Methodology

Last updated: September 29, 2026 · v1.4

How an Apprised.news brief is built, what it is not, and how to read each section. This page is versioned and dated. When the methodology changes, this page changes and the version increments; the changelog at the foot of the page records what changed.

One sentence

Seventy-three AI analyst personas across nine named-persona desks read the day's news corpus from Corvus Intel (800+ RSS sources) and produce structured briefs with required dissent, bias checks, presidential back-tests, and historical power lenses.

The corpus

Apprised.news does not crawl the open web directly. It consumes the corpus assembled by Corvus Intel, a separate aggregation pipeline. Apprised.news is published by J.A. Watte (jwatte.com), who also owns and operates Corvus Intel; the relationship is disclosed on the ethics page. Corvus reads roughly 800 RSS and Atom feeds across U.S. wire services, regional U.S. press, allied-Western press, non-Western regional press, state-aligned outlets, Reddit, Substack, financial newsletters, primary-source government feeds, and prediction-market endpoints.

U.S. coverage is weighted first. International coverage is included but framed through U.S. domestic impact. The corpus is rebuilt continuously and snapshotted at brief-generation time.

Apprised does not pay sources, does not have privileged access to any source, and does not embargo or pre-publish anything. Every fact in a brief should be traceable through a citation to a public article.

The desks

Nine named-persona editorial desks consume the corpus, each with a fixed cast of named AI personas and a fixed output schema; together they hold the seventy-three personas. Two further desks are derived and carry no named-persona roster — World (OSINT roundtable) and Local Wire (state-press cross-validation). The Politics desk runs three archetype voices that sit outside the seventy-three. Crypto is a market-data view.

The Politics desk (internally pressure) is not one of the nine named-persona desks. It tracks vote predictions, statement-vs-market gaps, and constituent alignment for U.S. federal legislation through three archetype voices — Whip Count, Constituent Impact, and Statement-vs-Vote Gap — that are not part of the seventy-three-persona roster. It pulls political data from Congress.gov, Kalshi, and Polymarket. Each day's brief lives at /desk/pressure, and the wider picture at Politics Landscape.

The personas

The personas are stylistic framings. Each one names a publicly known analytical tradition (Friedman-school structural geopolitics, Zeihan-school demographics, El-Erian-school macro, Alden-school monetary, Marks-school cycle psychology, Damodaran-school valuation, and so on) and writes from inside that tradition. The personas are not real people. They are not impersonations. They are recurring voices that hold consistent positions over time so that a reader can track how each tradition reads the day.

The model behind the seventy-three personas is currently Anthropic's Claude; the Politics desk's three voices and the Local Wire report are written by Moonshot AI's Kimi. The personas are constructed via system-prompt scaffolding, not fine-tuning.

Per-persona pages with bios, recent takes, and agreement history live under /analyst/.

Required dissent

Every Intelligence Desk brief includes a dissent field on every Tier 1 persona. The persona must articulate where they disagree with at least one other named voice in the same roundtable. If no genuine disagreement exists, the persona must say so explicitly. This forces structured friction into the output and prevents the desks from collapsing into a single consensus voice.

Council desks (Markets, Defense, Energy, Tech, Health, Culture, Sports, Insurance) carry the same constraint through synthesis.points_of_disagreement. The World Desk surfaces disagreement as narrative collisions across source families.

The cross-cutting Dissent view aggregates every disagreement signal across all desks for the day.

Back-tests: presidential and power lenses

The Intelligence Desk applies presidential lenses: 3 to 5 of nine U.S. presidents (Washington, Lincoln, T. Roosevelt, F.D. Roosevelt, Eisenhower, Kennedy, Nixon, Reagan, Obama) are selected each day and asked to read the day's signals through their documented decision-making frameworks.

All desks apply historical power lenses: 3 to 5 of fifteen historical figures (Napoleon, Nero, Genghis Khan, Cleopatra, Catherine the Great, Machiavelli, Alexander, Elizabeth I, Sun Tzu, Caesar, Hearst, Morgan, Carnegie, Bell, Edison) are selected and asked to read the day's signals through their documented strategic frameworks.

Back-tests are interpretive exercises, not endorsements. The lens of Hearst does not endorse Hearst's tabloid methods; it asks what Hearst's documented use of media as power would notice in today's news.

Roster rotation (added 25 July 2026). The power lenses are no longer offered as one fixed list. Each desk is shown a rotating subset of seven of the fifteen, sequential in the date and offset per desk. Two things forced this. The council desks' hard-coded roster listed only eleven figures and omitted Nero, Catherine the Great, Alexander and Elizabeth I outright, so those four could never be selected outside the Intelligence Desk. And free choice from a fixed list is salience-ranked: measured across 109 lens slots, four figures took 70% of them and Sun Tzu appeared on all nine desks. The rotation offers every figure equally often and stops any one figure occupying every desk on the same day.

Standing doctrines: the contemporary lens

The presidential and power lenses both look backwards. As of 25 July 2026 the Intelligence Desk also applies standing doctrines: two to three of six live, contested schools of thought — the Brussels Effect, State-Capital Fusion, Integrated Deterrence, the Precautionary Principle, the Energy Trilemma, and Shareholder Primacy vs Stakeholder Capitalism. Each names its institutional lineage, reads the day's specific stories through that framework, and must state where the doctrine and the desk's own read part company — a doctrine entry that declines to take a side is a failed entry.

Why doctrines and not living analysts. A contemporary lens could have been built from named living commentators. We did not do that, deliberately. Our ethics policy sets a no-impersonation floor for living people, and putting invented words in a real person's mouth every day would cross it. A doctrine has no mouth: it is a documented institutional position held by many and disputed by many, so attributing a reading to it is a claim about a school of thought rather than about an individual. Doctrine entries never name a living person as the source of a view.

Persona vs. lens overlap. A few historical figures appear in two roles. Sun Tzu is both one of the fifteen power lenses and the standing Intelligence Desk persona Sun Tzu Assessment; Otto von Bismarck lends his name to the standing persona Bismarck Frame but is not one of the fifteen power lenses. In every case the persona is a recurring voice that authors briefs, while the power lens is a rotating one-off back-test — separate constructs that share a historical namesake.

Threat level

The Intelligence Desk emits a single global threat level each day:

Calibration scorecard

Predictions, threat-level moves, and vote forecasts are scored against settled outcomes. The Track Record view shows per-day verdicts (HIT, PARTIAL, MISS, NULL_COVERAGE, OPEN) and a 30-day calibration breakdown by stated confidence (HIGH, MODERATE, LOW). A well-calibrated desk's HIGH-confidence calls should resolve to HIT at a higher rate than its MODERATE calls.

NULL_COVERAGE means the desk made no predictive claim on that surface for that day. It is not a miss.

The AI Disagreement Index

The briefs Claude writes are independently cross-checked by a second model, Kimi. The AI Disagreement Index turns that cross-check into a single dated, citable number: across the day's contested events that both models rated, how often did the two assign a different call on how settled the facts are?

This is agreement across models, not across outlets. It is the companion to the Consensus Index (which measures corroboration across sources using one model); this one holds two independent models against the same events.

How it is computed — pure read-side aggregation, with no extra AI calls:

• Matched events. On the World Desk, Claude records each contested event in its narrative collisions with a certainty call (Consensus, Contested, or Developing). Kimi, reading the same corpus, records its own certainty call per event. Each Claude event is paired with Kimi's closest event.
• Distinctive-token pairing. Events are paired only when they share at least two distinctive terms — proper nouns and specifics, not generic words like “strike,” “talks,” or “missile,” which are deliberately ignored because they otherwise pair unrelated stories. An event with no confident match is discarded, never guessed.
• Settled vs unsettled. Both calls collapse to two classes — settled (Consensus) versus unsettled (Contested or Developing) — so an adjacent grade like Developing-vs-Contested is not counted as a real disagreement.
• 30-day window with a minimum-sample gate. Daily samples are only a handful of matched events, so the headline is a rolling 30-day rate, and the page shows no number until at least a dozen events have been matched.

An honest caveat: the matching is heuristic. The index page publishes its own match confidence — the median number of shared distinctive terms and the share of pairs sharing three or more — so the fuzziness is visible rather than hidden. To see exactly which stories the two models split on, the AI Disagreement Index lists each flagged pair with a link to that day's World Desk brief, where Claude's collision calls and Kimi's independent reads render side by side with their source outlets — so you can follow any disagreement back to the underlying coverage. Live data: /api/disagreement-index.

The Bias Risk Ledger

Before a desk brief publishes, the same independent model that did not write it (Kimi) reads a condensed version and grades its bias risk — political lean, source-diversity skew, loaded framing — as LOW, MODERATE, or HIGH, with specific flags. The Bias Risk Ledger is the public rolling tally of those pre-publish verdicts across the eight audited desks, broken out per desk.

It is a deterministic count of labels Kimi already wrote at publish time — no new judgment is added when the ledger is built, and it costs no additional AI calls. The ledger surfaces the MODERATE and HIGH counts and the most common flags prominently, on purpose: a record that was 100% clean would mean the audit was theater rather than a real check. Briefs that get flagged are part of the record, not hidden from it.

The headline — the share of audited briefs that cleared as LOW over the last 30 days — renders only after enough briefs have been audited, so a thin first week cannot over-claim. The Intelligence Desk and the cross-desk intelligence report are not in this ledger; they carry their own bias-check and presidential-lens structure instead, so they correctly fall out rather than being counted. Live data: /api/bias-ledger.

Cadences

Every desk supports five cadences:

• Daily — built every day from the prior 24h corpus.
• Weekly — built on Mondays UTC from the prior 7 days.
• Monthly — built on the first of each month UTC from the prior 30 days.
• Quarterly — built on Jan/Apr/Jul/Oct 1 UTC.
• Yearly — built on Jan 1 UTC.

Most briefs are generated in one run through the Anthropic Message Batches API at about 04:00 UTC. The Local Wire report is generated on Kimi at about 03:00 UTC, with an Anthropic batch as its fallback, and the Politics desk on Kimi at about 06:00 UTC. The Intelligence Desk brief and the Daily Intelligence Report can be regenerated once more at about 13:00 UTC, and only when the news corpus has moved enough since the morning run. A watchdog probes every desk five times a day and re-triggers generation for any that is stale.

What a brief is not

Errors and corrections

When a brief is materially wrong, we issue a correction on the corrections page with the date, the original statement, the corrected statement, the reason, and the supporting source. Inline corrections also appear on the affected brief permalink. A brief that is regenerated after publication, such as the midday refresh, keeps its URL and carries a dated “Updated” notice. The corrections page lists every status label. See the ethics page for the full correction policy.

Open questions

This methodology is not finished. Live deliberations include: how to surface persona-level track records in real time, whether to expose the system prompts directly, how to integrate primary documents (executive orders, UN resolutions, SEC filings) alongside news citations, and how to weight prediction-market signals against analyst opinion when the two diverge.

Changelog

• v1.4 · 29 September 2026. Says which model writes what: Claude writes most desk briefs and the Daily Intelligence Report, and Kimi writes the Politics desk brief and the Local Wire report, with Claude as the fallback. The previous version said every brief was written by Claude and that Kimi never wrote a published brief, which stopped being true when those two moved to Kimi in July and August 2026. Also: nine named-persona desks, not eight; Markets has twelve voices, not eleven; the Politics desk gets its own row; the cadence paragraph now describes the real schedule, including the midday refresh; corrections now record the original and corrected statement.
• v1.0 to v1.3 · 16 May to 3 June 2026. The changes between these versions, and the additions dated 25 July 2026 above, were not logged individually. From v1.4 every change is logged here.