Non-record: Technique Taxonomy — Tier List, Interaction Effects, and BPB Verification Tools

[robbiebusinessacc]

Non-record research submission

Synthesized taxonomy from 20+ merged PRs and 880+ open PRs. Curated reference, not primary research.

What is in this document (and not elsewhere):

• Technique tier list — S/A/B/C ranked with measured BPB deltas and source PRs
• Interaction effects matrix — which combos are sub-additive (e.g. XSA-all + TTT loses 0.006 from overlap)
• BPB verification checklist — formula checks, common errors, n-gram causality verification
• N-gram legality status — organizer rulings from @valerio-oai (issue Illegal submissions megathread #677, PR Record: 5-gram Eval Cache + LeakyReLU² + Parallel Muon val_bpb: 1.0920 (3-seed mean, std 0.0007) | ~15.9 MB | 8×H100 SXM #659) and Will DePue (Discord) collected in one place
• Negative results index — pointers to PR Non-record: Depth Recurrence in Parameter-Constrained Transformers — What Works, What Doesn't, and Why #363, Research: Why Novel Architectures Fail at 16MB — Throughput-Quantization Co-optimization #831, Non-record: Negative results — hardware alignment & quantization on 8xH100 #670, Non-record: Negative results — quantization algorithms, TTT, architecture, and self-generated GPTQ calibration study #756
• Parameter budget calculator with verified configs

What this is NOT:

• Not original experiments (our best: 1.1428 BPB, PR Record: 10L Int5-MLP + BigramHash(10240) + SWA(0.4) + WD=0.04 (val_bpb=1.1428, mean 3 seeds) #180)
• Not replacing existing research PRs — references them throughout

Related PRs: #363, #831, #670, #756, #579, #549