the recipe log

Recipes, not vibes.

A recipe names the failure mode, the data, the eval gate, the slice gate, and the stop rule. "Try DPO" is not a recipe — it's a wish. These are the named, reproducible post-training and packaging recipes the factory runs, plus cookbooks for real integrations. Companion logs: the devlog and theexperiment lab.

Post-training recipes

Cookbooks — real integrations

Complete technique → recipe registry

Every retained technique has a final project disposition and a concrete lab contract: target, failure mode, data, method, eval, regressions, budget, stop rule, decision rule, exercise, and mastery gate. These are ready learning labs, not an unfinished AI backlog.

validated with caveat

Constrained generation

A capable model emits invalid syntax or forbidden output shapes.

Lab
Run a tiny legal-token or JSON-schema fixture and compare raw versus constrained validity.
Mastery
Explain why legal output is not the same as correct output.
Open evidence and method →
validated with caveat

Specialist distillation

A small base lacks a narrow behavior that a qualified teacher can demonstrate.

Lab
Trace the file-ops lineage from stock 4B through distilled, multibackend, and ReST candidates.
Mastery
State the target gain and breadth cost that force routing.
Open evidence and method →
reference only

Evolution strategies

The objective cannot provide useful gradients or the model/policy is treated as a black box.

Lab
Simulate one two-parameter ES update on a deterministic toy reward.
Mastery
Explain the evaluation-cost tradeoff versus backpropagation.
Open evidence and method →
validated with caveat

GaLore and training stability

Optimizer state or unstable gradients block an otherwise valid training recipe.

Lab
Compare the recorded GaLore smoke with the gradient-checkpointing result and identify what each actually saves.
Mastery
Distinguish optimizer-state memory from activation memory.
Open evidence and method →
reference only

Interpretability probes

Aggregate scores cannot explain which layers, tokens, or representations drive a failure.

Lab
Inspect one logit-lens or activation-patching fixture and write the causal claim it does and does not support.
Mastery
Separate correlation, localization, and causal intervention.
Open evidence and method →
validated with caveat

LoRA, DoRA, and adapter geometry

The base misses a bounded task and full fine-tuning is unnecessary or too costly.

Lab
Run or inspect LoRA geometry for one successful and one failed adapter.
Mastery
Explain rank, target modules, effective rank, and why lower loss is insufficient.
Open evidence and method →
validated with caveat

Sparse mixture of experts

A dense model lacks capacity and a true sparse execution path is available.

Lab
Calculate active versus total parameters for top-k expert routing and identify why dense execution saves nothing.
Mastery
Explain the difference between sparse parameters and sparse compute.
Open evidence and method →
reference only

Multi-token prediction

Single-token supervision or decode throughput is the measured bottleneck.

Lab
Draw the label shift for two future-token heads and compute the added loss terms.
Mastery
Explain why training more heads does not guarantee faster decoding.
Open evidence and method →
validated with caveat

Optimizer and schedule selection

Loss spikes, under-updating, or optimizer memory blocks a frozen recipe.

Lab
Diagnose three recorded loss curves as learning-rate, capacity, or data failures.
Mastery
Choose an optimizer change from evidence rather than novelty.
Open evidence and method →
validated with caveat

Parameter-efficient fine-tuning variants

Plain LoRA has a measured geometry, capacity, quantization, or many-adapter limitation.

Lab
Create a one-page comparison of LoRA, DoRA, QLoRA, and VeRA parameter/memory assumptions.
Mastery
Select a PEFT variant from the bottleneck, not from its name.
Open evidence and method →
validated with caveat

Precision and numerical gates

Precision choice causes overflow, drift, excess memory, or slow execution.

Lab
Measure fp32/fp16 error on a tiny matrix operation and choose a tolerance before looking at the result.
Mastery
Explain why a faster kernel with wrong numerics is not an optimization.
Open evidence and method →
validated with caveat

Unstructured and structured pruning

Artifact size or dense topology is the measured deployment bottleneck.

Lab
Compare 50% zero weights with physical layer removal and predict which metrics change.
Mastery
Distinguish statistical sparsity, compressed storage, and runtime sparsity.
Open evidence and method →
validated with caveat

Quantization and storage formats

Full precision exceeds the target's distribution or memory budget.

Lab
Trace one fp16-to-int4 artifact from calibration through load and decode measurement.
Mastery
Explain why four-bit storage can still execute with wider arithmetic.
Open evidence and method →
validated with caveat

Speculative heads and decoding

Autoregressive decode latency is the measured user-visible bottleneck.

Lab
Compute expected speedup from draft length, acceptance rate, and verification cost.
Mastery
Identify when speculative decoding becomes slower than ordinary decoding.
Open evidence and method →
validated with caveat

StreamingLLM and KIVI cache compression

KV cache growth, not model weights, is the measured memory blocker.

Lab
Calculate full, windowed, int8, and int4 KV bytes for one model/context.
Mastery
Choose between eviction and quantization from the actual context failure mode.
Open evidence and method →
closed lineage

SQL specialist lineage

Small-model SQL gains fail to generalize across public and synthetic domains or emit prose wrappers.

Lab
Reconstruct the SQL lineage and mark which variable changed at each step.
Mastery
Explain why routing beat blending and why DPO failed to remove the wrapper.
Open evidence and method →
closed experiment

Needle 45M successor factorial

Tiny overfit can pass while held-out selection remains below stock and destructive confirmation behavior fails completely.

Lab
Reconstruct the three factorial effects, apply the logical early-stop rule, and design the next 1.7B capacity test without changing the public gate.
Mastery
Explain why 100% tiny overfit coexists with a conclusive failed promotion and why sealed V2 stayed closed.
Open evidence and method →
closed experiment

Parakeet WGSL browser ASR

Browser audio ingestion may be too slow, too large, nondeterministic, or network-dependent.

Lab
Re-score the two prediction files, identify which WhisperKit errors come from numeric normalization, and compare aggregate with median per-clip RTFx.
Mastery
Explain why 0.0% WER and a 3.51x paired latency win can coexist with a correct reject decision under the frozen 50x gate.
Open evidence and method →

Technique registry