The generator is not allowed to grade itself.
AlertProof treats a notice like a versioned source artifact: extract source-grounded fact candidates, confirm the claims that cannot drift, manage language and channel drafts, then expose rule evidence before an accountable human makes any publication decision.
Evidence moves in one direction
The current build uses a deterministic local analyzer, editable drafts, exact-value rules, English rule-based coverage, and qualified-review flags for other languages. Prepared evaluation datasets cover the interface without a network connection.
Claims become inspectable objects
Each locked fact carries a category, source span, normalized value when applicable, polarity, severity, and mandatory channels. That makes a changed time a checkable failure instead of a vague feeling.
Four signals, transparent rules
URLs, phone numbers, dates, times, timezones, and quantities are detected before drafting and checked after reinsertion.
Normalized dates, numbers, phone digits, URL domains, negations, and modal strength can hard-fail a draft.
Paraphrases can pass, but missing scope, weakened obligations, changed timing, and contradictions return fail or review.
English shows reading ease and grade estimates. Other languages require language-specific review rather than a pretend English score.
Sources ground. Rules check. People decide.
Today, arbitrary source analysis is local and deterministic; custom translations are never fabricated. A future live provider can sit behind the documented server-side boundary, but it is not configured in this build and would never become the sole judge of exact high-risk values.
One word can change the result
Source fact: "Do not use unboiled tap water." A fluent draft that drops "not" can still look polished. The prohibition check treats that polarity loss as a hard failure and points to the exact output span.
Do not use unboiled tap water.
Use unboiled tap water.
What AlertProof does not claim
It does not publish alerts, replace qualified translators or interpreters, certify language accuracy, guarantee safety, or establish official alert authority or system integration.
Reference outputs are prepared evaluation data. Operational adoption requires authorized source integrations, qualified language review, security assessment, accessibility testing, and an accountable approval policy.