Diagnose F2 saved errors and optimizer trajectory without refitting #157

Open
elsie wants to merge 2 commits from research/f2-saved-diagnosis-20261020 into main
Member

Diagnose the already failed F2 reference from saved files only, without repeating its fit or unsealing18MODEL_LOCKEDcontrols. Frozen source/metadata/CSV/trajectory pins checked before and after reading. No private code or neural framework execution.

Findings: shallow depth1 diagnostic4/8, middle6/8, deep/large6/8 each; no depth-only failure. Train ends22/32, validation6/16; views agree exactly. Preupdate training BCE decreases throughout .69720→.64322 while validation worsens. All six parameter blocks have finite nonzero whole-block gradients and steps; savedSGDresidual exactly0 across512updates. This rejects a global zero-gradient/update-stall description, not optimization/capacity explanations generally. No unique LR/width/supervision correction is justified.

Six exposed roots are not six independent trials; all sinksXOR. MUX/role and repeated-input cases remain locked, so selective gate/multiplicity failure is unidentifiable. No retune, new fit, router/headroom claim or automatic followup.

Nine strengthened tests pass, including synthetic block-norm/SGD corruption, BCE/join/view controls; initial red and compressed-fixture repair under unchanged2MiB cap retained. Independent read-only scientific/code/evidence/report review found no blocker. Final guarded total5.118377CPU/14.826301335wall including publication checks, within60/120 with20reserve; allgroupsreaped/no survivors. Docs366files and generated result index pass. No GPU/spending/access/deployment or original artifact edits.

Report research/f2_saved_diagnosis_v1/REPORT.md; final DIAGNOSIS_006.json + matching source pins, card/exactcommand, ledger and failures. Output003 remains preliminary. Merge only after required CI succeeds.

Diagnose the already failed F2 reference from saved files only, without repeating its fit or unsealing18MODEL_LOCKEDcontrols. Frozen source/metadata/CSV/trajectory pins checked before and after reading. No private code or neural framework execution. Findings: shallow depth1 diagnostic4/8, middle6/8, deep/large6/8 each; no depth-only failure. Train ends22/32, validation6/16; views agree exactly. Preupdate training BCE decreases throughout .69720→.64322 while validation worsens. All six parameter blocks have finite nonzero whole-block gradients and steps; savedSGDresidual exactly0 across512updates. This rejects a global zero-gradient/update-stall description, not optimization/capacity explanations generally. No unique LR/width/supervision correction is justified. Six exposed roots are not six independent trials; all sinksXOR. MUX/role and repeated-input cases remain locked, so selective gate/multiplicity failure is unidentifiable. No retune, new fit, router/headroom claim or automatic followup. Nine strengthened tests pass, including synthetic block-norm/SGD corruption, BCE/join/view controls; initial red and compressed-fixture repair under unchanged2MiB cap retained. Independent read-only scientific/code/evidence/report review found no blocker. Final guarded total5.118377CPU/14.826301335wall including publication checks, within60/120 with20reserve; allgroupsreaped/no survivors. Docs366files and generated result index pass. No GPU/spending/access/deployment or original artifact edits. Report research/f2_saved_diagnosis_v1/REPORT.md; final DIAGNOSIS_006.json + matching source pins, card/exactcommand, ledger and failures. Output003 remains preliminary. Merge only after required CI succeeds.
research: record reviewed F2 diagnosis publication checks
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 8s
fe3383b2e5
Owner

The code is unchanged and still looks ready; test_diagnose.py passes for me (9 passed). Main has moved, though, and docs/results.md now conflicts: your row at line 34 sits where main added the BND1, BTR1, HLT1, LRM1, OSD1, PJP1, RTA1, SEL1, TCR1 and f3_reference_v1 rows. Please keep both sets of rows. Note main also has an f3_reference_v1 / saved-diagnosis-20261020 row, which is a different study that happens to share the run suffix. A one-word study prefix in the report path or the row would make that easier to tell apart.

Tiny: research/f2_saved_diagnosis_v1/.long_guard.lock is committed as an empty file, but guard.py:78 creates it at runtime. Consider adding it to .gitignore instead.

The code is unchanged and still looks ready; `test_diagnose.py` passes for me (9 passed). Main has moved, though, and `docs/results.md` now conflicts: your row at line 34 sits where main added the BND1, BTR1, HLT1, LRM1, OSD1, PJP1, RTA1, SEL1, TCR1 and f3_reference_v1 rows. Please keep both sets of rows. Note main also has an `f3_reference_v1 / saved-diagnosis-20261020` row, which is a different study that happens to share the run suffix. A one-word study prefix in the report path or the row would make that easier to tell apart. Tiny: `research/f2_saved_diagnosis_v1/.long_guard.lock` is committed as an empty file, but `guard.py:78` creates it at runtime. Consider adding it to `.gitignore` instead.
All checks were successful
checks / select relevant integrity checks (pull_request) Successful in 10s
checks / E01 candidate surface sufficiency (pull_request) Has been skipped
checks / DMS-003 protocol freeze integrity (pull_request) Has been skipped
checks / DMS-003 CPU successor integrity (pull_request) Has been skipped
checks / resource accounting and workflow consistency (pull_request) Has been skipped
checks / documentation integrity (pull_request) Successful in 8s
Required
Details
This pull request has changes conflicting with the target branch.
  • docs/results.md
View command line instructions

Manual merge helper

Use this merge commit message when completing the merge manually.

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u origin research/f2-saved-diagnosis-20261020:research/f2-saved-diagnosis-20261020
git switch research/f2-saved-diagnosis-20261020
Sign in to join this conversation.
No reviewers
No labels
R03
research-b
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
smolmodelco/thesmolmodelcompany!157
No description provided.