Audit 2026 Technique Inventory
This is the structured companion to ../audit_2026.md.
audit_2026.md is a prose audit. It mixes defaults, experimental features,
flagged alternatives, and a few duplicate mentions where the same method is
useful under different conditions. This file makes that treatment explicit.
Inventory Standard
An audit row is not automatically an attempt.
| Surface | Meaning | Where it belongs |
|---|---|---|
| Technique row | A method, feature, or implementation surface from audit_2026.md | This file |
| Measured experiment | A concrete before/after run with evidence, lesson, and next action | ../attempt-ledger.md |
| Recipe | A concrete application of a method to a target/eval | Target-specific backlog, such as sql-technique-backlog.md |
| Product cleanup | CLI, source layout, or UX simplification work | PRD/backlog, not attempt ledger |
Coverage Summary
Source: docs/audit_2026.md.
| Bucket | Audit rows | Treatment |
|---|---|---|
| Keep/default | 45 | Default or required capability, but still needs run-level evidence before quality claims |
| Experimental | 8 | Accessible for learning or future recipes, not default |
| Flagged | 30 | Kept with caveats; use only when the condition matches |
| Delete | 0 | Empty under the stated conviction bar |
| Tracked audit rows | 83 | Row-level inventory, including intentional duplicate mentions |
Duplicate / Overlap Notes
Some techniques appear in more than one bucket because the audit records both current utility and future conditions:
| Technique | Why duplicated | Exact treatment |
|---|---|---|
| YOCO | Listed as experimental and also flagged under architecture variants | Keep as long-context memory experiment; not a short-context speed default |
| BPE-dropout | Listed in data-side results and flagged as training-time exotic | Stable implementation; quality benefit still needs behavioral eval |
| MoE | Experimental architecture and separate measured dense-compute smoke | Capacity experiment until sparse dispatch exists |
| Gradient checkpointing | Default keep at large scale, but scale-dependent | Use for memory-bound large runs, not tiny/small defaults |
Keep / Default Rows
| Area | Technique | Evidence in audit | Ledger treatment |
|---|---|---|---|
| Training | AdamW optimizer | Outperformed Lion/Sophia/Muon/Adafactor in 200-step tests | Technique row; optimizer comparison rows stay flagged unless promoted to attempts |
| Training | bf16 dtype | Memory + range win; matches flagship training | Technique row |
| Training | Cosine LR + warmup | Used in flagship training runs | Technique row |
| Training | Gradient clipping | Prevents bf16 blowups; no measured downside | Technique row |
| Training | Gradient checkpointing | Behemoth B=4 ctx=1024 27.7GB -> 17.8GB | Normalized in attempt ledger as gradient-checkpointing-mac-training |
| Training | Sample packing for SFT | CoV(length*freq) 0.582 -> 0.061 | Normalized in attempt ledger as data-perf-sample-packing-bpe-dropout |
| Training | Persistent token cache | 10-30 min saved per re-run | Technique row; not enough standalone run detail here |
| Training | CPU speedup bundle | 5.0 -> 6.8 step/s on small B=16 | Normalized in attempt ledger as cpu-speedup-bundle |
| Tokenization | BPE via smollm2 | Used for real-text training | Technique row |
| Tokenization | Byte-level vocab=256 | Powers browser gallery | Technique row |
| Tokenization | HFTokenizer wrapper | Loads HF-family models | Technique row |
| Alignment | SFT with response masking | ChatML, Alpaca, Llama, plain templates work | Technique row |
| Alignment | DPO | Smoke tested; loss converges | Technique row |
| Alignment | SimPO | Reference-free, lower memory | Technique row; failed SQL hygiene SimPO is a separate attempt |
| Alignment | ORPO | Merges SFT + DPO stage | Technique row |
| Alignment | KTO | Single-side feedback path | Technique row |
| PEFT | LoRA | Standard base adapter | Technique row |
| PEFT | DoRA | 5-10% better than LoRA at same rank in smoke | Technique row |
| PEFT | LoRA-FA | Halves trainable params | Technique row |
| PEFT | LoRA+ | B-LR multiplier verified | Technique row |
| PEFT | NEFTune | Paper-backed one-line SFT win; smoke tested | Technique row |
| PEFT | Adapter file format | .lora round-trip safety | Technique row |
| PEFT | Multi-LoRA composition | Composition implementation exists | SQL static composition failure is normalized separately |
| Inference | KV cache | 470 vs 209 tok/s on flagship | Technique row; later KV optimization bundle normalized |
| Inference | KIVI int8 KV | 100% greedy-prefix match vs fp32 on flagship | Normalized in attempt ledger as streamingllm-kivi-cache-compression |
| Inference | Prefix caching | System-prompt reuse | Normalized in attempt ledger as kv-cache-optimization-bundle where measured |
| Inference | StreamingLLM sink | Quality preserved at 500 tokens | Normalized in attempt ledger as streamingllm-kivi-cache-compression |
| Inference | Speculative decoding, vanilla draft | Standard decode technique works | Technique row; Medusa/EAGLE heads normalized separately |
| Inference | HF model loading | Loads Qwen, Llama, Mistral, Phi | Technique row |
| Inference | AWQ reader | Loads AWQ-quantized HF model | Technique row |
| Inference | ANE Core ML inference path | 365 tok/s on Shakespeare via Core ML | Technique row; parked unless deploy path reactivates |
| Inference | OpenAI-compatible HTTP serve | Curl-tested; lm-eval-harness compatible | Technique row |
| Eval/bench | posttrainllm eval | Perplexity 4.71 on flagship matches val | Technique row |
| Eval/bench | posttrainllm bench | TTFT 1.91ms, 794 tok/s on Shakespeare | Technique row |
| Eval/bench | posttrainllm score-bench + manifest patcher | Browser leaderboard pipeline works | Technique row |
| Eval/bench | lm-evaluation-harness HTTP adapter | OpenAI-compatible serve verified | Technique row |
| Quality | 40 XCTests | CI gate with core coverage | Technique row |
| Quality | swiftformat + CI lint | 0 violations on 76 files | Technique row |
| Quality | Crash-recovery tests | SIGTERM-race verified | Technique row |
| Quality | GitHub Actions CI | Mac + Ubuntu runners | Technique row |
| Infrastructure | Atomic save-every + resume | SIGINT pause of v5 demonstrated | Technique row |
| Infrastructure | OOMGuard | Pre-flight memory aborts doomed configs | Technique row |
| Web playground | WebGPU + WASM browser training | Gallery models trained in browser | Browser/product attempts normalized where concrete |
| Web playground | Dynamic doc route | Docs web-visible | Factory/docs enforcement surface |
| Web playground | Leaderboard page | Scored entries exist | Technique row |
Experimental Rows
| Technique | Audit reason | Treatment |
|---|---|---|
| MoE | Pedagogical paper reimplementation | Normalized dense-compute smoke; future sparse dispatch needs new attempt |
| Distillation | Standard, not used at scale | Candidate method for agent factory recipes |
| Magpie synthetic data generation | Useful for future agent traces | Future data method, not attempted here |
| Evolution Strategies | Research curiosity | Parked learning/research surface |
| Tuned lens | Educational interpretability tool | Reference/learning surface |
| Logit lens / attention heatmap / activation patching / layer ablation | Educational interpretability tools | Reference/learning surface |
| YOCO | Long-context cache halving | Normalized smoke; future long-context gate needed |
| Sliding window attention | Bounded long-context attention | Future long-context method |
Flagged Rows
| Area | Technique | Tested / observed | When useful |
|---|---|---|---|
| Optimizer | Lion | 200 steps tiny; loss 3.18 vs AdamW 2.62 | Longer convergence runs |
| Optimizer | Sophia | 200-step Sophia-light; slightly behind AdamW | Full Sophia variant may differ |
| Optimizer | Muon | Tiny preset; 5.2 vs 16.3 step/s | Larger models where overhead amortizes |
| Optimizer | Adafactor | Huge preset 50 steps; 2x slower but lower state memory | Big memory-bound training |
| Architecture | DiffAttention | 22M smoke; no measured benefit | Larger or long-context reasoning runs |
| Architecture | MoD soft routing | No compute savings | Hard top-k plus scatter_add |
| Architecture | MTP | Smoke train; marginal regularization | Much larger bases |
| Architecture | ALiBi | Not tested at long context | Context extrapolation |
| Architecture | Sliding window attention | Not tested above ctx 4k | Long agent histories |
| Architecture | YOCO | -51% cache, -12% short-context decode | Long contexts above short-ctx crossover |
| Stability | DeepNorm | Untested in flagship runs | Very deep networks |
| Stability | Layer-wise LR decay | Not wired to a real run | Fine-tuning specialists |
| Stability | Embedding RMSNorm | v4/v5 with step-1 spike plus small lift | Needs longer controlled comparison |
| Training exotic | GaLore | 100-step loss descends; memory win unrealized | Optimizer-state surgery lands |
| Training exotic | BPE-dropout | 100 steps; regularization cost | Robustness evals |
| PEFT variant | VeRA | 30 steps; 512x fewer trainable params | Many-specialist factory |
| PEFT variant | LoftQ | 30-step simulated int4 init | Real int4 base model exists |
| PEFT variant | AdaLoRA | 30-step importance scoring | Rank reallocation implemented |
| PEFT variant | RsLoRA | Scale applied | High rank adapters |
| PEFT variant | PISSA init | Faster early convergence | SFT default candidate |
| PEFT variant | LayerDrop | Degraded fine-tune quality | Pretraining at depth |
| Quantization | SmoothQuant | Calibration works; no float matmul gain | int8 matmul kernel lands |
| Quantization | HQQ storage-only | Roundtrip shrinks file; no runtime win | Packed int4 matmul kernel |
| Quantization | GPTQ from-scratch | Rel error 0.1064; loads and samples | Own-model export path |
| Quantization | QAT | 30-step loss descends; qat-err bounded | int4/int8 specialist deployment |
| Pruning | Unstructured pruning | 50% sparsity; gzip -38% | Download-size only until sparse matmul |
| Pruning | Structured head pruning | Drop 4/8 heads; quality degrades | Physical removal implemented |
| Pruning | Structured layer pruning | 9.6M -> 8.0M, coherent samples | Real topology/wallclock path |
| Spec heads | Medusa heads | 50 head-train steps; 21-23% acceptance | Long head training with production recipe |
| Spec heads | EAGLE-2 | 50 head-train steps; 26.5% acceptance | Long head training with production recipe |
Delete Rows
The delete bucket is intentionally empty in audit_2026.md.
The actionable cleanup is not source deletion. It is:
- Inline
AUDIT FLAGnotes at entry points. - CLI curation into simple defaults plus experimental flags.
- Help text and docs that present one recipe per capability.
Those are product cleanup tasks, not attempt records.