separate a direct proof decision from the
downstream atoms that depend on it, so one missing root premise remains visible
without artificially counting the same reasoning error many times.

Require metamorphic holdouts that preserve a semantic obligation while varying
surface syntax, terminology, order, and irrelevant structure. This tests the
general proof method rather than recall of frozen fixtures. Canonicalize report
selection, validation, retry handling, and output caps; preregister objective
machine-materializable gates and retain operational attempts separately from
the scored corpus.

Define a finite sequence of diagnostic and terminal rounds with an explicit
maximum round Nmax. Each revision must trace to a general failure class, and
the final VN decision must use frozen unseen tasks and absolute gates. Stop only
when the preregistered terminal criteria pass or Nmax is reached and the
remaining limitations are reported; do not keep adapting tests until a desired
answer appears.

This commit changes the evaluation protocol, not the V5 runtime instructions,
and records no V5 evaluation outcome.

gherrit-pr-id: Gbwo6xhidw7eogojb7aokfi2gthtzr6wq
Agent-Authored-By: AI agent acting on Josh Liebow-Feeser's behalf
@joshlf
joshlf force-pushed the Gbwo6xhidw7eogojb7aokfi2gthtzr6wq branch from 46e28aa to fd2adbd Compare September 3, 2026 06:47
@joshlf
joshlf force-pushed the Gj24u57i5qq6wbmrnwypu7cy7dupt2mt3 branch from 18f87cf to 8376fca Compare September 3, 2026 06:47
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.

3 participants