issue-1417: framing-cell scope test for the shared context→answer map#1245
Merged
Conversation
…+ dispatcher, figures, render pins
Registry + verbatim renders + offset-mapping spans (issue1417_render), vLLM
gen with per-cell rollout upload (gen), teacher-forced token-id-concat capture
(extract), Batch-API graded compliance judging with {question}/{answer} slots
fixed live by gate G3 (judge), frozen fit825 fits + cell-vs-reference identity
battery with G1/G2/PC-3 gates (battery), issue1335-pattern CVD-pinned
work-conserving dispatcher with PASS_UNIFIED smoke (run.sh), hero + exploratory
figures with clamped CI offsets (figures), and span/registry/G1-anchor pins
(tests + sparse cone for eval_results/issue_825).
…lignment _random_orthogonal drew the Gaussian AND ran torch.linalg.qr on CPU, moving to the fit device only afterwards — two serial 3584x3584 fp64 CPU QRs per null draw (800/pair at production 200+200 draws), the dominant cost of the 961 s/pair identity-battery cells (#1335 P3 silent-CPU class). Keep the CPU randn (preserves the seeded Haar stream vs prior #825 runs), move only QR + sign-fix to the fit device via a new optional dev param; CPU environments are bitwise identical (self-tested). Sibling 2-arg callers unchanged by default. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…l device routing) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…ws 200->100 (env-overridable) per epm:compute-deviation v3; post-fix pilot 8236s/lane > 7200s abort threshold
…ble) — production pilot measured 601s/pair WITH the 100-draw descope; rc-7 abort at 15:14Z was the old 1.0h basis, refuted by two measured pilots
…rprint, ceiling-vs-numerator, healthy pair, manipulation checks)
… numerators are negative; qualify 'wherever the cell's own map is healthy')
…RSS collapse, not duplicate rows Free-analysis follow-up: per-fold GCV curves + Gram spectra for the broken (c1_helpful_ctrl, kept-fit L19 R2 -1.481) vs healthy (c2_rude, +0.546) instruct context-arm cells. Broken cell selects the grid-floor lambda=0.01 in 5/5 folds (dof_frac 0.96, train RSS 6.6e-5 of total); healthy selects interior lambda=1e3 in 5/5. Both carry comparable exact-duplicate context-row clusters (958 vs 1173 pairs), so duplicates do not discriminate. Reproduces the committed kept-fit R2 to <4e-12. Also fixes the analyzer_figs load_dotenv-before-heavy-import gate red + its pre-existing ruff B905/B007.
…; HF-mirrored at issue1417_framing_cells/raw_completions/judge/raw/)
superkaiba
marked this pull request as ready for review
July 18, 2026 22:49
# Conflicts: # eval_results/INDEX.md # tests/sparse_cones.txt
superkaiba
added a commit
that referenced
this pull request
Jul 18, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes task #1417. Experiment code: framing cells C1-C5 over the #825 pool; frozen ridge recipe + identity battery; judge compliance filter.