# Host validation and review

Akera recomputes deterministic checks from the candidate, prepared transcripts, and acknowledged context. Producer claims that a check passed do not count.

Hard checks include:

- exact workflow, skill, context, source, and release-receipt bindings;
- one complete coverage ledger per claimed source;
- exact evidence-class inventories and the configured coverage floor;
- one-to-one decision/article mapping, valid supporting segments, and unique slugs;
- 500–800 words per article;
- claim and first-person evidence binding;
- full paragraph-to-claim ledger coverage so uncited prose cannot bypass evidence review;
- exact audience, pillar, offer, acquisition intent, destination, and archetype binding;
- an actual Voice-kind revision with the exact acknowledged SHA-256;
- banned mechanics and phrases;
- estimated grade level at or below 8.0 and reading ease at or above 55;
- pairwise title/claim similarity no higher than 0.70.

Reading ease of 60 is an editorial target and warning until the approved fixture set proves it can be a stable hard gate. Any failed hard check fails the whole batch. Fix the reported rule IDs; do not delete a difficult source or pad a weak article to make the batch green.

Peer review is independent and hash-bound. The reviewer scores extraction depth (25), source fidelity (25), audience/offer fit (20), article quality (20), and batch distinctness (10). Acceptance requires at least 85/100 and no automatic failure.

Every rejection supplies `target`, `criterion`, observed `evidence`, and an observable `acceptanceCondition`. The same producer revises the exact rejected candidate. Three rounds is the ceiling; the third non-acceptance is terminal.

Peer acceptance only seals the immutable article and makes it eligible for human Gate 2. It does not approve, multiply, schedule, or publish the article.