01
Construct
Canonical data
Chat records, embedding pairs, trajectories, provenance, rights, and split groups.
finetuning
open-source infrastructure for reproducible model training
Fine-tune models.
Keep the entire story.
Build datasets, distill teacher behavior, train, evaluate, and export—with provenance, resume identity, and artifact hashes intact from beginning to end.
source records → training workflow → verified artifact
every stage leaves a receiptTHE PROBLEM
A training script is easy. The hard part is remembering what actually happened.
Which records went in? Which prompt shaped them? Did related examples leak across the split? Which model generated a candidate, which model judged it, and did resume quietly buy the same provider response twice?
finetuning turns those questions into contracts. Every stage hands the next one a stable identity and an inspectable receipt instead of another unlabeled directory.
THE WORKFLOW
One path from behavior to proof.
TypeScript owns data and orchestration. Python owns model execution. Versioned contracts keep the boundary honest.
02
Distill
Teacher candidates
Generate responses, vectors, scores, or rankings—with separate judges and explicit budgets.
03
Train
Bound recipes
Pin datasets, templates, model revisions, objectives, and every field that makes resume honest.
04
Evaluate
Comparable evidence
Keep held-out answers held out. Compare only reports that share compatible identities.
05
Export
Verifiable artifacts
Ship model files with relative paths, byte sizes, SHA-256 hashes, and the receipt that produced them.
FULL TOOL TRAJECTORIES
The row does not stop at the tool call.
Choosing check_availability is only half the behavior. The student also needs to read the result, preserve the user’s constraints, and close the loop naturally.
WHAT IT PRESERVES
The machinery around the model.
The product is not another wrapper around a trainer. It is the control plane between every step.
Full tool trajectories
Train the call, the result, and the final answer—not merely the decision to use a tool.
Task-native records
Retrieval, similarity, classification, and clustering keep their own semantics instead of sharing a vague universal schema.
Fail-closed execution
Network, spend, downloads, remote code, uploads, GPU paths, and mutation require explicit authorization.
Identity-aware resume
A checkpoint resumes only when the immutable dataset and recipe identity still match.
Evidence before claims
Production recipes remain unavailable until license, hardware, reload, and evaluation gates have passed.
A CLI that explains itself
Predictable noun-and-verb commands, JSON on stdout, diagnostics on stderr, and dry-runs before writes.
01 / CHAT
Teach behavior.
Generate conversations, complete tool trajectories, response candidates, judge decisions, and canonical SFT data.
- Scenario-driven self-play
- OpenAI and Anthropic teachers
- Separate generation and judging budgets
- Held-out answer locks
!
HONEST BY DEFAULT
A credential is not permission.
Provider calls require an exact model, an environment-variable reference, network opt-in, and a positive budget.
Downloads, uploads, GPU recipes, remote code, overwrites, and live infrastructure mutation stay explicit—or unavailable.
sha256: evidence-over-vibes
Build the model.
Keep the history.
Start entirely offline with deterministic fixtures. No credentials, network, model download, GPU, upload, or remote code required.