Non-record: SP8192 D 5-seed base and R-series evidence package

records/track_non_record_16mb/2026-04-13_sp8192_d_rseries_evidence/ARTIFACT_MAP.md

@@ -0,0 +1,66 @@
+# Artifact Map
+
+This folder is designed to stand on its own as a non-record records package.
+
+## Packaged Review Artifacts
+
+- report: [REPORT.md](REPORT.md)
+- overview: [README.md](README.md)
+- metadata: [submission.json](submission.json)
+- runtime manifest: [requirements.txt](requirements.txt)
+- packaged canonical script: [train_gpt.py](train_gpt.py)
+
+## Packaged Canonical D Evidence
+
+- machine-readable canonical summary:
+  [d_submission_summary.tsv](d_submission_summary.tsv)
+- canonical train logs:
+  [train_seed0.log](train_seed0.log)
+  [train_seed42.log](train_seed42.log)
+  [train_seed1234.log](train_seed1234.log)
+  [train_seed1337.log](train_seed1337.log)
+  [train_seed2025.log](train_seed2025.log)
+
+## Packaged R-Series Evidence
+
+- best measured single-seed follow-up eval log:
+  [r1_e_baseline.log](r1_e_baseline.log)
+- machine-readable R-series summary:
+  [r_series_combined_summary.tsv](r_series_combined_summary.tsv)
+
+## Script Provenance
+
+- packaged script SHA256:
+  `4f2ab2ca43105e94ea1b09924a7580a5446c72be47c2ff1d580c9c604fba69dd`
+- package-local script role:
+  single-file consolidation of the archived seed-0 `train_gpt.py` plus its archived helper chain, produced to keep counted code inside `train_gpt.py`
+- archived source paths used to create the package-local script:
+  `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed0/pr1413_combo_s0/train_gpt.py`
+  `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed0/pr1413_combo_s0/ngram_tilt.py`
+  `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed0/pr1413_combo_s0/fused_expert_kernel.cpp`
+- archived source component SHA256s:
+  - `train_gpt.py`: `db19d2a078354bd861e425965badbdb41ad644a2aec9c1c9a4f6984fca4c7019`
+  - `ngram_tilt.py`: `065ced48efcd5ae633f4307d254a0d3e475641878a0dc580f8e677b6e56aa379`
+  - `fused_expert_kernel.cpp`: `6b11646609508a84f7c2d9ddd9cdb4c133c2474ec83a50b78313d96664984056`
+- additional archived canonical `D` script paths:
+  - `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed42/pr1413_combo_s42/train_gpt.py`
+  - `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed1234/pr1413_combo_s1234/train_gpt.py`
+  - `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed1337/pr1413_combo_s1337/train_gpt.py`
+  - `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed2025/pr1413_combo_s2025/train_gpt.py`
+
+## External Archive Roots Used To Build This Package
+
+These paths are provenance references and are not required to review the packaged folder itself.
+
+- canonical `D` archive root:
+  `artifacts/runpod_pull/pr1413_archive_20260407_213205/`
+- R-series archive root:
+  `artifacts/runpod_pull/runpod_r_experiments_20260409_182045/`
+
+## Important Note
+
+Do not treat the mutable working-tree file at
+`records/track_10min_16mb/2026-04-07_SP8192_QK5_LegalTTT_ParallelResid7_TiltPrep/train_gpt.py`
+as the canonical measured artifact for this package.
+
+That path is a mutable prep surface. The packaged `train_gpt.py` is the self-contained review copy, and the archived seed-0 path above is the provenance anchor for that copy.

records/track_non_record_16mb/2026-04-13_sp8192_d_rseries_evidence/README.md

@@ -0,0 +1,84 @@
+# SP8192 D/R-Series Evidence Package
+
+This folder is a non-record submission package for the SP8192 `D` branch and the 2026-04-09 R-series sweep.
+
+It packages the strongest evidence line from this branch in a reviewer-friendly shape:
+
+- canonical base: the 5-seed `D` bundle built from the SP8192 `#1413` family
+- canonical result: `1.08128837` score-first TTT BPB (5-seed mean, `sigma = 0.00058943`)
+- best measured single-seed follow-up: `R1_e_baseline = 1.08078562`
+- important caveat: `R1_e_baseline` ran at `605s`, so it is not a clean lead submission number
+- primary contribution: a non-record evidence package documenting A/B/C/D/E, R1-R9, the fixed-Brotli rate-distortion finding, and 12+ negative results
+
+## What This Package Claims
+
+- `D` is the canonical, best-supported base from this branch because it is backed by a clean 5-seed RunPod bundle.
+- `R1_e_baseline` is a real measured single-seed follow-up signal on top of `D`, but only as a follow-up signal.
+- OWC/CDQuant create a fixed-Brotli compression-entropy penalty on this stack that overwhelms their raw BPB gain under the 16 MB cap.
+- The negative-results inventory is substantial enough to be useful to nearby Track A efforts.
+- Pegasus validation attempts are operational context only. The main evidence claim in this package is already anchored to the RunPod `D` 5-seed bundle.
+
+## What This Package Does Not Claim
+
+- It does not claim a record or a submission-valid lead result.
+- It does not claim multi-seed confirmation for `R1_e_baseline`.
+- It does not claim Pegasus produced completion-valid validation evidence.
+- It does not claim grouped OWC, CDQuant salvage, or other follow-up ideas are already demonstrated on this stack.
+- It does not treat missing Pegasus reruns as a gap in the main evidence line.
+
+## Primary Documents
+
+- Longform report: [REPORT.md](REPORT.md)
+- Artifact inventory: [ARTIFACT_MAP.md](ARTIFACT_MAP.md)
+- Submission metadata: [submission.json](submission.json)
+
+## Included Files
+
+- `README.md`
+- `REPORT.md`
+- `submission.json`
+- `requirements.txt`
+- packaged canonical `train_gpt.py`
+- canonical `D` train logs:
+  `train_seed0.log`, `train_seed42.log`, `train_seed1234.log`, `train_seed1337.log`, `train_seed2025.log`
+- machine-readable summaries:
+  `d_submission_summary.tsv`, `r_series_combined_summary.tsv`
+- best follow-up eval log:
+  `r1_e_baseline.log`
+- `ARTIFACT_MAP.md`
+
+## Canonical Packaged Script
+
+- packaged script: [train_gpt.py](train_gpt.py)
+- SHA256:
+  `4f2ab2ca43105e94ea1b09924a7580a5446c72be47c2ff1d580c9c604fba69dd`
+- archived source paths used to create the package-local script:
+  `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed0/pr1413_combo_s0/train_gpt.py`
+  `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed0/pr1413_combo_s0/ngram_tilt.py`
+  `artifacts/runpod_pull/pr1413_archive_20260407_213205/seed0/pr1413_combo_s0/fused_expert_kernel.cpp`
+- identity note:
+  the package-local `train_gpt.py` is a single-file consolidation of the archived seed-0 `train_gpt.py` plus its archived helper chain; this preserves the submission-shaped requirement that counted code live in `train_gpt.py`
+
+## Runtime Notes
+
+- A minimal dependency list is provided in [requirements.txt](requirements.txt).
+- The package-local `train_gpt.py` inlines the archived n-gram helper chain instead of shipping separate local code files.
+- The measured runs used the official Parameter Golf / RunPod CUDA environment.
+- The packaged script expects a FlashAttention runtime exposing `flash_attn_interface`; use the challenge image or an equivalent Hopper-compatible install.
+- The inlined n-gram helper writes `fused_expert_kernel.cpp` to the records folder at runtime if it is absent, then compiles `libfused_ngram.so` with `g++`.
+
+## Reviewer Snapshot
+
+- [x] Canonical `D` evidence is packaged with train logs for the five reported seeds
+- [x] `R1_e_baseline` evidence is packaged only as a follow-up eval log, not as the canonical submission basis
+- [x] The package-local `train_gpt.py` consolidates the archived seed-0 script and helper chain into a single counted code file
+- [x] The package-local `train_gpt.py` is checksum-verified as the packaged submission-shaped review artifact
+- [x] Non-record framing remains explicit; this folder does not claim a clean lead result or a grouped-OWC success
+
+## Provenance Note
+
+Do not treat the mutable working-tree file at
+`records/track_10min_16mb/2026-04-07_SP8192_QK5_LegalTTT_ParallelResid7_TiltPrep/train_gpt.py`
+as the canonical measured artifact for this package.
+
+That path is a mutable prep surface. The provenance anchors for the packaged script are the archived seed-0 source paths listed above, while the package-local `train_gpt.py` is the self-contained single-file review artifact derived from them.