# posttrainllm docs

> Mac-local LLM factory documentation — training, inference, evals, systems notes, and learning paths.

## Docs

- [posttrainllm docs](https://posttrainllm.com/docs/): Mac-local LLM specialist factory — training, inference, evals, systems notes, and learning paths.
- [`posttrainllm quickstart` — your first specialist in one command](https://posttrainllm.com/docs/quickstart/)
- [Agent runtime](https://posttrainllm.com/docs/agent_runtime/)
- [Async tool-call dispatch — investigated, skipped](https://posttrainllm.com/docs/async_tool_dispatch/): Why async tool dispatch in AgentLoop doesn't justify the work at current bottleneck shape. Logged so the decision doesn't get re-litigated.
- [Attempt Ledger](https://posttrainllm.com/docs/attempt-ledger/)
- [BPE-aware leaderboard scoring](https://posttrainllm.com/docs/bpe_browser_scoring/)
- [Browser notes — WASM, Workers, OPFS, WebGPU](https://posttrainllm.com/docs/browser_notes/)
- [Capability matrix — modalities × actions](https://posttrainllm.com/docs/capability_matrix/): Exhaustive map of input modalities posttrainllm supports and what operations (train, distill, tune, quantize, etc.) are available for each. Honest status per cell — what's shipped, what's partial, what's only roadmapped.
- [Citations](https://posttrainllm.com/docs/CITATIONS/)
- [CLI lab reference](https://posttrainllm.com/docs/cli-reference/)
- [Dataset inventory](https://posttrainllm.com/docs/dataset-inventory/)
- [Dataset inventory — what's available, sizes, schemas](https://posttrainllm.com/docs/data_inventory/): Reference doc for every dataset wired into posttrainllm — registry entries, what they're for, how to pull them, what the records look like after conversion, and known gotchas (gated datasets, parquet decode).
- [Doc map — where things moved](https://posttrainllm.com/docs/MAP/)
- [Document Status Registry](https://posttrainllm.com/docs/doc-status/)
- [External Products And Research Reviewed](https://posttrainllm.com/docs/external-products-reviewed/)
- [Feature ideas — interactive learning content](https://posttrainllm.com/docs/feature_ideas/)
- [Gallery v1.5 and v2 — build plan](https://posttrainllm.com/docs/gallery_v2_plan/)
- [GRPO-Clarify v1 — RL on the ambig dimension](https://posttrainllm.com/docs/GRPO_CLARIFY/)
- [Industry learning roadmap](https://posttrainllm.com/docs/industry_learning_roadmap/)
- [Leaderboard — benchmarks for tiny models](https://posttrainllm.com/docs/leaderboard/)
- [Learn posttrainllm — a guided path for software engineers new to AI](https://posttrainllm.com/docs/learn/)
- [Learning Pipeline](https://posttrainllm.com/docs/learning-pipeline/)
- [Learning Progress Tracker](https://posttrainllm.com/docs/learning-progress/)
- [Native macOS app — build plan](https://posttrainllm.com/docs/mac_app_plan/)
- [Pace landing page — pattern mine + draft](https://posttrainllm.com/docs/pace-landing-draft/)
- [Pace model manifest — every model role, decided and on disk](https://posttrainllm.com/docs/pace-model-manifest/)
- [posttrainllm — a detailed write-up](https://posttrainllm.com/docs/notes/)
- [posttrainllm — master plan (shipped / skipped / TODO)](https://posttrainllm.com/docs/PLAN/): Single source of truth for what's shipped, skipped, and still to build. Consolidated from docs/roadmap/*, docs/progress.md, docs/backlog.md, docs/audits/feature_audit_2026_05_31.md, and docs/roadmap/recent_research.md (paper catalogue). Replaces them as the canonical reference; the older docs are now pointer stubs or archived under docs/archive/.
- [posttrainllm Closure and Fresh-Experiment Gate](https://posttrainllm.com/docs/NEXT/)
- [posttrainllm Decision Log](https://posttrainllm.com/docs/decision_log/)
- [posttrainllm Docs](https://posttrainllm.com/docs/README/)
- [Progress dashboard — moved](https://posttrainllm.com/docs/progress/): Consolidated into docs/PLAN.md §1 (shipped). Headline metrics live there.
- [Project Recommendation Context](https://posttrainllm.com/docs/PROJECT_RECOMMENDATION_CONTEXT/)
- [Quality Bundle: tests, lint, crash recovery](https://posttrainllm.com/docs/quality_bundle/)
- [ROI-ordered backlog — moved](https://posttrainllm.com/docs/backlog/): Consolidated into docs/PLAN.md §3 (TODO). Tier A/B/C/D ordering preserved.
- [Tool-call extractor (mini-router)](https://posttrainllm.com/docs/tool_call_extractor/): Design + training recipe for posttrainllm's tiny encoder model that picks which tool a user query needs, before the full LM forward pass.

## Architecture

- [How it works: the specialist factory end-to-end](https://posttrainllm.com/docs/architecture/how-it-works/): A learning-first walkthrough of posttrainllm's Mac-local specialist factory — the target → data → post-training → eval → package → report loop, the native MLX/HF CLI, the browser WebGPU training path, the eval harnesses, and the key design decisions and why.

## Archive

- [Annotated transcript](https://posttrainllm.com/docs/archive/annotated_transcript/)
- [Evaluation & safety](https://posttrainllm.com/docs/archive/evaluation/)
- [I built a GPT-2 in the browser, then made it 2.6×–12.1× faster](https://posttrainllm.com/docs/archive/BLOG/)
- [Learning roadmap](https://posttrainllm.com/docs/archive/learning_roadmap/)
- [Lessons from building posttrainllm](https://posttrainllm.com/docs/archive/lessons/)
- [Milestones](https://posttrainllm.com/docs/archive/MILESTONES/)
- [Parked: the four "multi-model" directions](https://posttrainllm.com/docs/archive/parked_multi_model/)
- [Phase 9 + 10 — status and follow-up design](https://posttrainllm.com/docs/archive/phase_9_10_status/)
- [Project status — 2026 update](https://posttrainllm.com/docs/archive/status/)
- [Roadmap — recent research (2024-2026 highlights)](https://posttrainllm.com/docs/archive/recent_research/)
- [Session handoff — pick this up cleanly](https://posttrainllm.com/docs/archive/HANDOFF/)
- [Shared vs. native — how the browser and macOS paths relate](https://posttrainllm.com/docs/archive/shared_vs_native/)
- [Watch the model think](https://posttrainllm.com/docs/archive/watch_the_model_think/)
- [While you slept](https://posttrainllm.com/docs/archive/WHILE_YOU_SLEPT/)

## Artifacts

- [GitHub Corpus Parity Decision](https://posttrainllm.com/docs/artifacts/github-corpus-parity/)

## Audits

- [Audit 2026 — what we tried, what worked, what we flagged](https://posttrainllm.com/docs/audits/audit_2026/)
- [Docs Quality Audit](https://posttrainllm.com/docs/audits/docs-quality-audit/)
- [Exactness Completion Audit](https://posttrainllm.com/docs/audits/exactness-completion-audit/)
- [Feature audit (2026-05-31) — moved](https://posttrainllm.com/docs/audits/feature_audit_2026_05_31/): Verification results absorbed into docs/PLAN.md §1. Kept here only as the audit's historical anchor.
- [History Coverage Audit](https://posttrainllm.com/docs/audits/history-coverage-audit/)
- [Phase 1-10 validation + end-to-end workflows](https://posttrainllm.com/docs/audits/validation_report/)
- [Test Coverage Goals — posttrainllm](https://posttrainllm.com/docs/audits/test-coverage/)

## Factory

- [Autocorrect adapter recipe and training path](https://posttrainllm.com/docs/factory/autocorrect-adapter-recipe/)
- [Autocorrect base-model gate](https://posttrainllm.com/docs/factory/autocorrect-model-shortlist/)
- [Batch-First Post-Training](https://posttrainllm.com/docs/factory/batch-posttraining/)
- [Factory Case Study Template](https://posttrainllm.com/docs/factory/case-study-template/)
- [Factory Docs](https://posttrainllm.com/docs/factory/README/)
- [Factory Enforcement](https://posttrainllm.com/docs/factory/enforcement/)
- [Factory Eval Protocol](https://posttrainllm.com/docs/factory/eval-protocol/)
- [Factory Overview](https://posttrainllm.com/docs/factory/overview/)
- [Factory Packaging](https://posttrainllm.com/docs/factory/packaging/)
- [Factory Reports](https://posttrainllm.com/docs/factory/reports/)
- [Factory Run Lifecycle](https://posttrainllm.com/docs/factory/run-lifecycle/)
- [Factory Run Schema](https://posttrainllm.com/docs/factory/run-schema/)
- [Fine-Tune Report Card](https://posttrainllm.com/docs/factory/report-card/)
- [Foundry Evidence Contract — posttrainllm](https://posttrainllm.com/docs/factory/foundry-evidence/)
- [Hugging Face Artifact Storage](https://posttrainllm.com/docs/factory/huggingface-artifact-storage/)
- [LoRA Geometry Diagnostics](https://posttrainllm.com/docs/factory/lora-geometry/)
- [Mac-local autocorrect foundation](https://posttrainllm.com/docs/factory/autocorrect-foundation/)
- [Post-Training Factory Positioning](https://posttrainllm.com/docs/factory/post-training-factory/)
- [Public Artifacts](https://posttrainllm.com/docs/factory/public-artifacts/)
- [Report Card Cohort and Review](https://posttrainllm.com/docs/factory/report-card-cohort/)

## Guides

- [Model guide — building posttrainllm from scratch](https://posttrainllm.com/docs/guides/model_guide/)
- [posttrainllm study guide](https://posttrainllm.com/docs/guides/study_guide/)
- [Training guide](https://posttrainllm.com/docs/guides/training_guide/)

## Hf Cards

- [Qwen3-4B File-Ops Distilled — model card](https://posttrainllm.com/docs/hf-cards/qwen3-4b-file-ops-distilled/)
- [Qwen3-4B ReST Fused — model card](https://posttrainllm.com/docs/hf-cards/qwen3-4b-rest-fused/)

## Integrations

- [Deploying the playground](https://posttrainllm.com/docs/integrations/deploy/)
- [GitHub data integration](https://posttrainllm.com/docs/integrations/github_data_integration/)
- [HuggingFace Datasets Hub integration](https://posttrainllm.com/docs/integrations/hf_datasets_integration/)
- [lm-evaluation-harness integration for posttrainllm](https://posttrainllm.com/docs/integrations/lm_eval_integration/)
- [Using posttrainllm with Continue.dev / Cline / Aider](https://posttrainllm.com/docs/integrations/continue_provider/): posttrainllm exposes an Ollama-compatible HTTP surface so it drops straight into Continue.dev, Cline, and Aider configs as a local provider. Setup + caveats.

## Learn

- [Advanced architecture, RAG/agents, eval & ML system design — interview-grade map](https://posttrainllm.com/docs/learn/advanced-ml-systems-eval/): Senior/staff topics spanning modern transformer architecture, the linear-algebra↔transformers whiteboard, RAG, agents, evaluation (LLM-as-judge, perplexity, contamination), ML system-design rounds, and classic-ML depth — each with the best external source and repo anchor.
- [Advanced LLM inference & serving — interview-grade map](https://posttrainllm.com/docs/learn/advanced-llm-inference/): Senior/staff inference-optimization interview topics — KV cache & paging, batching, speculative decoding, quantization, attention kernels/variants, long context, serving architecture — mapped to the best external source and to where this repo touches it.
- [Advanced LLM training & post-training — interview-grade map](https://posttrainllm.com/docs/learn/advanced-llm-training/): Senior/staff interview topics for LLM training at scale (parallelism, precision, MoE, data) and post-training (RLHF/DPO/GRPO/reward modeling/distillation), each mapped to the best external source and to where this repo touches it.
- [Agent context as a memory hierarchy](https://posttrainllm.com/docs/learn/agent-context-hierarchy/): The L1/L2/L3 framing for agent context engineering — what we stole from the Shortcut vertical-agents essay and where each steal lives in this repo.
- [App Intents comparison for Planner v7](https://posttrainllm.com/docs/learn/app-intents-comparison/)
- [Apple on-device Foundation Models — what they are, and where they fit (and don't)](https://posttrainllm.com/docs/learn/apple-on-device-foundation-models/): What Apple's FoundationModels framework + on-device model give you, the bridge we built to score it on our gates, the measured verdict (can't ground actions; 4096-token context can't hold a tool catalog; not faster), and the strategic decision — use it as a free floor for routing, never as a dependency.
- [Castform's RL fine-tune platform — what we stole](https://posttrainllm.com/docs/learn/castform-rl-finetune/): Patterns lifted from castform.com (RL fine-tune SaaS) into posttrainllm's specialist-training surface. Sibling page to docs/learn/agent-context-hierarchy.md.
- [Competitive landscape — fine-tune, eval, interpretability](https://posttrainllm.com/docs/learn/competitive-landscape/): A 2026 map of the products around "make a model good at your task" and "is my agent any good," with the Mac-first whitespace called out. Evidence behind docs/sessions/2026-06-13-market-landscape-mac-first.md.
- [Diversity-driven small-model reasoning (Spectrum-to-Signal)](https://posttrainllm.com/docs/learn/diversity-driven-small-model-reasoning/): Three transferable training methods from the VibeThinker papers — Diversity-Exploring Distillation, MaxEnt-Guided Policy Optimization (MGPO), and specialist weight-merging — mapped to where each upgrades our own self-improvement loop. The "how a 3B matches flagship reasoning" recipe, and why one of its steps is the direct antidote to the negative transfer we measured.
- [Encoder-decoder adapters: LoRA on a seq2seq model for text repair](https://posttrainllm.com/docs/learn/encoder-decoder-adapters/)
- [Eval matrix — what every Pace LoRA actually scores](https://posttrainllm.com/docs/learn/eval-matrix-2026-06-08/)
- [Eval methodology — the gate finding](https://posttrainllm.com/docs/learn/eval-methodology-2026-06-08/)
- [External references — articles worth reading](https://posttrainllm.com/docs/learn/external-references/)
- [How an LLM actually works (the matmul-first explanation)](https://posttrainllm.com/docs/learn/llm-mechanics-fundamentals/)
- [Learn](https://posttrainllm.com/docs/learn/README/): posttrainllm's learning corpus — a ground-up curriculum from "what's a neural net" to modern training mechanics, plus reference + research notes.
- [Learning coverage map — every subsystem has a home](https://posttrainllm.com/docs/learn/coverage-map/)
- [Learning journal](https://posttrainllm.com/docs/learn/journal/)
- [Mac-local AI mastery map — what's buildable, what's built, the distributed boundary](https://posttrainllm.com/docs/learn/mac-mastery-map/): The learning + build agenda for mastering Mac-local AI. Every capability buildable on a single Mac, annotated with fleet coverage (built / partial / to-learn), plus an explicit single-machine ↔ distributed boundary. The spine for "learn everything + build everything buildable on this Mac."
- [Mathematically essential vs engineering optimization](https://posttrainllm.com/docs/learn/essential-vs-optimization/): The single most useful lens on this project — which parts compute the model's function (irreducible math, the correctness oracle) vs which parts only make that function cheaper/faster (the optimization layer). Plus the one number that polices the boundary: loss drift.
- [Model vs agent — what's actually different](https://posttrainllm.com/docs/learn/model-vs-agent/): The architectural distinction between a fine-tuned model and the agent that runs it, and what "making a good model into an agent" actually requires. Mapped onto this repo.
- [posttrainllm — study queue](https://posttrainllm.com/docs/learn/new-things/)
- [posttrainllm learning curriculum — ground up](https://posttrainllm.com/docs/learn/curriculum/)
- [Qwen3-VL mRoPE + DeepStack — math spec from HF reference](https://posttrainllm.com/docs/learn/qwen3-vl-mrope-deepstack/)
- [Reproducing Qwen Chess under a 50M-parameter ceiling](https://posttrainllm.com/docs/learn/reproducing-qwen-chess-under-50m/): What the external Qwen Chess project actually proves, what posttrainllm can reproduce, and where the 8B-to-44M compression experiment begins.
- [RL + multi-autonomous-agents roadmap — the AI game as a Mac-local RL environment](https://posttrainllm.com/docs/learn/rl-multi-agent-roadmap/): Parked blueprint. The fleet's AI game (autonomous characters in a world) is already a multi-agent RL environment — the scarcest RL ingredient. This is the plan to turn it into a self-improving-agents testbed on the Mac, and the staged roadmap toward multi-agent dynamics. Revisit after distillation + the first GRPO loop.
- [Session 1 — From a line to a learned line](https://posttrainllm.com/docs/learn/session-01-neural-net-basics/)
- [Session 10 — Attention and transformer blocks: routing information](https://posttrainllm.com/docs/learn/session-10-attention/)
- [Session 11 — Evals, rewards, and self-improvement: making progress real](https://posttrainllm.com/docs/learn/session-11-evals-rewards/)
- [Session 2 — How the search for `m` and `b` actually works](https://posttrainllm.com/docs/learn/session-02-gradient-descent/)
- [Session 3 — What makes a neural net more than linear regression](https://posttrainllm.com/docs/learn/session-03-non-linearities/)
- [Session 4 — The taxonomy of ML approaches (and where transformers fit)](https://posttrainllm.com/docs/learn/session-04-ml-paradigms/)
- [Session 5 — Scaling: why bigger models know more (and what scaling DOESN'T cover)](https://posttrainllm.com/docs/learn/session-05-scaling/)
- [Session 6 — Tokenization + embeddings: how text becomes numbers](https://posttrainllm.com/docs/learn/session-06-tokenization-embeddings/)
- [Session 7 — How models learn behavior (supervised, imitation, reinforcement)](https://posttrainllm.com/docs/learn/session-07-behavior-learning/)
- [Session 8 — Training mechanics: the actual loop](https://posttrainllm.com/docs/learn/session-08-training-mechanics/)
- [Session 9 — Vectors, matrices, tensors: the shapes that carry everything](https://posttrainllm.com/docs/learn/session-09-tensors/)
- [Small-model tool-calling: the SOTA playbook (what others do)](https://posttrainllm.com/docs/learn/small-model-tool-calling-playbook/)
- [Speech & systems topics — interview-grade map](https://posttrainllm.com/docs/learn/speech-and-systems-topics/): Voice-pipeline latency, WER, speech-to-speech, fine-tuning debugging, feature selection, queues vs websockets, FSDP2 — each mapped to the best external source and to where this codebase (or Pace) actually does it.
- [The buildable AI journey](https://posttrainllm.com/docs/learn/artifact-journey/): A nine-stage journey from transformer foundations to local specialist models, agents, packages, runtime decisions, and causal probes.
- [The WebGPU execution model (for reading our shaders)](https://posttrainllm.com/docs/learn/webgpu-execution-model/): The mental model you need to read posttrainllm's WGSL compute kernels and explain them to another engineer — device/queue, pipeline, dispatch, workgroups, invocations, bind groups, the memory hierarchy — then how our matmul + attention shaders map onto it.
- [Tool-calling: how close can a Mac-local small model get to frontier?](https://posttrainllm.com/docs/learn/tool-calling-frontier-parity/)

### Ane Research

- [ANE research dossier for posttrainllm](https://posttrainllm.com/docs/learn/ane-research/dossier/)
- [M6 — ANE Bisect Findings (2026-06-08)](https://posttrainllm.com/docs/learn/ane-research/m6-findings/)

### Archive

- [Session 1 — What's a neural net (the building block)](https://posttrainllm.com/docs/learn/archive/session-01-neural-net-basics-dense/)

## Parked

- [Parked Lanes](https://posttrainllm.com/docs/parked/README/)
- [Parked: ANE and CoreML](https://posttrainllm.com/docs/parked/ane-coreml/)
- [Parked: Broad App Polish](https://posttrainllm.com/docs/parked/app-polish/)
- [Parked: Browser and WebGPU](https://posttrainllm.com/docs/parked/browser/)
- [Parked: Tier 5 Research](https://posttrainllm.com/docs/parked/tier5/)
- [Parked: VLM](https://posttrainllm.com/docs/parked/vlm/)

## Performance

- [Cold-start bundle — results](https://posttrainllm.com/docs/performance/cold_start_results/)
- [CPU speedup bundle — measured results](https://posttrainllm.com/docs/performance/cpu_speedup_results/)
- [CPU utilization research — what we're leaving on the table](https://posttrainllm.com/docs/performance/cpu_utilization_research/)
- [Data-side performance levers](https://posttrainllm.com/docs/performance/data_perf/)
- [Determinism contract](https://posttrainllm.com/docs/performance/determinism/)
- [First Worked Run — `posttrainllm bench`](https://posttrainllm.com/docs/performance/benchmark_first_run/)
- [Flash Attention 2 — backward pass notes](https://posttrainllm.com/docs/performance/fa2_backward_notes/)
- [Flash Attention 2 — forward pass notes](https://posttrainllm.com/docs/performance/fa2_forward_notes/)
- [Gradient checkpointing — worked example](https://posttrainllm.com/docs/performance/gradient_checkpointing_results/)
- [KV Cache Optimization: GQA Audit, In-Place Buffers, Persistent Prompts](https://posttrainllm.com/docs/performance/kv_cache_optimization/)
- [Memory tradeoffs — bf16, gradient accumulation, gradient checkpointing](https://posttrainllm.com/docs/performance/memory_tradeoffs/)
- [MLXFast SDPA & tied-embedding audit (2026-05-30)](https://posttrainllm.com/docs/performance/perf_audit_mlxfast_tied/)
- [Online softmax in attention — why and how](https://posttrainllm.com/docs/performance/online_softmax_in_attention/)
- [Performance notes](https://posttrainllm.com/docs/performance/performance/)
- [Performance research — what's done, what's plausibly next, what's mythology](https://posttrainllm.com/docs/performance/perf_research/)
- [posttrainllm Inference Benchmark Harness — Design](https://posttrainllm.com/docs/performance/benchmark_harness_design/)
- [The browser-frontier performance quest](https://posttrainllm.com/docs/performance/perf_quest/)
- [YOCO — "You Only Cache Once" worked example](https://posttrainllm.com/docs/performance/yoco_results/)

## Prds

- [5.1 reasoning training on a 22M model](https://posttrainllm.com/docs/prds/1-reasoning-on-22M/)
- [5.2 test-time compute scaling experiment](https://posttrainllm.com/docs/prds/2-testtime-compute-scaling/)
- [5.3 vision-language toy](https://posttrainllm.com/docs/prds/3-vision-language-toy/)
- [5.4 diffusion LM micro-implementation](https://posttrainllm.com/docs/prds/4-diffusion-lm-micro/)
- [5.5 real sparse MoE kernels (Metal)](https://posttrainllm.com/docs/prds/5-sparse-moe-kernels/)
- [5.6 TTS toy (audio-token GPT)](https://posttrainllm.com/docs/prds/6-tts-toy/)
- [5.7 specialized explainer-video model (Lamina-like)](https://posttrainllm.com/docs/prds/7-explainer-video-model/)
- [A1 first specialist end-to-end (tool-caller)](https://posttrainllm.com/docs/prds/A1-first-specialist-tool-caller/)
- [B1 second specialist (SQL POC)](https://posttrainllm.com/docs/prds/B1-second-specialist-shell-or-sql/)
- [B10 quality classifier on pretrain data](https://posttrainllm.com/docs/prds/B10-quality-classifier/)
- [B11 WSD (warmup-stable-decay) schedule](https://posttrainllm.com/docs/prds/B11-wsd-schedule/)
- [B12 loss-spike recovery + replay](https://posttrainllm.com/docs/prds/B12-loss-spike-recovery/)
- [B13 interp-on-checkpoints methodology](https://posttrainllm.com/docs/prds/B13-interp-on-checkpoints/)
- [B14 speculative decoding (Mini-Llama draft for Mega target)](https://posttrainllm.com/docs/prds/B14-speculative-decoding/)
- [B15 layer-wise LR decay for SFT](https://posttrainllm.com/docs/prds/B15-layerwise-lr-decay-sft/)
- [B16 M5 Neural Accelerator prefill benchmark](https://posttrainllm.com/docs/prds/B16-m5-na-prefill-bench/)
- [B17 SAE Lens interop / Neuronpedia format export](https://posttrainllm.com/docs/prds/B17-saelens-interop/)
- [B18 nanochat-style `--depth` single-knob HP derivation](https://posttrainllm.com/docs/prds/B18-nanochat-depth-knob/)
- [B19 Group-SAE (layer-group SAE training)](https://posttrainllm.com/docs/prds/B19-group-sae/)
- [B2/B2b/B3/B7 mini-router family (bundled)](https://posttrainllm.com/docs/prds/B2-B7-router-family/)
- [B21 Micro-AutoMixer for specialist data mixes](https://posttrainllm.com/docs/prds/B21-micro-automixer/)
- [B23 agent eval protocol hardening](https://posttrainllm.com/docs/prds/B23-agent-eval-protocol/)
- [B25 ScaleDown Challenge specialist](https://posttrainllm.com/docs/prds/B25-scaledown-specialist/)
- [B26 server-side deferred tools](https://posttrainllm.com/docs/prds/B26-deferred-tools/)
- [B28 composite reward framework](https://posttrainllm.com/docs/prds/B28-composite-reward-framework/)
- [B31 unified model gallery + project-level model pins](https://posttrainllm.com/docs/prds/B31-gallery-and-project-pins/)
- [B32 `posttrainllm eval` as a CI / pre-commit gate](https://posttrainllm.com/docs/prds/B32-eval-ci-gate/)
- [B33 one-command laptop-finetune onboarding](https://posttrainllm.com/docs/prds/B33-laptop-finetune-onboarding/)
- [B34 batched eval-runtime + pluggable MLX backend (oMLX steals)](https://posttrainllm.com/docs/prds/B34-batched-eval-runtime/)
- [B35 local-agent vertical PoC — code reviewer on a Mac](https://posttrainllm.com/docs/prds/B35-local-agent-vertical-poc/)
- [B5 cloud-escalation training signal](https://posttrainllm.com/docs/prds/B5-cloud-escalate-training/)
- [B6 Mac app demo](https://posttrainllm.com/docs/prds/B6-mac-app-demo/)
- [B8 multilingual specialist (Indic-focus first)](https://posttrainllm.com/docs/prds/B8-multilingual-specialist/)
- [B9 energy J/token measurement](https://posttrainllm.com/docs/prds/B9-energy-per-token/)
- [C10 training-run dashboard](https://posttrainllm.com/docs/prds/C10-train-run-dashboard/)
- [C3 DoRA on-disk adapter format](https://posttrainllm.com/docs/prds/C3-dora-ondisk-format/)
- [C4 tool-call extractor — BPE tokenizer support](https://posttrainllm.com/docs/prds/C4-tool-extractor-bpe/)
- [C5 decode jitter under thermal load](https://posttrainllm.com/docs/prds/C5-decode-jitter-thermal/)
- [C9 determinism harness — bit-exact replay of step N](https://posttrainllm.com/docs/prds/C9-determinism-harness/)
- [Capability retention under fine-tuning (PRD — backlog)](https://posttrainllm.com/docs/prds/capability-retention/): Measure and preserve a small model's GENERAL intelligence when we specialize it (tool-calling, style, persona). This session proved specialization erodes breadth — a retention eval battery + retention techniques, run on every fine-tune. Future work; captured for later.
- [docs/prds/ — Product Requirement Briefs](https://posttrainllm.com/docs/prds/README/)
- [E6 `posttrainllm eval-scaledown` harness](https://posttrainllm.com/docs/prds/E6-eval-scaledown/)
- [Everyday specialist benchmark](https://posttrainllm.com/docs/prds/everyday-specialist-benchmark/)
- [Factory completeness tracker — meta-PRD](https://posttrainllm.com/docs/prds/factory-completeness-tracker/)
- [Game-as-RL-environment PoC (PRD)](https://posttrainllm.com/docs/prds/game-rl-environment-poc/): Turn the fleet's AI game (autonomous characters in a world) into a Mac-local RL environment and train a self-improving NPC with GRPO. Environments are the scarcest RL ingredient — we own a live one. The parked trigger (distillation closed + a working GRPO loop) is now MET.
- [GEPA prompt evolution — automated system-prompt iteration](https://posttrainllm.com/docs/prds/gepa-prompt-evolution/)
- [GPU / research backlog — what's left, and how to do it](https://posttrainllm.com/docs/prds/GPU-RESEARCH-BACKLOG/)
- [Local-model arena — turn-based strategy, self-play RLVR (PRD)](https://posttrainllm.com/docs/prds/local-model-arena-selfplay/): Pit Mac-local models against each other and against frontier models in turn-based strategy games, then self-play-RL a local model until it beats a frontier model playing zero-shot. The match win/lose IS a verifiable reward — RLVR with no hand-authored golds or checker.
- [Multi-turn / agentic tool-calling eval (PRD)](https://posttrainllm.com/docs/prds/multi-turn-agentic-eval/): A Mac-local, frontier-gated, STATEFUL multi-turn eval — the single biggest blind spot in our tool-calling work. Single-turn numbers (88.7 for our best 4B) overstate agentic ability; small models are known to cliff hard on multi-turn. This is the prerequisite for judging any model as an agent/planner.
- [Pace planner specialist — first factory customer](https://posttrainllm.com/docs/prds/specialist-pace-planner/)
- [Pace planner v11 — ship gate](https://posttrainllm.com/docs/prds/pace-planner-v11-ship-gate/)
- [Pace planner v11 — unhappy-path training data PRD](https://posttrainllm.com/docs/prds/pace-planner-v11-training-data/)
- [Pace Task Loop v1 — long-horizon tasks + the bulk-download benchmark](https://posttrainllm.com/docs/prds/pace-task-loop-v1/)
- [Planner v7 — tools-in-prompt (generalizable function calling)](https://posttrainllm.com/docs/prds/factory-planner-v7-tools-in-prompt/)
- [Port macOS 26 int8 direct ANE array handoff into M8 chain](https://posttrainllm.com/docs/prds/macos26-int8-ane-handoff-port/)
- [posttrainllm product thesis — the embedded local-AI runtime for Mac apps](https://posttrainllm.com/docs/prds/tinygpt-product-thesis/)
- [PRD — On-device continual-learning loop](https://posttrainllm.com/docs/prds/continual-learning-loop/)
- [PRD Priority Triage](https://posttrainllm.com/docs/prds/PRIORITY/)
- [PRD status — code-verified audit (2026-06-20)](https://posttrainllm.com/docs/prds/STATUS/)
- [QLoRA on Mac — fine-tuning larger open models on 48 GB](https://posttrainllm.com/docs/prds/qlora-large-model-finetune/)
- [Quantized inference in posttrainllm's HFModel — Swift-side QuantizedLinear](https://posttrainllm.com/docs/prds/quantized-inference-swift/)
- [Self-improving agents — the closed loop (PRD)](https://posttrainllm.com/docs/prds/self-improving-agents/): A verifiable act, score, learn, and curriculum loop for testing whether a Mac-local agent can improve without a teacher.
- [Specialist capability graph and verified cascade](https://posttrainllm.com/docs/prds/specialist-capability-graph/)
- [Vision-language specialist for Pace screen reading](https://posttrainllm.com/docs/prds/factory-vision-specialist/)
- [VLM A/B — UI-Venus-1.5-2B vs Qwen3-VL-2B for Pace's vision pillar](https://posttrainllm.com/docs/prds/vlm-ab-uivenus-vs-qwen3vl/)
- [VLM M4 — Qwen3-VL implementation plan](https://posttrainllm.com/docs/prds/factory-vision-m4-impl-plan/)
- [VLM M4 architecture decision — Qwen3-VL port vs LLaVA fallback](https://posttrainllm.com/docs/prds/factory-vision-m4-architecture-decision/)

## Recipes

- [Balanced training mix via reasoning-depth classification](https://posttrainllm.com/docs/recipes/balanced-training-mix/): Use B30 ahead of B29 to control single-hop / multi-hop / comparison ratios in the SFT corpus.
- [Cookbook - Character Specialist](https://posttrainllm.com/docs/recipes/cookbook-character-specialist/)
- [Cookbook - Personal Code Specialist](https://posttrainllm.com/docs/recipes/cookbook-personal-code-specialist/)
- [Cookbook - posttrainllm with Pydantic AI](https://posttrainllm.com/docs/recipes/cookbook-pydantic-ai/)
- [Cookbook - posttrainllm with smolagents](https://posttrainllm.com/docs/recipes/cookbook-smolagents/)
- [Eval gate (CI / pre-commit)](https://posttrainllm.com/docs/recipes/eval-gate/): Gate a posttrainllm specialist in CI — `posttrainllm eval-gate` exits non-zero when any declared eval suite regresses past threshold, on a self-hosted Mac runner so the model never leaves the device.
- [From .atraj rollouts to a trained specialist](https://posttrainllm.com/docs/recipes/from-traces/): B22 + B29 + B30 + posttrainllm sft chained — the closed substrate→training loop, V1.
- [MLX Export](https://posttrainllm.com/docs/recipes/mlx-export/)
- [posttrainllm Recipes](https://posttrainllm.com/docs/recipes/README/)
- [Recipe - Pace Planner Eval Protocol](https://posttrainllm.com/docs/recipes/eval_planner/)
- [Recipe — B25 ScaleDown specialist (extractive context compression)](https://posttrainllm.com/docs/recipes/b25-scaledown/)
- [Recipe — distill function-calling from Phi-3-mini → posttrainllm-Huge](https://posttrainllm.com/docs/recipes/distillation-fc/)
- [Recipe — micro-AutoMixer for data-mix ratios (B21)](https://posttrainllm.com/docs/recipes/automix/)
- [Recipe — Pace planner specialist](https://posttrainllm.com/docs/recipes/pace-planner/)
- [Recipe: distill a big model into a small local specialist](https://posttrainllm.com/docs/recipes/distill-specialist/)

## Research

- [B20 — Learnable cross-stream attention evaluation](https://posttrainllm.com/docs/research/cross-stream-attention-evaluation/): Read-and-evaluate write-up on the modded-nanogpt speedrun's "cross-stream attention" trick; decide whether to adopt for posttrainllm.
- [Indic-language evals — MILU + IndicGenBench wiring](https://posttrainllm.com/docs/research/indic_evals/): How posttrainllm scores Indic-language ability — MILU multi-choice across 11 langs, IndicGenBench XQuAD extractive QA. The Wave 4 gate before claiming Hindi support.
- [LLM Quality Benchmark Landscape, May 2026 — Survey for posttrainllm](https://posttrainllm.com/docs/research/quality_benchmarks_may_2026/)
- [Mac inference baseline — M5 Pro / 48GB](https://posttrainllm.com/docs/research/mac_decode_baseline_m5pro/)
- [Mac SLM agentic leaderboard v0](https://posttrainllm.com/docs/research/mac_slm_leaderboard_v0/): One artifact that cross-cuts decode speed, BFCL, τ-bench, and Pace unhappy-paths — the publication-shape view we're missing.
- [State of LLM Inference Benchmarks (May 2026)](https://posttrainllm.com/docs/research/inference_benchmarks_may_2026/)
- [Wave 2.5 kernel audit — BUILD / DEFER / DROP](https://posttrainllm.com/docs/research/wave_2_5_kernel_audit/)
- [Wave 4 landscape — TML / Apple FM / code agents / Indic](https://posttrainllm.com/docs/research/wave_4_landscape/)

## Roadmap

- [Single-machine roadmap — moved](https://posttrainllm.com/docs/roadmap/): Consolidated into docs/PLAN.md.
- [North Star — refined (2026-05-31)](https://posttrainllm.com/docs/roadmap/north_star_refined/)
- [Roadmap — open-source datasets](https://posttrainllm.com/docs/roadmap/datasets/)
- [Roadmap blockers — moved](https://posttrainllm.com/docs/roadmap/blockers/): Consolidated into docs/PLAN.md §2.
- [Roadmap honest summary — moved](https://posttrainllm.com/docs/roadmap/honest_summary/): Consolidated into docs/PLAN.md.
- [Roadmap orthogonal categories — moved](https://posttrainllm.com/docs/roadmap/categories/): Consolidated into docs/PLAN.md. Several markers in the old version were stale; the merged doc is verified against code.
- [Roadmap phased plan — moved](https://posttrainllm.com/docs/roadmap/phased_plan/): Consolidated into docs/PLAN.md. Most phases shipped; remainder in §3.
- [Roadmap recommended order — moved](https://posttrainllm.com/docs/roadmap/recommended_order/): Consolidated into docs/PLAN.md §3.
- [Roadmap Tier 1 — moved](https://posttrainllm.com/docs/roadmap/tier1/): Consolidated into docs/PLAN.md.
- [Roadmap Tier 2 — moved](https://posttrainllm.com/docs/roadmap/tier2/): Consolidated into docs/PLAN.md.
- [Roadmap Tier 3 — moved](https://posttrainllm.com/docs/roadmap/tier3/): Consolidated into docs/PLAN.md.
- [Roadmap Tier 4 (skip) — moved](https://posttrainllm.com/docs/roadmap/tier4_skip/): Consolidated into docs/PLAN.md §2.
- [Roadmap Tier 5 (2026 frontier) — moved](https://posttrainllm.com/docs/roadmap/tier5_frontier_2026/): Consolidated into docs/PLAN.md §3 Tier 5.

## Sessions

- [Session — eval-first prep before the 2-day training window](https://posttrainllm.com/docs/sessions/06-05-eval-first/)
- [Strategy session — what posttrainllm actually is, what's possible on a Mac](https://posttrainllm.com/docs/sessions/06-06-mac-specialist-platform/)
- [Specialist Embedder PRD Blocker](https://posttrainllm.com/docs/sessions/06-06-specialist-embedder-blocker/)
- [vllm-mlx Investigation](https://posttrainllm.com/docs/sessions/06-06-vllm-mlx-investigation/)
- [Session — 2026-06-08, the factory-validation night](https://posttrainllm.com/docs/sessions/06-08-factory-validation-night/)
- [Strategy session — the market landscape and the Mac-first wedge](https://posttrainllm.com/docs/sessions/06-13-market-landscape-mac-first/)
- [Session — from frontier-parity distillation to the self-improving loop](https://posttrainllm.com/docs/sessions/06-16-distill-to-self-improvement/)
- [Step-back: what we have, the wall, and the ROI menu](https://posttrainllm.com/docs/sessions/06-17-stepback-inventory-roi/)
- [Vercel Eve and the local-agent wedge](https://posttrainllm.com/docs/sessions/06-17-vercel-eve-and-local-agents/)
- [Drilldown — the experiments left untried, run to bedrock](https://posttrainllm.com/docs/sessions/DRILLDOWN/)
- [First specialist run — toolcall-v1 findings](https://posttrainllm.com/docs/sessions/specialist_v1_findings/): End-to-end first specialist training on M5 Pro. What worked, what didn't, what was unblocked, and what to try next.
- [Pace handoff — posttrainllm deliverables status (2026-06-10)](https://posttrainllm.com/docs/sessions/pace-handoff-2026-06-10/)
- [Planner lock - 2026-06-19](https://posttrainllm.com/docs/sessions/planner-lock-2026-06-19/)
- [posttrainllm — retrospective and forward plan](https://posttrainllm.com/docs/sessions/RETROSPECTIVE/)
- [posttrainllm — Working Session Q&A Log](https://posttrainllm.com/docs/sessions/qa_log/)
- [Session retrospective — 2026-05-31](https://posttrainllm.com/docs/sessions/session_2026_05_31/): Multi-hour session that took posttrainllm from "Wave 2.5 stalled" to "every CLI verified + first specialist trained end-to-end + ROI-ordered backlog." What shipped, what was learned, what to pick up next.
- [Session Retrospective: posttrainllm, the Long Working Session](https://posttrainllm.com/docs/sessions/session_retrospective/)
- [v11 baselines — committed 2026-06-09](https://posttrainllm.com/docs/sessions/v11-baselines-2026-06-09/)
- [WWDC 2026 (June 8–9) — impact on posttrainllm + Pace](https://posttrainllm.com/docs/sessions/wwdc-2026-impact/)

## Specialists

- [A1 — first tool-calling specialist](https://posttrainllm.com/docs/specialists/a1-tool-caller/)
- [B1 — second specialist: text-to-SQL](https://posttrainllm.com/docs/specialists/b1-sql/)
- [B1 SQL POC](https://posttrainllm.com/docs/specialists/b1-sql-poc/)

## Techniques

- [Audit 2026 Technique Inventory](https://posttrainllm.com/docs/techniques/audit-inventory/)
- [Constrained generation: JSON-mode for posttrainllm](https://posttrainllm.com/docs/techniques/constrained_generation/)
- [Evolution Strategies — gradient-free training](https://posttrainllm.com/docs/techniques/evolution_strategies/)
- [GaLore + training-stability bells (Tier 2)](https://posttrainllm.com/docs/techniques/galore_and_stability/)
- [Interpretability tools — what is the model thinking?](https://posttrainllm.com/docs/techniques/interpretability/)
- [Knowledge distillation — making a tiny model punch above its weight](https://posttrainllm.com/docs/techniques/distillation/)
- [LoRA guide — fine-tuning the tiny model](https://posttrainllm.com/docs/techniques/lora_guide/)
- [Method vs Recipe](https://posttrainllm.com/docs/techniques/method-vs-recipe/)
- [Mixture-of-Experts — more capacity per byte of weight](https://posttrainllm.com/docs/techniques/moe/)
- [Multi-Token Prediction — better training signal per step](https://posttrainllm.com/docs/techniques/mtp/)
- [Needle 2 baseline review](https://posttrainllm.com/docs/techniques/needle2-baseline-review/)
- [Numerics & precision — the gate framework](https://posttrainllm.com/docs/techniques/precision/)
- [Optimizers](https://posttrainllm.com/docs/techniques/optimizers/)
- [parakeet.wgsl browser smoke](https://posttrainllm.com/docs/techniques/parakeet-wgsl-browser-smoke/)
- [PEFT Variants in posttrainllm](https://posttrainllm.com/docs/techniques/peft_variants/)
- [Pruning — magnitude masks and structured surgery](https://posttrainllm.com/docs/techniques/pruning/)
- [Quantization Expansion](https://posttrainllm.com/docs/techniques/quantization_expansion/)
- [Speculative-decoding heads: Medusa + EAGLE-2](https://posttrainllm.com/docs/techniques/speculative_heads/)
- [SQL Technique Lineage — Closed](https://posttrainllm.com/docs/techniques/sql-technique-backlog/)
- [StreamingLLM + KIVI — long-context decode without growing the KV cache](https://posttrainllm.com/docs/techniques/streaming_llm_kivi/)
- [Technique Registry](https://posttrainllm.com/docs/techniques/README/)
- [TrainLoop Teardown](https://posttrainllm.com/docs/techniques/trainloop-teardown/)

## Training

- [The three phases of training — pretrain, SFT, DPO](https://posttrainllm.com/docs/training/)
- [DPO training: direct preference optimization for LLMs](https://posttrainllm.com/docs/training/dpo/)
- [posttrainllm — nightly training queue](https://posttrainllm.com/docs/training/nightly/)
- [Pretraining](https://posttrainllm.com/docs/training/pretrain/)
- [Supervised fine-tuning (SFT)](https://posttrainllm.com/docs/training/sft/)