Add LR0.85 prefix2750 legal TTT record

[ZanePeycke]

Record submission: AutoZany LR0.85 + prefix2750 legal phased TTT

This PR adds one new 10min/16MB record folder:

records/track_10min_16mb/2026-04-30_AutoZany_LR085_Prefix2750_LegalTTT_1.05908/

Result

3-seed mean val_bpb: 1.05907559
Population std: 0.00041335
Mean val_loss: 2.31764997
Hardware: 8x H100 SXM
Track: 10min_16mb

Seed	Stop step	Train ms	Pre-quant BPB	Quant no-TTT BPB	Final BPB	Eval ms	Artifact bytes
42	4889	596017	1.06193713	1.07022525	1.05849788	473829	15,976,870
0	4888	596020	1.06282344	1.07113858	1.05928718	464754	15,980,787
1234	4906	596115	1.06264173	1.07122159	1.05944171	468273	15,984,508

What changed

This is a conservative final-day variant on the public PR #1953 / PR #1945 lineage. It keeps the same legal score-first phased TTT path and changes the final TTT evaluation neighborhood:

TTT_LOCAL_LR_MULT=0.85
PHASED_TTT_PREFIX_DOCS=2750
PHASED_TTT_NUM_PHASES=3
EVAL_SEQ_LEN=2560
TTT_EVAL_SEQ_LEN=2560
TTT_MASK=no_qv
TTT_Q_LORA=0
TTT_V_LORA=0
QK_GAIN_INIT=5.25
ASYM_LOGIT_RESCALE=1
AWQ_LITE_ENABLED=1
COMPRESSOR=pergroup

The submitted train_gpt.py is the PR #1953 stack source used for the verified runs. The final BPBs above come from TTT_EVAL_ONLY=1 re-evaluations of the saved artifacts with PHASED_TTT_PREFIX_DOCS=2750.

Compliance checklist

• Adds only one new folder under records/track_10min_16mb/.
• Includes README.md.
• Includes submission.json with author, GitHub ID, score, seeds, and metadata.
• Includes train_gpt.py that runs from inside the record folder.
• Includes full training logs for seeds 42, 0, and 1234.
• Includes final prefix2750 TTT eval logs for seeds 42, 0, and 1234.
• Training is strictly under 600s for all seeds. Max observed: 596115ms.
• Evaluation is under 600s for all seeds. Max observed: 473829ms.
• Artifact is under decimal 16MB for all seeds. Max observed: 15,984,508 bytes.
• Full fixed validation set with CaseOps byte sidecar.
• Score-first phased TTT only: validation chunks are scored before adaptation updates.
• No n-gram cache, PPM mixture, SLOT, validation pretraining, pre-quant TTT, validation lookahead, external data, or network access.

Files included

• README.md: method summary, results table, compliance notes, reproduction command, lineage.
• submission.json: structured metadata and per-seed results.
• train_gpt.py: executable training/eval script.
• train_seed42.log, train_seed0.log, train_seed1234.log: full train + quantization logs.
• ttt_prefix2750_seed42.log, ttt_prefix2750_seed0.log, ttt_prefix2750_seed1234.log: final TTT_EVAL_ONLY=1 prefix2750 eval logs.

Reproduction

Run the script once per seed with the config shown in the record README. To reproduce the final reported score from a saved artifact, rerun with:

TTT_EVAL_ONLY=1 \
PHASED_TTT_PREFIX_DOCS=2750 \
TTT_LOCAL_LR_MULT=0.85 \
TTT_MASK=no_qv \
TTT_Q_LORA=0 \
TTT_V_LORA=0 \
TTT_EVAL_SEQ_LEN=2560 \
TTT_CHUNK_SIZE=48 \
COMPRESSOR=pergroup \
torchrun --standalone --nproc_per_node=8 train_gpt.py

Lineage

Built on the public PR #1953 / PR #1945 / PR #1855 lineage: AWQ-lite, Asymmetric Logit Rescale, CaseOps tokenizer, SparseAttnGate, SmearGate, LQER, QK gain, and legal score-first phased TTT. This PR contributes the final-day TTT_LOCAL_LR_MULT=0.85 + PHASED_TTT_PREFIX_DOCS=2750 legal eval selection and 3-seed verification under the hard time and artifact limits.