> ## Documentation Index > Fetch the complete documentation index at: https://docs.while.ai/llms.txt > Use this file to discover all available pages before exploring further. > ## Agent Instructions > Install with `uv add whileai`; import as `import whileai as wai`. > Run the offline path first (`simulator=False`, `wai.seeded_agent`, a callable judge); no key is needed for it. > Report every pass rate with its interval and n, as `scored.pass_at` prints it. # Constitution > What whileai is, what we believe, and where each belief is enforced in the code. Repeatable science, replicated papers as proof, your keys, ergonomics as the product. This page mirrors [`CONSTITUTION.md`](https://github.com/whilehq/whileai-sdk/blob/main/CONSTITUTION.md) at the repository root, which is the canonical copy; a test keeps the two identical. The three agents that apply it on demand live in [`.claude/agents/`](https://github.com/whilehq/whileai-sdk/tree/main/.claude/agents): `style-guide` (the SDK surface against the [style guide](/reference/style) and the book), `docs-designer` (these docs and the website, with a designer's page contract), `user-sim` (a researcher running a recipe on their own keys and compute, filing what got in the way). What `whileai` is, what we believe, and how that shows up in the code. Read it before you add a public name, write a page, or run a recipe. The routines that maintain this repo read it too. ## What we are `whileai` is a scientific post-training library for language models: SFT and RL, on open models, with the measurement that says whether training helped. Simulate, grade, measure with intervals, select, train, prove on a held-out set, serve, and feed the new traces back in. Build self-improving systems. It is for AI researchers, ML engineers and applied-AI developers, and the goal is that it sits in every applied-AI and research department the way PyTorch does. The platform (`whileai.platform`) is a separate, optional service for hosted training and serving. The library needs no account. ## 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](https://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 | Belief | Where it is enforced | | ------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Repeatable science | `recipes/papers/check.py` refuses "moved" without an interval that excludes zero, three base re-runs, a clean holdout, and a proxy-vs-target verdict | | Replicated papers | `recipes/papers/README.md`: one paper, one recipe, one command, one `post.md` | | Sourced defaults | `scripts/check_no_hardcoding.py` in CI; `tests/api/test_readme_defaults.py` | | Ergonomics | `docs/reference/style.md`; `tests/api/test_style_ratchet.py` pins the retired shapes | | Docs order | `docs/` on Mintlify; the docs routine's one PR a day; the five-line loop list on the website home (`components/quickstart.tsx` in whilehq/website) is the reference wording | | README skeleton | `tests/api/test_readme_skeleton.py` pins the section order, the five-line loop before the first heading, and the 900-word prose budget; `tests/api/test_readme_offline_quickstart.py` runs the quick start as written | | Bring your own keys | `wai.configure(agent=, judge=, api_key=)`, backend objects whose repr names the key source; the Modal and Prime Intellect researcher routines run on their own accounts twice a day | | One name | `scripts/check_old_name.py` in CI lint pins the count of the old name per file from `scripts/old_name_baseline.json`; the docs, site and style routines fix any old-name string in a file they touch | | One counter | `.github/scripts/check_version.py` fails any version that is not the last published counter plus one, or that is not its own PEP 440 normal form; `release.py` only cuts N+1 and refuses to start from a padded or rolled-over version; `tests/api/test_version_string.py` pins `wai.__version__` to the top `CHANGELOG.md` heading; `RELEASING.md` is the long form | ## Who reads this People: contributors, before their first public name. Agents: the style guide routine, the docs and site routines, and the Modal researcher routine (the Prime Intellect one when it has a key), at the top of every run. When this file and another file disagree, this file wins and the other file gets a PR.