Community Review — SP8192 + Banking + Triple Recurrence + Parallel Residuals + Muon 0.97 + TTT

Thanks @aryanbhosale — this is the same general stack family as your other submissions today (#1540 LoRA TTT + varlen) and the #1493 lineage. Clean compliance read below.

What I found (head SHA 9304879bb01246a0f8afa11e9f54e13f7e8f246b, records/track_10min_16mb/2026-04-11_SP8192_Banking_ParResid_TripleRecur_Muon97_TTT/train_gpt.py, decoded from the import lzma as L,base64 as B self-extracting shim — 58,957 bytes / 622 lines of actual source):

• eval_val_sliding — standard stride-64 sliding window under torch.inference_mode(), stock byte-accounting
• eval_val_ttt — score-first-per-chunk (chunk scored in no_grad pass, then SGD on scored chunk, is_last_chunk guard, cosine LR across chunks)
• Banking via parameter sharing across triple-loop passes (per PR title) — structural, no eval-time leakage concern
• Triple recurrence following the Record: ParallelResiduals + MiniDepthRecurrence, 1.1063 BPB / 1.8679 nats, -0.0072 vs PR #1179, -0.0143 vs merged SOTA #1204 (msisovic) + Record: Triple Loop + Fused Kernels + Parallel Residuals + N-gram Tilt; val_bpb 1.08309 (5-seed mean) #1420 (abaybektursun) pattern
• Parallel residuals following the Record: ParallelResiduals + MiniDepthRecurrence, 1.1063 BPB / 1.8679 nats, -0.0072 vs PR #1179, -0.0143 vs merged SOTA #1204 + GPT-J lineage
• Muon 0.97 momentum — optimizer tuning delta
• No prequant_ttt_adapt_adamw(val_tokens, ...) call, no scored-region SLOT, no full_key n-gram cache

CPU smoke test (CT2038 proteus-engine, 2026-04-11):

IMPORT_OK               seconds=6.00
HP_NUM_LAYERS           11
HP_MODEL_DIM            512
HP_VOCAB_SIZE           8192
HP_QK_GAIN_INIT         5.0
HP_CODE_BYTES           19760 (shim; decoded 58,957 B)
SMOKE_TEST_PASS

Compliance summary:

• TTT: legal (score-first-per-chunk, same shape as Record: SP8192 + Pre-Quant TTT — val_bpb 1.07948 (3-seed mean) #1416 / Record: SP8192 + Pre-Quant TTT + QK-Gain 5.0 + Depth Recurrence + MuonEq-R — val_bpb 1.0791 (3-seed mean) #1423 / Record: SP8192 + Improved Parallel Residuals + Muon 0.97 + LR 0.03 + Legal TTT — val_bpb 1.07785 (3-seed mean) #1541)
• SLOT: not present
• N-gram cache: not present
• Pre-Quant TTT on val: not present
• Byte accounting: stock SP8192 — clean
• Credit lineage: banking / triple recurrence / parallel residuals are architecturally iterative on Record: ParallelResiduals + MiniDepthRecurrence, 1.1063 BPB / 1.8679 nats, -0.0072 vs PR #1179, -0.0143 vs merged SOTA #1204 / Record: Triple Loop + Fused Kernels + Parallel Residuals + N-gram Tilt; val_bpb 1.08309 (5-seed mean) #1420 / Record: SP8192 + 3-Layer Recurrence + Parallel Residuals + QK-Gain 5.25 + Legal TTT — val_bpb 1.0810 (3-seed mean) #1493 — worth an explicit credit line in the README if it isn't already there

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE. Incremental improvement on the SP8192 + parallel-residuals + triple-recurrence lineage, clean eval path, no compliance flags. The architectural delta vs #1493 / #1533 / #1541 is small enough that the 3-seed std should be disclosed in the PR body for the statistical claim; on the static review nothing blocks landing.

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): IMPORT_OK 6.00s, SMOKE_TEST_PASS. Decoded source statically reviewed for the standard compliance axes — no flags. Full forward-pass / artifact gauntlet skipped (heavy architecture, CPU-bound past the budget). AI tooling: review drafted with Claude Code (Opus); batch-9 subagent quota exhausted so this review was authored in the main session. SHA 9304879bb01246a0f8afa11e9f54e13f7e8f246b.

Community Review — Record: SP8192 + Banking + Triple Recurrence + Parallel Residuals + Muon 0.97 + TTT — val_bpb 1.0790 (5-seed mean)

BPB: 1.0790 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1413 dexhunter pattern)

What I found in the code (head SHA 9304879bb012, file records/track_10min_16mb/2026-04-11_SP8192_Banking_ParResid_TripleRecur_Muon97_TTT/train_gpt.py):

The TTT path at line 490 implements the score-first-per-chunk pattern: each chunk is scored under torch.no_grad() / inference_mode() before the base_model.train() + SGD adaptation runs on that same chunk, with an is_last_chunk guard so the final chunk gets no adaptation pass. This is the structural shape of the current leaderboard's legal frontier (PR #1413 dexhunter, the 1.0828 SP8192 + QK-Gain 5 + Legal TTT entry — verified at its head SHA against the is_last_chunk + torch.no_grad() score-first accumulator pattern).

Per Issue #402 and Issue #677, TTT is legal when each token is scored before the adapter updates on it, and that's what the code does here — chunk ci is scored under weights adapted only on chunks 0..ci-1. No prequant_ttt_adapt_adamw(val_tokens, ...) multi-epoch fine-tune, no scored-region SLOT, no target-in-key n-gram cache.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 6.00s, dim=512, layers=11, vocab=8192, code=19760 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending standard checks (3-seed validation, 16MB artifact cap, 10-min wallclock on 8×H100 SXM). The compliance picture matches the legal reference frontier and no flags were raised by the classification pass.

Auto-classification caveat: this review was drafted by the AST-based classifier against a template derived from manually-reviewed cluster PRs (#1420, #1450, #1487, #1541, #1529, #1533, #1518). If I've misread a subtlety in your eval path — e.g., multi-epoch TTT that I mistook for single-pass, or a target-in-key lookup I missed in a helper function — please flag it and I'll re-run the audit manually.

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 6.00s, dim=512, layers=11, vocab=8192, code=19760 B, SMOKE_TEST_PASS. Classification via deterministic AST-based classify_prs.py (pattern bank derived from ~65 manually-reviewed PRs earlier in the 2026-04-11 sweep). This review was auto-drafted from a template and spot-checked before posting — if the template misread your code, please call it out so I can iterate the classifier.