newlife

Turn a curiosity into a question that can be settled, freeze what would count as an answer before running anything, and leave a verdict that can be audited from git alone.

You talk to your AI; it runs the commands. Setup is two lines:

python -m pip install newlife
newlife start my-research

Then open my-research in Claude Code, Codex, or any AI tool that reads SKILL.md files, and say:

Explore this with me: your curiosity

NEXT.md in that folder says what happens from here. If you get lost, ask your AI "how does this work" or "where are we?" — the installed newlife skill runs one authoritative status check and answers in your terms.

Requires Python 3.12 or newer and git (the freeze is a git commit; that is what makes the timestamp real). start tells you if git has no identity to sign with, rather than guessing one.

What happens, in your words

1. Explore. Your AI asks one or two honing questions, then walks one edge at a time and keeps a map silently. When several explanations fit, it uses idea-lab's Draw–Attack–Compare–Check reasoning inside the same conversation; you never choose a second workflow. You answer and decide. (exloop and idea-lab ship together in the exloop package; newlife depends on it, so the same install brings both.)
2. Hand off. When a boundary can be stated as "what measurement would make the answer different", the AI says so once. If you agree, a question folder appears with the exploration record inside it.
3. Is it worth asking? The AI asks only what the map cannot supply: who would bet the other way, is the answer settled by design or by running, who changes what they do. Six anchors go into goal.md. One you cannot fill yet means keep exploring or narrow — not a failure.
4. World and pilot. Your simulator goes into verdict.py; a pilot run shows you every quantity a criterion will name. Emitted values are seen; a named withheld unit stays blind only when its value is absent. Failed executions never enter the pilot ledger.
5. Criteria. Every judgement unit is marked seen / blind / mechanical, and at least one is blind — a confirmatory round in which every number was already seen carries no information.
6. Freeze. The one irreversible step; the AI asks you first. Two gates refuse on their own if the goal is unfilled or a criterion was never piloted. Your environment and any downloaded data are pinned by hash at this moment.
7. Execute and close. newlife status first names the one next step from the real repository state. The execution skill runs and verifies the frozen question, diagnoses anomalies without changing its rules, audits it, and delivers the report. H1, H0 or INVALID is computed by the runner, never written by hand. The AI asks separately before committing results/.

Two moments are yours alone: the freeze, and committing the results.

What the record can prove afterwards

newlife audit shows, from commit ancestry rather than dates anyone could type: the registration has a freeze commit; the stamped hash matches it; the criteria were never edited afterwards; every committed output post-dates the freeze; and env.lock and every pinned data file still hash to what was frozen. The runner's own units add the other half: the run reproduces itself byte for byte, the packages installed at run time are exactly the ones recorded, and every criterion proves at run time that it can go red.

A registration whose criteria turn out to be defective is left as INVALID, exactly as it stands, and a new one is opened. Never re-frozen. That record is how a conclusion was reached, not a failure to tidy away.

If you would rather type

newlife init <slug>       newlife pilot <folder>      newlife freeze <folder> [--data FILE...]
newlife run <folder>      newlife status [folder]
newlife audit <folder>
newlife skills install    newlife skills path         newlife blocks

Never git add -A before the freeze: prereg.md must be committed by the freeze itself, and newlife init leaves exactly that file uncommitted for that reason.

Underneath

Optional simulation backends: pip install "newlife[spatio-flux]" (Monod kinetics, dynamic FBA with GLPK, diffusion, particles) or "newlife[process-bigraph]" for the runtime alone; newlife blocks lists what is installed. Third-party processes run unmodified inside a declared contract — a mechanism says which state paths it owns and which effects it may emit, and writing anywhere else is refused by the contract layer, not by your solver. proofroot is installed transitively and supplies RNG stream derivation, canonical serialization and the evidence-tier vocabulary.

Source, design rationale, and the record of every milestone judged so far: https://github.com/hzshen88/newlife