Transparency

Summaries are generated by a local model, then filtered by deterministic rules before anything publishes. Prompts ask for restraint; these rules enforce it. Published in full — a tool that demands transparency from institutions owes the same about itself.

• Name scrub. Published AI summaries of scanner audio never contain a personal name. Any sentence naming someone is deleted before publication; people appear only as “a subject”, “the caller”, “a juvenile”.
• Casualty guard. A summary cannot claim anyone was injured or killed unless the transcript itself plainly says so. Unsupported casualty sentences are deleted.
• Escalation guard. Claims of a stand-off, barricade, hostage, pursuit, weapons, explosion, or “dangerous/critical condition” are deleted unless the transcript contains matching language. Mundane radio traffic must read as mundane.
• Radio-code definition ban. The model may never assert what a numeric radio code “means” or “indicates”. Only codes decoded by the local Sangamon/Illinois code table are explained; everything else has no known meaning here.
• Scrub-to-fallback. If the guards delete a summary's substance, the whole blurb is replaced with an honest “details are unclear from the audio” — never published half-hallucinated.
• Merge veto. The AI's decision to thread a follow-up transmission into an existing incident is vetoed in code when the two mention contradictory places. Closeness in time alone never links two events, and vague housekeeping traffic (“updated location”, “radio check”) never becomes an incident.

Every record carries a “dispute this record” button and people named here can request removal — anonymously. Every resolution publishes, including rejections and why.