AI you can check.

A diplomatic cable arrives in the inbox: 9 paragraphs, 612 words, 4 referenced annexes. No structure, no provenance markers, no temporal order — and somewhere inside it, one quiet contradiction of last week’s reporting that a tired reader will miss.

The engine reads the prose and proposes typed records: actors, claims, denials, commitments, escalation signals — every one bound to its exact source span with a confidence score and a hash-addressed receipt. Disagreement is preserved as data. Nothing is promoted yet; these are proposals awaiting a named human gate.

Every claim gets two timestamps — when it happened, when this desk learned it — so the analysis can never quietly rewrite its own past. Episodes lock into Allen-13 interval relations: the offensive OVERLAPS the talks; the rejection MEETS the deadline. “What did we know on the 14th?” becomes a query, not an argument.

An expert promotes, caveats, or rejects each proposal — decisions recorded forever. The compiler then assembles the capsule deterministically: evidence at the bottom, claims with trust tiers, the graph, both clocks, the reviewer’s open questions — and the contradiction, kept as an explicit red edge. Same inputs, same bytes, one hash.

With the capsule attached, PRAXIS drafts the SITREP under its contract: T1 asserted, T2 attributed, T3 hedged, the dispute surfaced in line. The margin names which expert device shaped each judgment. Every paragraph carries receipts; every open question is visible. The analyst argues with the model — not the other way round.