Constitution

​

What we are

​

What we believe

1. Repeatable science. A number is a result only with its interval, its noise floor, its seed and the versions that produced it. A mean alone is not a result. A flat result is a result. (pass_at, eval_variance, delta_report, holdout_size.)
2. Replicated papers are the proof. We show the library works by reproducing recent post-training research in it, one recipe per paper, under an hour on one GPU, with the number it moved and the number it did not. Every reproduced paper is a post. The proof point is the recipe, not the pitch. (recipes/papers/.)
3. The book is the map, the paper is the citation. Every default is named, sourced and tunable from the call. rlhfbook.com (Lambert) is the map of the field; the originating paper is the reference. A default with no source says “convention, untested”. (defaults.py, scripts/check_no_hardcoding.py.)
4. Bring your own keys. Your models, your compute, your accounts. Modal and Prime Intellect are first-class: a whileai environment becomes a verifiers environment and back, selected rows become a trainer’s prompt set, eval results flow back into measurement with intervals. Nothing in the loop requires our hosting.
5. Developer ergonomics are the product. The code reads like PyTorch, DSPy and Unsloth: one import, objects carry configuration, calls carry data, reports print themselves, errors name the fix, and a first-time reader can guess the next line. Rigor lives behind a default, never behind a flag. (docs/reference/style.md, the ratchet test.)
6. Plain words, then the mechanism, then the proof. Every page, every docstring, every README section in that order. The first thing a reader sees is the loop as five lines, one per step, each the step’s name and one sentence saying why the step exists in the reader’s own words (“Measure. One run proves nothing. Ask whether a change is real or noise before you ship it or train on it.”). Nothing goes in front of that list: a manufactured hook (“Break it. Score it. Prove it.”) and a paragraph about the problem both read worse than the list, and were cut the same day they shipped. Book vocabulary (interval, rollout, band, gradient) stays in the docstring that cites the chapter, never in a lead sentence or a public name. The step names are Simulate, Grade, Measure, Select, Train. Measure is evaluation with intervals and Select is data curation; they are not merged, and neither is renamed to Eval. The README is a front page, not a paper. It follows the skeleton the most-used Python repos share (Polars, TRL, vLLM, uv, decomposed 2026-09-19): logo, one-line tagline, badges, link bar; the five-line loop; Install; Quick start, offline, with its printed output; With your agent; the call table; Why the numbers hold, as bullets of term, mechanism, citation; Recipes; Platform; Documentation; Development; Cite; References, collapsed; License. Prose outside code, tables and the references stays under 900 words. A code block carries the explanation wherever one can; anything longer than a paragraph moves to docs/.
7. Mass experimentation. A PhD or an engineer runs many experiments from one import, on their own compute, and every run leaves a record that a person can decide from.
8. Never big-bang. The internals carry the science and the tests. Change the front door, migrate callers mechanically, keep the old name working for one release with a warning that says the new one.
9. One name. The company is While, the package is whileai, the import is import whileai as wai, the command is wai (whileai runs the same entry point; zp is gone), the hosts are while.ai (site and platform), api.while.ai and docs.while.ai, the variables are WHILEAI_*, the config dir is ~/.whileai. app.withwhile.com was retired the same day; nothing links to it. ZeroProof was the name before 2026-09-16; the cutover finished on 2026-09-19 and nothing new is written under it: no code, page, prompt, routine, dataset card or post. What still carries the old name is wire protocol and infrastructure that would break users if renamed (the zp_ key prefix, zeroproof.* span attribute keys, Modal app hostnames, volume and table names) and the history in CHANGELOG.md. Those are pinned, not permitted: scripts/old_name_baseline.json counts them per file, a count may fall and never rise, and a new file may not add one. The ZEROPROOF_* variables and ~/.zeroproof are not read.
10. One counter. The version is 0.N. N goes up by one per release and never rolls over, resets or pads: 0.99 then 0.100 then 0.101. There is no 1.0; a number says how many releases came before it, nothing about maturity. PEP 440 drops leading zeros, which is why the counter is never padded (1.07 is 1.7 on PyPI). On 2026-09-20 the bump script rolled 0.99 over to 1.00 and nothing in this file said it could not; those uploads are yanked and re-cut as 0.100 to 0.109.

​

How it shows up

​

Who reads this