restatement Check

How much of a post restates its sources, in code. The second half of the explainer gate, beside the helpful-content check. Ported from the source site's restatement script (10 October 2026), built because a draft explainer that retold one vendor announcement had passed every helpful floor and was discarded by a person the day before. Measured against the pages the writer read.

Copied share of the post's eight-word runs found verbatim in a source paraphrased share of the post's sentences whose content words are at least 60% one source sentence's (a reworded sentence) lead the same, counting only the single source most sentences come from (a retelling of one announcement) numbers of the post's figures found in the sources, the share found in only one of them (reported, never gated)

Code blocks are left out on both sides. A documented request shown as an example is fair.

Calibration (source site, 10 October 2026, thirteen explainers a person approved and the one discarded). None of these measures separated the discarded draft from the approved ones. It sat mid-range on every one (copied 0.015, paraphrased 0.17, lead 0.08; approved up to 0.058, 0.31 and 0.28). Its fault was adding nothing, not copying.

So each limit is the highest any approved explainer reached plus half again, and the check catches a post that copies or rewords its sources more than any approved explainer did. A post that retells one announcement in fresh words still passes; that stays a person's call. An install that calibrates on its own approved posts sets site.json restatement.limits.

Two ways in. The writer agent measures every draft against the pages it read before the draft can be proposed or published (lib/agents/writer/run.mjs). As a check it is opt-in (--only restatement, or site.json checks.include), because it fetches every page a post lists under "## Sources"; fetched pages are cached in .heddle-cache/restatement so a run reads each once a day.