/nexus

Nexus Newsroom Recipe Reference

"An article is only as true as its weakest uncited claim."

Grounded article production and audit, run like an editorial newsroom: the writer drafts, the fact-check desk verifies. Every factual claim traces to a cited, trust-tiered source; speculation, inference, and opinion are explicitly labeled or removed; wrong views are hunted by adversarial cross-checking — before the article ships, not after a reader catches it.

Contents

• Overview
• Invocation and Modes
• When to Use Newsroom
• Claim Taxonomy (the core contract)
• Ledger Membership
• Phase Contracts
• Gates
• AUTORUN Chain Template
• Failure Modes Prevented
• Cost and Scale
• Boundaries — vs Neighbors

Overview

Newsroom is a claim-grounding maximization recipe for articles (blog posts, tech articles, announcements, explainers). Where podium maximizes content polish across formats (prose + visuals + slides), newsroom maximizes content truth provenance for a single written artifact — the editorial fact-check-desk discipline: it builds a cited Evidence Ledger before (or against) the text, decomposes the article into atomic claims, verifies each claim's source actually supports it (citation-support, not citation-existence), attacks load-bearing claims with an adversarial skeptic panel, and ships with a named Provenance Report in which no claim is silently unaccounted for.

Two modes, one verification core:

• compose (default when no existing article is supplied) — evidence-first authoring. The Evidence Sweep runs before writing; the writer is ledger-bound: it may assert as fact only what a ledger entry backs, and everything else must be phrased as labeled opinion/hypothesis or dropped. Verification then audits the writer's own output (producer ≠ verifier).
• audit — an existing article is the input. Phase 2 is skipped; the sweep is claim-driven (research what the article asserts), and the remediation loop corrects, re-cites, downgrades, or marks the text.

Key design decisions:

• Evidence Ledger is the single source of truth for facts — schema, trust tiers (T1–T4), and per-claim verification status are inherited from reference/research-grounding.md (§3), with one adaptation: newsroom has no captured artifact, so the authority rule becomes primary-source-authoritative — a claim's ground is its highest-tier source, cross-corroborated; T4 (community/unverified) is never a sufficient ground alone.
• Citation-support over citation-existence — the auditor re-opens each cited source (WebFetch) and confirms it says what the article says it says. A real URL that doesn't support the sentence is a miscited defect, not a pass.
• Producer ≠ verifier (reference/autonomy-quality-protocol.md Q9) — the Phase 3 claim auditor and Phase 4 skeptics are never the Phase 2 writer.
• Wrong views are a first-class target — Phase 4 doesn't only check "is there a source"; it hunts incorrect-despite-cited content: outdated best practices, deprecated APIs presented as current, misread statistics, causal claims a correlational source can't carry, and staleness relative to the publication date.
• Transparency over cosmetic completeness (shared with podium's hard rule) — a claim that cannot be grounded and cannot be sourced by the user ships explicitly marked [UNVERIFIED] or is removed with a log entry; it is never silently kept or silently deleted.

Invocation and Modes

/nexus newsroom "<article goal>"              # compose mode: evidence-first article creation
/nexus newsroom audit <path-or-url>           # audit mode: verify an existing article
/nexus newsroom "<goal>" --risk external      # risk_tier override (routine | external-facing | release-critical)

Mode detection when unstated: an existing article path/URL/pasted text → audit; a topic/goal only → compose.

When to Use Newsroom

Use Newsroom for

• Tech articles / blog posts where factual credibility carries the piece (benchmarks, comparisons, "state of X" surveys, security write-ups)
• Announcements or explainers that make verifiable claims (statistics, version numbers, API behavior, quotes, pricing)
• Fact-checking an existing draft before external publication
• Any article where the user asks for "no speculation", "every claim sourced", or "check this is actually correct"

Do NOT use Newsroom for

• Article authoring with no grounding requirement → tome direct
• Doc + slide unified package with polish maximization → podium (its Phase 4 claim-grounding branch covers grounding within that pipeline)
• Spec-vs-implementation conformance → attest direct
• Standards compliance (WCAG/OWASP/style) → canon direct
• Opinion pieces / personal essays where the value is the subjective take — newsroom would strip the point (still usable in audit mode to label facts vs opinion, if asked)
• Reproducing an external product's surface → clone (its Evidence Ledger serves parity, not prose)

Claim Taxonomy (the core contract)

Every assertion in the article is classified into exactly one class. The classes carry different obligations:

Claim-tolerance contract (fixed in Phase 0): which statement kinds require a source at this risk tier. At external-facing and above, every fact does. At routine, widely-known background facts ("HTTP has status codes") may pass as common-knowledge — but any number, name, date, quote, or comparative/causal statement is never common-knowledge.

Per-claim audit verdicts (Phase 3/4 output vocabulary):

grounded | miscited (source exists but doesn't support the sentence) | unsupported (no source) | contradicted (a higher/equal-tier source says otherwise) | stale (was true; no longer true at publication date) | mislabeled (opinion/speculation phrased as fact)

Ledger Membership

Newsroom is a member of the external-reviewer-to-zero family: its completion oracle is an evaluator's finding set driven to zero, not a rubric score. The shared machinery is _common/FINDING_LEDGER.md; the Evidence Ledger + claim_audit.json are that file's ledger, with claims as the findings. Newsroom's five declaration slots:

C4 holds (FINDING_LEDGER.md §1): the claim set is the article's own assertions — finite, and each has a right answer — so newsroom needs no split oracle, which is why loop ≤ 2 cycles suffices where quell needs 6. The craft axis of prose (structure, altitude, register, redundancy) fails C4 and is not in scope here; it lives in reference/doc-quality-protocol.md W7-W11 as a single-pass gate.

Verdict → disposition. A verdict is what the evaluation found; a disposition is why the claim is closed. Both appear in claim_audit.json.

DOWNGRADED and DELETED (logged) are newsroom's self-dismissal analogue (FINDING_LEDGER.md §6): weakening a claim into vacuity and deleting the inconvenient paragraph both close a finding by moving the standard rather than meeting it. That is why both are logged, both are re-audited, and a hollowed-out thesis is a BLOCK rather than a ship.

Phase Contracts

Phase 0: FRAMING (Nexus internal, 0-1 agents)

Detect mode (compose | audit), scope, audience, and risk_tier. Fix the claim-tolerance contract (§ taxonomy) and the article's load-bearing thesis (the 3-7 claims the piece stands on — these get the ≥2-source + skeptic-panel treatment).

Output: newsroom_charter.yaml (mode, topic/article ref, risk_tier, claim-tolerance contract, load-bearing thesis list, loop budget).

Gate: risk_tier == release-critical → Confirm before launch (conditional confirm — intentional parity with podium, not summit's unconditional gate).

Phase 1: EVIDENCE SWEEP (parallel, 2-4 agents)

Build the Evidence Ledger per reference/research-grounding.md §3 (schema: claim / source_url + retrieved / trust_tier T1–T4 / describes_version / category / verification). Adaptation: authority is primary-source-authoritative (highest-tier source wins ties; no capture oracle exists here).

parallel:
  - branch: external_grounding
    agents: [field]            # WebSearch/WebFetch sweep of T1 (official docs, papers, changelogs) → T3
    mission: exhaust T1 sources for every in-scope topic; cite + tier + date every entry
  - branch: internal_sources        # conditional: article derived from a codebase / releases
    agents: [lens | launch]
    mission: code citations, PR/release facts — internal claims get repo-path/SHA provenance, same ledger schema
  - branch: competitive_claims      # conditional: article compares products/tools
    agents: [compete]
    mission: comparison facts from each vendor's own T1 sources, never from a rival's marketing

In audit mode the sweep is claim-driven: a pre-pass extracts the existing article's claims (Phase 3 machinery run early, extraction only) and the sweep researches exactly those.

Gate (Research Completeness, inherited): T1 exhausted for in-scope topics; every ledger entry cited + dated; uncited entries rejected. Critical gaps (a load-bearing thesis has no findable source) → pause and present to the user before writing.

Phase 2: COMPOSE (compose mode only, 1-2 agents)

tome writes the article ledger-bound: every fact sentence carries an inline ledger reference ([E12]); anything not in the ledger is written as labeled opinion/hypothesis or not written. prose? for headings/hook. reference/doc-quality-protocol.md W1-W6 apply (reader contract + UNKNOWN-over-fabrication).

In audit mode this phase is skipped; the existing article is ingested verbatim.

Phase 3: CLAIM AUDIT (producer ≠ verifier, 1-2 agents)

A separately spawned auditor (never the writer):

1. Decomposes the article into atomic claims and classifies each per the taxonomy.
2. For every fact: locates its ledger entry, re-opens the cited source (WebFetch) and confirms citation-support — the source must state what the sentence states, at the claimed strength.
3. Flags mislabeled phrasing (speculation dressed as fact) and inference steps that outrun their sources.

Output: claim_audit.json — one row per claim: { claim_id, claim, class, ledger_ref, verdict, disposition, evidence_note, first_seen_cycle }. No claim may be absent from the table — unaccounted claims fail the gate (Acceptance-Provenance discipline, Q15).

claim_id is assigned at this first decomposition and never re-derived (§ Ledger Membership slot (c)). Remediation rewrites sentences, so a claim re-extracted from the fixed text would otherwise arrive as a new row with no history — and a claim that has already been refuted once would be indistinguishable from a fresh one.

Phase 4: ADVERSARIAL VERIFY (parallel, 2-4 agents)

Skeptic panel per _common/ADVERSARIAL_REFUTATION.md on the load-bearing thesis claims + any Phase 3 grounded claim the auditor marked shaky:

• Each skeptic independently tries to refute the claim via fresh search: contradicting T1-T3 sources, newer versions/deprecations, misread statistics (base rates, survivorship, denominator games), causal overreach.
• Corroboration rule: load-bearing facts require ≥2 independent sources (independent = not citing each other / not the same origin).
• Staleness check: every date-sensitive claim is tested against "still true at publication date?" — a 2023 benchmark presented as current is stale.
• Verdicts merge into claim_audit.json (a refuted claim's verdict is overwritten to contradicted/stale with the refuting citation attached).

Phase 5: REMEDIATE — loop ≤ N cycles (default N=2)

Fix per verdict, generator-evaluator separated per reference/evaluator-loop-protocol.md (writer fixes, auditor re-verifies the diff — never self-graded):

The writer states the claim_id it acted on for every remedy, so the re-audit re-attaches the claim's verdict history instead of re-discovering it (Phase 3). A claim returning to OPEN twice after a FIXED-VERIFIED is FROZEN + BLOCK, never a third attempt.

Re-run Phase 3 on changed text only, then the Thesis-Integrity gate (§ Gates). Exit reasons: ACCEPT (zero unresolved fact verdicts and the thesis gate green) | cap-reached (deliver best-so-far with residual claims explicitly marked [UNVERIFIED] + the residual gap listed) | BLOCK (a load-bearing thesis is refuted and the article's premise fails → escalate to the user; do not ship a corrected-into-meaninglessness piece).

Phase 6: DELIVER

NEXUS_COMPLETE with ## Nexus Execution Report plus the named Provenance Report:

### Provenance Report
- Mode: compose | audit          Risk tier: <tier>
- Evidence Ledger: N entries (T1: n / T2: n / T3: n / T4-leads: n)
- Claims: N total — fact n / inference n / opinion n / speculation n
- Verdicts: grounded n · miscited→fixed n · unsupported→(sourced n / downgraded n / deleted n) · contradicted→corrected n · stale→updated n
- Load-bearing thesis: each claim + its ≥2 independent sources + skeptic verdict
- Dispositions: every claim carries exactly one (§ Ledger Membership); `DOWNGRADED` / `DELETED (logged)` / `FROZEN` listed individually with the ratifying auditor
- Thesis-Integrity gate: result per cycle
- Residuals: [UNVERIFIED] claims shipped (0 required for external-facing) + deletions log
- Loop: N/2 cycles, exit reason

Gates

Resume: checkpoint-resume — recipes with ≥ 4 phases persist phase outputs at each boundary (ledger, claim_audit.json, remediation diffs), so an interrupted run resumes from the last checkpoint (newsroom resume).

AUTORUN Chain Template

recipe: newsroom
mode: AUTORUN_FULL
required_confirmation: only risk_tier == release-critical
phase_chain:
  - { phase: 0_framing,   owner: nexus_internal, gate: release_critical_confirm }
  - { phase: 1_evidence,  parallel: [field, lens|launch?, compete?], gate: research_completeness }
  - { phase: 2_compose,   agents: [tome, prose?], if: mode == compose }   # ledger-bound writing
  - { phase: 3_claim_audit, agents: [auditor(spawned, != writer)], output: claim_audit.json }
  - { phase: 4_adversarial, parallel: [skeptic ×2-3 per _common/ADVERSARIAL_REFUTATION.md, staleness_check] }
  - { phase: 5_remediate, loop: "≤ 2 cycles", gate: thesis_integrity, exit: [ACCEPT, cap-reached, BLOCK] }
  - { phase: 6_deliver,   output: NEXUS_COMPLETE + Provenance Report }

Failure Modes Prevented

Cost and Scale

Scale: 6-16 agents × ≤2 remediation cycles; 20-60 min; 2-5× a feature chain (compose mode upper range; audit of a short article sits at the bottom). Roughly half of podium (16-53 agents) — newsroom has no visual/layout/slide tracks.

Boundaries — vs Neighbors

• vs podium — podium is the package polish recipe (doc + slides + visuals, five teams); its claim-grounding is one verification branch. Newsroom inverts the weighting: grounding IS the deliverable's core guarantee, polish is incidental. An article that also needs a slide deck and hero imagery → podium (optionally with newsroom-grade rigor requested); an article whose selling point is "every claim checked" → newsroom.
• vs tome direct — Tome authors without Newsroom's Evidence Ledger, independent auditor, or refutation loop. Article with no grounding requirement → Tome direct.
• vs attest — attest verifies implementation against a normative spec (AC conformance). Newsroom verifies prose against the world (external sources).
• vs canon — canon checks compliance with named standards (WCAG/OWASP/style guides); newsroom checks factual truth provenance.
• vs clone/fuse/graft research sweep — same Evidence Ledger machinery (reference/research-grounding.md), different oracle: reproduction recipes stay capture-authoritative; newsroom is primary-source-authoritative because there is no artifact to capture.
• vs wish/marquee — ceiling-quality one-shot wrappers; newsroom is a repeatable grounding pipeline with a standard bar (zero ungrounded facts), not a scarcity-gated ceiling.

Decision tree

Is the deliverable prose that makes factual claims?
  └─ NO  → not newsroom (code → judge/acceptance; spec conformance → attest)
  └─ YES ↓
Does it need slides/visuals/multi-format packaging as the main ask?
  └─ YES → podium (grounding runs as its Phase 4 branch)
  └─ NO ↓
Is "every claim sourced / no speculation / verify correctness" part of the ask
  (or the artifact is external-facing with verifiable claims)?
  └─ NO  → tome direct
  └─ YES → newsroom  (existing draft → newsroom audit; new article → newsroom compose)

Shared protocols cited (not re-derived): _common/FINDING_LEDGER.md (the ledger machinery this recipe is a member of — scope freeze, identity across cycles, disposition vocabulary + integrity, oscillation, ZERO predicate, validity-gate requirement; newsroom declares its five slots in § Ledger Membership) · reference/research-grounding.md (Evidence Ledger schema, trust tiers, Research Completeness gate) · _common/ADVERSARIAL_REFUTATION.md (skeptic panel discipline) · reference/evaluator-loop-protocol.md (generator-evaluator separation, single termination oracle) · reference/autonomy-quality-protocol.md (Q9 producer≠verifier, Q10-Q11 evidence-bound claims, Q15 Acceptance Provenance) · reference/doc-quality-protocol.md (W4-W6 grounding / UNKNOWN-over-fabrication, W12 Doc Quality Gate — newsroom's Phase 3-4 gates subsume W12 and add citation-support + corroboration).