Method
The model judges. The code verifies.
WHAT THE INSTRUMENT MEASURES
Prism describes a disclosed body of coverage gathered during a fixed time window. It can say that an account appears in six supplied articles, that two qualify it, or that a particular frozen article does not contain it. It cannot turn prevalence into truth, silence into motive, or a source quotation into independent confirmation of the event being described.
The most important boundary is a division of labour. Models read, compare and judge meaning. Code establishes corpus facts: which article version was read, where a declared quotation occurs, whether identifiers and article rosters resolve, how many supplied articles carry a model-mapped account, and whether every required publication gate completed. Code does not decide whether two accounts agree, whether a framing is fair, or whether an omission is meaningful.
THE EDITORIAL SEQUENCE
57 sources across the political spectrum are read hourly. Each article is stored in full text, classified by genre, and embedded for clustering. A rolling window keeps roughly the most recent three days active.
Articles are grouped into story clusters: one event, gathered from its many tellings. Clustering is deliberately conservative. Merging two different stories poisons a briefing, while failing to merge merely fragments one, so the system prefers the recoverable error.
A model ranks the day’s clusters by editorial value. The selected roster is then frozen: every article, body hash and article version that the edition will use is fixed before synthesis begins.
One frontier correspondent reads each full article and writes a prose source brief: the article’s account, attribution, chronology, uncertainty, framing and distinctive contribution. In a second turn it attaches evidence packets to the material propositions. Code locates every quoted span exactly in that article’s frozen body and returns failures for bounded repair.
A frontier editor receives the complete set of source briefs as prose, not a machine-generated claim table. In one independent judgment it authors the reader’s coverage map and the page’s entire prose layer: the read that opens the page, optional one-sentence notes conducting each movement of the map, and zero to five bounded analysis threads, each welded to at least two map items. Prose characterizes, conducts and connects the map; it never retells it. The editor judges which articles carry each account and which differences matter; code derives broad, clustered and sole-source circulation bands from those memberships. The editor may also distinguish true conflicts, compatible divergences, chronology or referent traps, framing, unevenness, uncertainty, source dependence and a small number of consequential questions the coverage itself makes visible but largely leaves unasked. A sparse correspondent hotdesk is available when a material ambiguity remains.
In parallel with the first draft, a blind atlas model independently separates common and clustered accounts, possible direct conflicts, non-conflict divergences, framing, uneven coverage, source dependencies and unique contributions without seeing the editor’s conclusions. A fresh critic then compares the complete authored bundle with that anatomy and files only concrete fidelity concerns.
The original editor accepts, rejects or qualifies the critic’s concerns and files one complete revision. A separate evidence pass maps material passages in the frozen publication — map and prose alike — back to the already verified source-brief packets; it cannot rewrite either.
Every article then receives one final raw-text audit against the revised publication, including its article rosters and bounded absence judgments. Only major or critical findings can trigger a narrow repair, followed by re-audit of affected articles. Final sanity sees the complete publication, evidence, atlas, critic and audits together.
A run persists atomically only after every selected story completes. New batches can be inspected without becoming the public edition; promotion is a separate, deliberate action.
PUBLICATION GATES
A report is withheld unless every selected article was briefed, every declared source-brief quotation was located in its frozen body, every report receipt resolves to a specific public field or paragraph, and every article completed a fresh final audit. A major or critical audit finding permits one narrow repair and an affected-article recheck. Final sanity may publish only when no load-bearing defect remains. Passing these gates means the evidence chain is complete; it does not mean the editorial judgment is certainly correct.
CURRENT SYSTEM · EDITORIAL-2
Source briefs and the blind atlas run on gpt-5.6-terra and gpt-5.6-terra. Independent synthesis and whole-report review run on gpt-5.6-sol and gpt-5.6-sol. Final article audits run on gpt-5.6-terra. All editorial work currently uses a local Codex harness with saved subscription authentication. Every run freezes its model assignments, prompts, response schemas, corpus hashes and upstream artifact hashes. The roles can change without changing the principles above; the exact newsroom that produced a report remains traceable in its run record.
The next live test is governed by whose selection rule, defect taxonomy and pass gates were fixed before the run. Reader-relevant changes to these contracts are recorded in the public method changelog. Calibration notes report observations; they do not silently promote a canary into validation.
LIMITATIONS, STATED IN ADVANCE
- The map is relative to the roster. Every claim of consensus, divergence, or absence means: within the sources Prism reads. See sources for the roster and its doctrine.
- Judgment can err even when receipts hold. Exact location proves that a source said the quoted words. It cannot prove that Prism’s reading is wise, complete or fair. The atlas, critic and article audits reduce that risk; they do not abolish it.
- Absence is bounded. Prism may say that a supplied frozen article does not contain something after the correspondent and reviewer inspect it. It does not infer that an outlet ignored a story merely because clustering or scraping failed to retrieve a page.
- One edition maps one day. Stories evolve; a briefing is a dated snapshot of the telling, not a living document. Cross-day story threads are designed but deliberately not shipped until they cannot lie.
- Retrieval and clustering shape the map. Paywalls, feed failures, body extraction, publication timing and conservative clustering can leave relevant coverage outside a story. The source list describes what entered the report, not everything that existed.
- Political position is multidimensional. The roster’s spectrum is a disclosed navigation aid, not a universal coordinate system or a reliability score.
OPENNESS
This page publishes the reasoning. The code, prompts and verification contracts are being prepared for open-source release after secrets and private deployment history are removed. Until that release, Prism should not imply that a reader can already inspect a public repository. A map you cannot eventually audit is just another telling.
