Facts use rules and semantic review
Every new Tibo post is semantically classified. Other official accounts first pass a broad relevance guard. The model separates completed hard resets, future promises, banked resets, limit changes, incidents, retrospective mentions, negations, and unrelated posts.
A fact is confirmed only when the source is authorized, Codex scope matches, the event is a completed hard reset, confidence clears the threshold, and every evidence span is an exact substring of the public post.
Conflicts fail safely
Rules remain guards for source, product scope, time and explicit grammar. When AI understands a novel sentence that the rules do not cover, a second model must agree before confirmation. Conflicts and ambiguity remain review context and cannot replace the last-known-good fact.
If AI is unavailable, only complete high-precision language can confirm. Metaphors, jokes, and unfamiliar phrasing remain pending and are retried by a later scan.
Forecasts are deterministic
AI never writes the displayed percentage. Score v3 combines a historical baseline, elapsed time, reset cooldown, source authority, explicit promises or negations, semantic agreement, freshness, and a capped contextual adjustment.
Every displayed factor participates in the final calculation. The published history records which factor and source URL moved the forecast during the previous seven days.