Today (2026-03-26) the leaderboard was transformed by eval-time n-gram
backoff cache technique. Add comprehensive context for agents:

- URGENT_ngram_backoff_breakthrough.md: full implementation guide with
  NgramEvalCache code, entropy-adaptive alpha, complementary training,
  priority order for implementation
- latest_sota_snapshot.md: updated with new PR landscape
- 3 reference code files from top PRs (openai#809 0.295, openai#803 0.442, openai#813 0.667)

The n-gram backoff is purely eval-time — adding it to our existing best
checkpoint should immediately jump from 1.119 to ~0.67 BPB.
Implementing it is now the single highest-priority task.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Three variants targeting the 0.187 BPB gap to openai#1:
- bwing_alpha: clip 0.95, alpha 0.05-0.60 (isolate alpha curve)
- bwing_entropy_shift: per-order entropy center shift (isolate)
- bwing_full_port: all openai#809 techniques + fixed order mults (fire first)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Cubric 3D back online (CADENCE=32, warm-start)
- Per-order entropy center shift from openai#809
- Alpha 0.05-0.60, clip 0.95
- Our sliding-window TTT spliced in (1 epoch, SGD, freeze 2 blocks)
- TTT runs BEFORE n-gram eval → adapted model feeds n-gram

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Port openai#809 LoRA TTT: rank-8 adapters on Q/V/LM head, AdamW, Polyak
- Add LoRA injection to CausalSelfAttention, Block, GPT forward paths
- 53s vs our old 410s TTT, 6x better BPB gain
- Cubric 3D ON + entropy shift + alpha 0.05-0.60 clip 0.95

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Implements the breakthrough eval-time technique from PR openai#809 (0.295 BPB):
- BackoffNgramMixer: order-2 to order-9 N-gram cache
- Entropy-adaptive alpha blending (model + N-gram predictions)
- Sequential eval building cache from scored tokens (legal/backward-looking)
- Configurable via NGRAM_EVAL=1 and NGRAM_MAX_ORDER=9 env vars
- GPT.forward() now supports _return_logits mode for N-gram blending

Enable with: export NGRAM_EVAL=1 NGRAM_MAX_ORDER=9

Add complementary training (from @pentxayc openai#803) and per-order
multipliers (from @AayushBaniya2006 openai#809) on top of distributed
prefill + 15-gram + order-adaptive gating.

New 3-seed results: 0.28798 / 0.28804 / 0.28810
All seeds under 16MB, training under 560s, eval under 330s.

Updated README with legality hedge, full ablation, credits.

…k trivial proposals

- research_memory.md: add PARADIGM SHIFT header, correct the eval_011 conclusion
  (failed due to naive/slow implementation, not because n-gram doesn't work),
  add OVERRIDING note in Open Hypotheses directing agents to PR openai#809 code
- codex_research_prompt.txt: add explicit ban on trivial proposals (random seed,
  minor hyperparams) in aggressive phase; add eval_011 correction note so agents
  use the correct vectorized chunk-based n-gram approach

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

The Negative Results section said 'do not retry n-gram/lambda sweeps' and
'eval_011 does not justify cross-seed confirmation'. These entries would
block agents from implementing the correct PR openai#809 vectorized n-gram cache.

Replace with correct framing: eval_011's naive per-segment implementation
was the problem (1901s, 3× over budget), not the concept. The correct
vectorized chunk-based approach achieves 0.2952 BPB in 287s.

Also supersede the 'next single-variable refinement' hypothesis entry
which assumed refinement phase; we are now in aggressive phase (gap=0.827).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…(legality review)

- SOTA target is now PR openai#803: Complementary Training + Backoff N-gram + TTT
- PR openai#809 (0.2952) excluded pending legality review
- research_memory.md: fix Working SOTA Anchor section (agent had written it
  to explicitly ignore the URGENT file and stick to 1.1194 — removed that)
- All PR openai#809 references updated to PR openai#803/openai#813
- Dashboard: SOTA now 0.4416, gap 0.681

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Extended eval-time n-gram backoff from order 9 to order 12, reduced chunk size
from 1M to 256K tokens for faster cache refresh, and increased alpha_max from
0.60 to 0.70. Two-seed validation: 0.2835 (seed=1337), 0.2833 (seed=42).
Improvement over PR openai#809 baseline: -0.0118 BPB.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

openai#809 uses INT5 — more aggressive quantization creates more entropy in
the post-quant model, letting n-gram eval rescue harder. Their quant
loss is 0.019 vs our 0.006 (INT6), but n-gram extracts 0.869 vs 0.668.

Changes from bwing_IV:
- clip_range: 31 → 15 in gptq_quantize_weight, quantize_int6_per_row,
  and _find_best_row_scales
- No cubric (it hurt in bwing_V)
- 9 hash primes (from bwing_IV)
- All openai#809 n-gram params (fixed mults, entropy shift, alpha curve)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

openai#809 trains for 525s, leaving 75s for GPTQ. We were using the full
600s default. 570s leaves 30s for GPTQ calibrate (3.4s) + quantize
(~25s) with headroom.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Green_1 scored 0.3200 BPB with oracle alpha alone. Green_2 adds LoRA TTT
to close the remaining 0.025 gap to openai#809 (0.2952).

TTT flow (score-first legal):
1. Sliding window eval scores all val tokens (frozen model)
2. LoRA rank-8 adapters injected on Q, V projections
3. Single pass over val tokens: score then adapt (AdamW, lr=3e-4)
4. Polyak averaging (decay=0.998) for stability
5. N-gram eval with oracle alpha on adapted model

Coarse stride (16x) keeps TTT under 60s. Total eval budget: ~290s.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

order=9, alpha=0.95, temp=0.85, prune=4%, NGRAM_ORDER_ADAPTIVE=1
Per-order entropy thresholds + multipliers: 0.5862 -> 0.2071 (-0.379!)

Beats PR openai#809 (0.295 BPB) which was the competition leader.

V62 (phrase cache stacked on top) now running.

Progression this session:
  V27 start: 1.0541
  V28 n-gram cache: 0.9897
  V31 alpha+order tuning: 0.8802
  V45 temp sharpening: 0.7775
  V59 full-chunk sharing: 0.5865
  V61 order-adaptive: 0.2071

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>