The truth standard

Every module carries its evidence with it: its sources, who reviewed it, what the red team found, what has been corrected since. A reader can check all of it. tools/check_modules.py enforces the parts a machine can check, and CI runs it on every push.

The rules

  1. No citation, no publish. Every factual claim cites a source in the module's sources.yml: a paper, a standard, official documentation or a primary record. Secondary write-ups are only for context, never the sole support for a claim.
  2. A separate agent checks every source. The source checker opens each source and confirms it says what the sentence claims. It never checks a module it helped write. Its report lives in the module's checks/source-check.md, and each source records the date it was checked.
  3. A red team tries to break it. A different system from the writer attacks every claim, example, chart and quiz answer. Findings go to the human red-team lead, who accepts or rejects each one. A module cannot go to review with an accepted finding still open.
  4. Two people review every module. One subject expert and one audience reader: someone at the level the module is for. Reviewers read the objectives and the standard gear in full, then check the plain and technical gears for drift.
  5. Agents never count as reviewers. Agents draft, check and attack. Only people sign. review.yml names each reviewer, and the checks refuse any reviewer recorded as an agent.
  6. The board signs each level. The review board (co-chaired by Bob and Ken) signs a level once every module in it has passed review. A module publishes only after its level is signed.
  7. Every module is versioned and has an errata log. A correction bumps the version, is logged in ERRATA.md with the date and what changed, and is re-signed by one reviewer for the changed passage only.
  8. Fast-moving facts carry an as-of date. Model names, prices and limits are dated in the text. The freshness agent re-checks modules weekly and opens an erratum when a fact has moved.
  9. Every figure and report names its runner. Each figures.yml entry carries a runner, and each source-check, red-team and art red-team report carries a Runner: line, naming the agent and session that produced it. Rules 2 and 3 ask for a separate agent; until 21 Nov 2026 nothing in the repository recorded which agent had run, so a report's independence was only a sentence it wrote about itself. The checks confirm the runner is recorded. They cannot confirm it is true. Work from before the rule reads "not recorded", and stays that way rather than being guessed (Bob Rapp's ruling, design/decisions-2026-11-21-group-b.md).
  10. Option length gives nothing away. In a module's quiz, the keyed answer is the longest option in no more than one-third of items, and the shortest in no more than one-third, counting ties and measuring by characters. A learner who always picks the longest answer, or always the shortest, should score about what guessing scores. The check warns today and does not yet block: on 21 Nov 2026 it fired on 26 published modules, whose rebalance is planned work (docs/reviews/2026-11-21-option-length.md; Bob Rapp's ruling, design/decisions-2026-11-21-group-e.md).

Module status

StatusMeansThe checks require
draftBeing writtenAll files present; quiz items map to objectives; every citation resolves to a source
in-reviewWith the two reviewersThe above, every source checked, a red-team report with no open accepted findings, at least five quiz items
publishedLiveThe above, two human reviewers signed, the level's board sign-off, an errata log

What the public stamp shows

Each published module shows its version, review date, the two reviewers' names and roles, the red-team lead, the board sign-off date, and a link to its errata log.