Non-record: PR1953 K+O-only TTT + QK_GAIN_INIT=5.35

[dexhunter]

Non-record submission documenting a clean PR #1953-family follow-up.

This is not a record claim. Seed 42 was promising, but the 3-seed mean did not justify a record submission.

Results:

• seed 42: 1.05767136 val_bpb, 15,978,954 bytes, 434.877s TTT eval
• seed 314: 1.05854316 val_bpb, 15,983,413 bytes, 390.025s TTT eval
• seed 7: 1.05951958 val_bpb, 15,978,698 bytes, 391.064s TTT eval
• 3-seed mean: 1.05857803 val_bpb, sample std 0.00092460

Mechanism summary:

• PR Record: PR #1945 base + 2560 long-context + no_qv TTT mask + TTT LR 0.75 + QK_GAIN 5.25 — val_bpb 1.05855 (3-seed mean) #1953-style long-context score-first phased TTT stack
• TTT restricted to K and O LoRA adapters
• TTT_MLP_LORA=0
• QK_GAIN_INIT=5.35
• EVAL_SEQ_LEN=2688 and TTT_EVAL_SEQ_LEN=2688

Verification run locally:

• uv run python -m json.tool submission.json
• uv run python -m py_compile train_gpt.py run.py
• tools/verify_rules.py <seed log> -s train_gpt.py --strict for seeds 42, 314, and 7

Strict verifier notes: script audit passed with 17 pass / 0 fail / 0 warn for each seed. Log audit passed caps and score-first TTT, with expected legacy-log warnings for missing VERIFY_JSON blocks and validation split coverage not being directly provable from the old log format.

Refs #1017 and #677.

[andrewbaggio1]

cool work

[simon-marcus]

Nice, dex. You've consistently produced really high quality material through the whole competition.