Non-record: Causal N-gram Logit Blend — Legal, Bug-Free, Null Result at Scale

[himanshudongre]

Summary

This is a rigorous negative result paired with the first clean legal reference implementation of an eval-time causal n-gram additive-logit blend — the technique every closed ngram-titled record PR in this repo was trying to implement.

• Clean legal implementation verified against the specific closure rulings — Record: 11L XSA + Mixed INT6 + Adaptive N-gram Cache (2->7 backoff) - val_bpb=0.9631, 3-seed #993 (hashed caches disallowed), [10min_16mb] 0.9641 BPB: LeakyReLU² + Score-First TTT + N-gram Backoff Cache #1185 (must renormalize over full vocab), Record: Nacrith Log-Bias + Full-Rescore N-gram — val_bpb 0.00000035 (3-seed mean) #959 (two-pass rescoring leaks) — via an 8-probe automated legality harness and 4-test integration suite.
• Scaling curve across 6 model configurations (2L/4L, 128d/256d, 800–4000 steps, sp1024/sp8192) showing the peak BPB improvement collapses from 0.0515 on a very weak baseline to 0.00018 on the strongest model tested — ~40× below the 0.0072 BPB record threshold, with peak α shifting from 0.5 → 0.05 as the baseline strengthens.
• Localized (bucketed) delta analysis showing 100% of the net benefit comes from cache hits whose context was first seen outside the 2048-token attention window. Section in the README explains why the sp1024 → sp8192 tokenizer transition erodes even that architectural floor.

Why this is useful

1. Every previous n-gram PR was closed for a specific C1/C2/C3/C4 violation per Issue A Field Guide to Valid Submissions #1017. This one passes all four. The automated legality harness is reusable for future eval-time adaptation techniques.
2. Saves other participants from running the same experiment and submitting the same variant. The scaling trend is monotonic and leaves no room for a "just train slightly longer" rescue.
3. Provides a per-bucket delta decomposition tool (code/localized_delta.py) that can be applied to any Track B technique to verify where its marginal gain comes from.

Legality

This is a non-record research submission. It does not claim a leaderboard position. Code, results, and logs are provided for reproduction of the reported negative result. See README.md for the full per-condition compliance table and the 12/12 passing automated tests (8 legality + 4 integration) on both CPU and CUDA.

No network calls. N-gram state is built entirely from already-scored eval tokens per Track B semantics. Single left-to-right pass. Score-before-update discipline enforced at chunk boundaries.

Test plan

• python3 code/legality_harness.py → 8/8 PASS (CPU + CUDA on A40)
• python3 code/test_integration.py → 4/4 PASS (CPU + CUDA on A40)
• Phase 1-A: 4L 256d sp8192 2000 steps alpha sweep → peak Δ 0.00223 BPB at α=0.10 (logs in training_logs/)
• Phase 1-B: 4L 256d sp8192 4000 steps α=0.05, 0.10 → peak 0.00018 BPB, hurts at α≥0.10
• Full reproduction instructions in README.md

Happy to ping the specific closed PRs (#993, #1026, #1185) to point to this as the legal reference after it lands, if that's useful.