Record: PR #1850 + Anti-Hijack Gate — val_bpb 0.99445 (full val)

[leon2k2k2k]

Summary

• Submission to track_10min_16mb building on PR Record: SP8192 + Strict Full-Val Byte PPM Mixture — 1.00495 BPB (3-seed mean) #1850's PPM-D byte-mixture infrastructure.
• 5-line anti-hijack patch in score_byte: high-λ branch is suppressed when the NN is already confident on the actual byte (-log p_NN < 0.40 bits, nn_skip_thr_nats = 0.277).
• Addresses the legality concern raised in Issue A Field Guide to Valid Submissions #1017 / Legality clarification: byte-level PPM-D mixture submissions (#1835 / #1850 / #1854 cluster) under Issue #1017 C2 #1872 about confidence-gated mixtures: when the NN already nails the byte, the PPM table cannot compound on top.
• Full-val measurement (all 47,851,520 tokens; no subsetting). PPM scoring is gathered across all 8 ranks via file-based gather, rank 0 runs the C scorer.
• Stackable with PR Record: PR #1797 base + PPM-D byte mixture (v2, full-val coverage) — mix_bpb 0.9019 / quantized_ttt 1.0621 #1881 / PR Record: SP8192 + Order-6 Strict Full-Val Byte PPM — 0.96255 BPB (3-seed mean) #1877 — the patch is local to score_byte; NN forward, PPM table construction, OMP scoring, and gather pattern are unchanged from Record: SP8192 + Strict Full-Val Byte PPM Mixture — 1.00495 BPB (3-seed mean) #1850.

Results (3 seeds, 8×H100 SXM, full 47.85M val)

Seed	mix_bpb_sidecar	gate_high_frac	Train	Eval	Artifact
42	0.99291	16.38%	595.98s	221s	15,917,572 B
7	0.99471	16.41%	596.04s	173s	15,914,567 B
1337	0.99572	16.43%	596.10s	170s	15,914,752 B
Mean	0.99445 (std 0.00141)	16.41%	596.04s	188s	15,915,630 B

All within 600s train cap, 600s eval cap, 16,000,000-byte decimal artifact cap.

Class disclosure

Same scoring scheme as PR #1850 — this submission is in the PPM byte-mixture class under discussion in Issue #1872. If #1872 disallows the class, the neural-only fallback is the diagnostic quantized number ≈1.073.

Test plan

• Reviewer reproduces 3 seeds; expects mix_bpb_sidecar within ±0.002 of the reported per-seed numbers (PPM is byte-deterministic given identical NLLs; only NN training noise contributes seed variance).
• Reviewer verifies all 3 seeds clear 600s/600s/16MB caps.
• Reviewer rules on Issue Legality clarification: byte-level PPM-D mixture submissions (#1835 / #1850 / #1854 cluster) under Issue #1017 C2 #1872 (PPM byte-mixture class admissibility).

[OE-GOD]

@leon2k2k2k — flagging a C1 (causality) concern with the anti-hijack gate, independent of the #1872 C2 discussion.

The patch:

int hi = hi_raw && !(nn_skip_thr > 0.0 && nn_logp > -nn_skip_thr);

uses nn_logp = -nll[i] / nb, where nll[i] = -log p_NN(target[i] | context) is the NN's NLL on the observed token at position i. That makes the gate's λ a function of the observation: at the same position with the same context, a different observed token would change nll[i], change nn_logp, and flip the high-λ branch. The gate's high-λ is suppressed exactly where NN happens to be confident on the actual token — which lowers NLL by construction.

This is the same failure mode that was caught on PR #1795 (where the original cf used the post-scoring prob of the observed byte). The fix there was to derive the gate purely from prefix-only state. The anti-hijack reintroduces target dependence on the NN side.

Empirical signature: thr was lowered from #1850's 0.9 to 0.76 (which on its own would push gate_high_frac higher), but the reported gate_high_frac ≈ 16.4% reflects the anti-hijack selectively suppressing high-λ on the "NN was right" cases.

A causal alternative would condition on NN's entropy H(p_NN | context) = -Σ_v p(v) log p(v) rather than -log p_NN(observed | context). That captures "is NN confident at this position" without peeking at the answer.

[leon2k2k2k]

@OE-GOD you are absolutely right. I will try to fix it with an update.