Cross-checked against real file content, real git commit timestamps, and dated ledger entries — not memory. This finds and closes the specific pattern that made the project feel chaotic: an investigation reaches a real conclusion, work moves on to something more urgent, and the original document is never updated to say so — leaving a confidently-wrong page sitting in the docs.
Real git commit timestamps, not recollection. The project started as a GPTQ-for-MLX port; it worked so well, so fast, that it snowballed straight into YAQA without a pause — the git history itself never even captured a standalone "GPTQ phase" commit, because by the time anyone ran git init here, both were already bundled into one blob.
gptq-main.zip downloaded. The plan was to build GPTQ for MLX first.GPTQ_MLX_Integration.md written: a real, first-party MLX GPTQ implementation already existed and needed no porting — just a per-tensor plan-integration wrapper. It passed its synthetic tests so cleanly the team moved straight to YAQA without ever running it as its own standalone milestone."YAQA->MLX port, Steps 1-4 complete." The snowball had already happened before version control even started.The documentation gaps found below aren't a carelessness problem — they're the direct, mechanical consequence of a team that kept succeeding faster than it could pause to write it down. Every individual investigation was documented in real depth at the time (the docs below are long, technical, and evidenced) — what broke was the connective tissue between them: the forward pointer saying "this got superseded," the ledger entry saying "this actually shipped." That's fixable without slowing the real work down, which is what the rest of this page does.
Every research/investigation doc in the project was read end to end and checked against what actually happened afterward, using dated ledger entries as ground truth.
Compiled from PORT_LEDGER.md's own dated section headers, CHANGELOG.md's entries, and each doc's own embedded date/status metadata — not reconstructed from memory. Amber dot = a conclusion that was later superseded; cyan dot = a standalone reference doc; mint = a real, standing fix.
LEDGER.md was never updated to say so until today.PORT_LEDGER.md and CHANGELOG.md as it happened — no gap here.Status: real, measured, complete — this is a real success.RESEARCH.md was never updated to say this until today's audit.ADAPTIVE_DAMPING_PLAIN_ENGLISH.md, GPTQ_LM_HEAD_SAFETY_GATE_PLAN.md) checked end to end for this audit — genuinely self-contained, still accurate, no update needed. A real safety-net mechanism, independent of which root cause it's protecting against.monitor.sh process found and fixed. None of this touches the curvature investigation above — logged promptly, no gap."Full-model reprocessing not yet started" / "Root-cause fix specification" — accurate when written. The real 363-tensor run finished the next day, reprocessing everything under the fix. Neither doc was updated to say so until today."Still open, real, and yours to decide: --fallback-bits 4" — framed as a minor tuning question. It wasn't — chasing it directly led into the next, much larger bug. Never updated until today.embed_tokens was being quantized unlike the reference. Both fixed, rebuilt via --assemble-from-resume (26h correction never re-run), independently re-verified: 0/1910 tensor mismatches, real 4-way weight-provenance chain-of-custody check, resume-manifest correction-quality analysis.LEDGER.md (stale since 09-03) and PORT_LEDGER.md's own status table (still said the full run "Not started") both updated. The two superseded-conclusion docs and two stale-status docs above corrected with real forward-pointing status notes, not silently overwritten — this page is the result.Each one is the same shape: a document reached a real, honest conclusion at the time it was written, reality moved past it, and nothing pointed forward. Fixed by adding a dated status update to the original document — not deleting or rewriting its real history.
YAQA_UMA_Hessian_Precision_Root_Cause.html| Document | Written | Original status | Now |
|---|---|---|---|
| RESEARCH.md | 09-05 | "real success" | Status update added |
| ADAPTIVE_DAMPING_PLAIN_ENGLISH.md | 09-06 | "complete" | Checked — still accurate |
| GPTQ_LM_HEAD_SAFETY_GATE_PLAN.md | 09-06 | "implemented" | Checked — still accurate |
| BF16_CURVATURE_BUG_PLAIN_ENGLISH.md | 09-06 | "not yet started" | Status line updated |
| YAQA_UMA_ROOT_CAUSE_FLOAT64_CURVATURE_FIX.md | 09-06 | "specification" | Status line updated |
| HESSIAN_ZERO_TO_EXPERT.md | 09-06 | reference/tutorial | No status concept — fine as is |
| AFFINE_MODE_CODE_REVIEW.md | 09-07 | "still open" | Resolved, forward-linked |
| RCA_scales_f32_bug.html | 09-07 | "pending benchmark" | Already updated with real post-fix results earlier today |
Last touched 2026-09-03. Missed 4 full days of real work — the entire YAQA port getting built, both major bugs, the rebuild, the benchmark now running. Its own 09-03 entry explicitly said YAQA-on-MLX was "not started" and stayed that way for 4 days after it was in fact well underway. A real summary entry, linking down into the detailed docs, is now appended in its correct place.
Its own "Status at a glance" table said the full 363-tensor run was "Not started" and Part 5c was "smoke test in progress" — both false as of today. Table corrected, and a full 2026-09-07 section added covering both bugs, the rebuild, the re-verification, and the benchmark.
The entire 2026-09-06 BF16 curvature bug — 5 real supporting documents — had zero changelog entry. Inserted in its correct chronological slot (not appended at the end), between the 09-07 and 09-06-later entries.