twelve stages, one manifest, one ship gate
01 / 02 Tokenizer
vocab, merges, hash
~1 hr · ~$5
03 Dataset
shards, dedup, rows
~hours · ~$50
04 / 05 Pre-training
base checkpoint
days · $$$
06 SFT
instruction tuned
~hours · $$
07 RLHF (PPO)
reward + policy
parallel w/ 08
08 DPO
preference pairs
parallel w/ 07
09 CAI + GRPO
self-improvement
rule rewards
10 Evaluation
MMLU, MATH, GPQA
held-out set
11 Quantization
GPTQ / AWQ / GGUF
4-bit · ~mins
12 Inference
vLLM, TRT-LLM
continuous batch
Ship gate
regression + KL + $
numeric thresholds
Artifact store
content-addressed
hash, not filename
Manifest
version, seed, git_commit
per-stage input/output hash
cost, wall clock, status
single file, sufficient to replay
Gates
regression: metric ≥ baseline
KL budget: drift < cap
safety: refusal rate < 5%
cost: total < budget
no subjective sign-off — every gate is a number, and every gate is versioned