Closed: superseded non-record draft

records/track_non_record_16mb/2026-05-01_LastDay_Frontier_Transfer_Autopsy/PR_BODY.md

@@ -0,0 +1,80 @@
+# Non-record: Last-Day Frontier Transfer Autopsy
+
+This is a non-record methodology submission.  It does **not** claim a new
+leaderboard score.
+
+It documents my final-day attempt to transfer several plausible orthogonal
+ideas onto the clean late frontier around PR #2018, plus the local probes that
+informed those attempts.  The main purpose is to leave reviewers and future
+participants with a precise negative-results package rather than a pile of
+unexplained failed pod runs.
+
+## Headline finding
+
+On the current frontier, **pre-quant BPB is the first kill gate**.  The final
+branches all failed before quantization/TTT could matter.
+
+| Run | Base | Change | Seed | Pre-quant BPB | Quant BPB | Final BPB | Verdict |
+|---|---|---|---:|---:|---:|---:|---|
+| PR #2018 reference | #2018 | none | 1337 | **1.05124428** | 1.05990331 | **1.04826351** | target |
+| Plan A | #2018 | Gate32 + q-aware token-only tilt | 1337 | 1.06385301 | 1.07199665 | 1.06057508 | no-go |
+| Plan B | #2018 | Gate32 + native n-gram | 1337 | 1.06434971 | not run | not run | killed at pre-quant |
+| Plan C | #2018 | native n-gram + BigramHash 512x4 + Path-A-v3 small | 1337 | 1.06471733 | not retained | not retained | killed at pre-quant |
+
+The most important result is **where** these branches failed.  Gate32 damaged
+the trained neural model itself.  Removing the q-aware patch did not fix it.
+Reverting to exact #2018 gate settings and adding a tiny BigramHash branch also
+failed to recover pre-quant quality.
+
+## What is included
+
+- full Plan A log and metrics,
+- Plan B log through the pre-quant kill gate,
+- terminal-observed Plan C pre-quant result,
+- exact runner and patch scripts used for the final attempts,
+- summary of earlier probes:
+  - A40 structural smoke matrix,
+  - CrossWS tokenizer train-proxy result,
+  - fixed Memento/copy-memory no-go,
+  - context horizon audit,
+  - online bias/temperature probe,
+  - artifact savings audit.
+
+## Why this is useful
+
+Several last-day ideas looked plausible from public PRs or local probes:
+
+- gate widening,
+- runtime n-gram/memory overlays,
+- BigramHash-style input features,
+- CrossWS tokenizer changes,
+- small artifact-routing tricks.
+
+This PR records which of those actually survived contact with the 8xH100
+competition surface.  The answer was mostly "mechanically viable, not
+BPB-transferable."
+
+## Compliance posture
+
+The executed runs use the CaseOps byte sidecar (`CASEOPS_ENABLED=1`), standard
+causal transformer scoring, full-vocab normalized softmax / closed-form
+renormalized token tilt where applicable, and inherited score-first TTT.  No
+validation PreQuantTTT, no score-after-update, and no byte-only PPM scoring are
+used.
+
+## Cost note
+
+After grant credits were exhausted, I used roughly `$150` of personal RunPod
+spend across the final pushes.  I include this only as compute-accounting
+context.  The real conclusion is methodological: when pre-quant is not
+competitive, stop before quantization and TTT.
+
+## Conclusion
+
+This is not a record.  It is a record-shaped autopsy.
+
+The clearest longer-horizon positive result remains CrossWS tokenization:
+`original_crossws` gave a stable train-proxy token ratio of `0.94842` versus
+default SP8192, with byte-denominator invariants passing.  It was not
+operationally safe to finish before the deadline, but it remains the most
+promising novel direction from this work.