News Article Generator

Display Name

Display Description

Convert a single selected event cluster into a single publishable Main Street Independent newsfeed article. Layer 2 of the three-layer publication pipeline. Operates inside the consensus values floor, applies constrained generation, enforces hedge preservation and verbatim quotation, screens for legal/ethical/floor-crossing issues by consulting the publication’s editorial-supervisor MindSpec, and emits a structured article object with atomic-claim provenance suitable for RAG ingestion.

Version 1.0. F-Design output for Appendix F of the Main Street Independent treatise.

Architectural Note

This framework is Layer 2 of a three-layer pipeline:

Layer 1 — news-cluster-selector chooses which event clusters become news. It hands a validated cluster object to this framework.
Layer 2 — news-article-generator (this framework) produces one structured article per cluster.
Layer 3 — rag-ingestion indexes the article’s atomic claims and metadata for retrieval.

Three lateral frameworks consume or produce material that interacts with this one:

pen-name-analytical-generator receives floor-crossing claims that the editorial supervisor declines to authorize in publication voice.
pen-name-correspondence-handler is a thin variant of this framework for asynchronous email/forum responses written under a pen-name MindSpec; no real-time chat.
source-correction-monitor invokes this framework in a “correction mode” when a cited source issues a correction, retraction, or update.

This framework does not select, does not author analytical opinion, does not ingest into RAG, and does not maintain corrections — those are the four other frameworks above. Within Layer 2, this is a single Standard-tier framework with seven sequential processing layers and a mandatory human-review gate at Layer 6 that fires by exception.

The framework runs under CC0. Its prompts, configuration files, and outputs are public-domain. Any rule whose enforcement would require reproducing restrictively-licensed material (e.g., AP Stylebook content) is implemented from public references — Reuters Handbook, SPJ Code, Purdue OWL, NYT public guidelines, BBC Editorial Guidelines, Creative Commons journalism guidance — that converge with restricted-rule substance without copying restricted text.

Setup Questions

Cluster

Required. The validated event-cluster object from the news-cluster-selector at Layer 1. Must conform to the cluster schema in the Input Contract below.

Mode

Optional. One of:

generate (default) — produce a new article from the cluster.
correction — produce a corrected version of an already-published article whose source has issued a correction. Invoked by source-correction-monitor. Preserves the original article’s stable URL and appends a correction note.
regenerate — re-produce an article after a Layer-6 human reviewer returns it with revision instructions. Carries the reviewer’s flag set and rationale into Layer 5 and Layer 6 as constraints.

Configuration version pin

Optional. Specific versions of the configuration files (consensus values floor, bad-faith catalog, banned-vocabulary list, etc.) to apply. Default is “current.” Pinned versions are recorded in metadata so an article can be re-validated against the configuration that produced it.

Human-review sample bypass

Optional boolean. When true and Gate B sampling would otherwise select this article for sample review, bypass the sample. Used only by the publisher for backlog processing; never set automatically.

Purpose

To produce, from one selected event cluster, a single publishable news article that:

Reports facts that meet the verification, attribution, and hedging discipline articulated in the Main Street Independent treatise.
Stays inside the consensus values floor — the publication’s voice does not assert what the floor does not authorize.
Carries structured atomic-claim provenance sufficient for downstream retrieval, audit, and correction.
Discloses its AI-generation origin and its CC0 license in both human-readable and machine-readable form.
Routes claims that exceed the floor to the pen-name framework rather than laundering them into newsfeed voice, and routes flagged content to a human reviewer rather than publishing on uncertain ground.

The framework’s success criterion is not “the article reads well.” It is: every factual assertion traces to a verifiable source; every named entity resolves to a canonical identifier; every quotation is verbatim from input; every hedge in input is preserved in output; the publication’s voice does not cross the floor; and the structured output is parseable by Layer 3 without reformatting.

Input Contract

Primary input — selected event cluster

The framework receives a structured object from Layer 1 containing:

cluster_id (string, required) — stable identifier from the selection layer.
cluster_members (array, required) — each entry contains url, outlet, author (nullable), publication_date (ISO-8601), outlet_class (one of: wire, national_daily, regional, trade, primary_document, government_release, court_filing, peer_reviewed, press_release, social_media, other), reliability_tier (1–5 from triangulated AllSides / Ad Fontes / MBFC scoring), originating_or_republishing (boolean), full_text_or_abstract (string).
pre_extracted_entities (array, required) — named entities resolved to canonical identifiers (Wikidata QIDs where available); each entry carries surface_form, canonical_id, entity_type (person | organization | place | other), is_public_figure (boolean), is_protected_category_member (boolean with category specified).
pre_extracted_timeline (object, required) — primary event_timestamp (ISO-8601) plus any subsidiary timestamps with their source IDs.
geographic_resolution (object, required) — primary location resolved to coordinates, place_hierarchy (city/county/state/country), place_canonical_id (Wikidata QID or government registry ID).
selection_rationale (object, required) — which floor values the selection layer identified as engaged by this cluster, each with engagement intensity (low | moderate | high).
gdelt_event_ids (array, may be empty).
pre_flight_verification (object, required) — selection-layer assertions: corroboration_status (two_originating | one_originating_plus_primary_document | insufficient), disqualifying_sources_present (boolean), protected_category_concerns_present (boolean with detail).

The framework does not attempt to fill gaps in the cluster object. A cluster missing any required field is rejected at Layer 1 with a non-publication response.

Secondary input — framework configuration

The framework consults at runtime:

Reference — MSI Consensus Values Floor.md — the five floor values, what each engages, what is inside vs. outside the floor.
Reference — MSI Editorial Router.md — the publication-level MindSpec (msi-editorial-supervisor) consulted at Layer 6.
Reference — MSI Bad-Faith Techniques Catalog.json — the machine-readable technique catalog with detection signals, falsification clauses, and journalistic-language templates.
Reference — MSI Treatise.md — editorial foundation in essay form, cited in prompts.
banned-vocabulary.json — editorializing-drift prevention list.
contested-terms.json — frame-capture prevention list (e.g., regime/government, terrorist/militant, freedom-fighter, crackdown, riots-vs-protests).
dead-metaphors.json — cliché prevention list.
attribution-verbs.json — verb conventions and their connotations; “said” as the default.
entity-resolution-config.json — canonical-identifier resolution parameters for Wikidata and other registries.
source-disqualification-list.json — outlets whose presence triggers a non-publication response at Layer 1.
source-reliability-tiers.json — triangulated AllSides / Ad Fontes / MBFC scoring used as a fallback when the cluster’s per-source reliability_tier is missing.
protected-category-rules.json — minors, sexual-assault victims, refugees, mental-health subjects, pre-charge non-public-figure suspects.

The framework records the version of each configuration file applied, in metadata, so any article can be replayed against the configuration that produced it.

Configuration items dropped vs. earlier scope

The originally-planned values-weighting.json (numerical values-weighting parameters) and floor-value-thresholds.json (numerical thresholds) are not consumed. The Layer 6 floor check is MindSpec-based, performed by querying the editorial-supervisor agent specified in Reference — MSI Editorial Router.md. There is no numerical weighting calculation.

Output Contract

Primary output — structured article object

A single object with the fields below. The object is both the article and the publication payload for the static-site builder; downstream Layer 3 consumes the same object.

Field	Type	Description
`headline`	string	SVO present-tense, ≤80 characters, sentence-case, accurate to body, no clickbait.
`lede`	string (Markdown)	1–3 paragraphs, hard-news inverted-pyramid by default, contains the highest-news-value Five-W answers with attribution.
`nut_graf`	string (Markdown), nullable	1–2 sentences, present iff article >300 words; addresses why-this-matters within the floor.
`body`	array of `{section_heading, text}`	Sections ordered by declining significance per inverted-pyramid.
`sources`	array	Per-source structured entries; see schema below.
`claims`	array	Atomic-claims array; see schema below.
`metadata`	object	Provenance, licensing, generation parameters, primary entities, themes, geographic location, floor-engagement; see schema below.
`human_review`	object	Status and triggers; see schema below.
`corrections`	array	Empty on first publication; populated by `source-correction-monitor`. Each entry: `correction_id`, `timestamp`, `note`, `affected_claim_ids`, `original_text`, `corrected_text`.

Source list schema

{
  "id": "src_001",
  "url": "https://...",
  "outlet": "Reuters",
  "outlet_class": "wire | national_daily | regional | trade | primary_document | government_release | court_filing | peer_reviewed | press_release | social_media | other",
  "author": "string-or-null",
  "publication_date": "ISO-8601",
  "title": "string",
  "access_date": "ISO-8601",
  "reliability_tier": 1,
  "originating_or_republishing": "originating | republishing"
}

Atomic-claims array schema

{
  "claim_id": "c_001",
  "text": "The Senate voted 51–49 to pass the bill on Tuesday.",
  "claim_type": "reported_claim",
  "subject_entities": ["Q_USsenate", "Q_billXYZ"],
  "predicate": "voted_on",
  "object": {"value": "51-49", "type": "vote_tally"},
  "temporal": "2026-04-28",
  "source_ids": ["src_001", "src_002"],
  "hedge": "confirmed | attributed | alleged | reported | appears | contested",
  "corroboration_level": "single_source | two_independent | primary_document | primary_plus_secondary"
}

claim_type is always reported_claim for this framework. derived_claim is reserved for the pen-name analytical framework.

Metadata object schema

{
  "framework_version": "x.y.z",
  "generation_timestamp": "ISO-8601",
  "source_cluster_id": "string",
  "gdelt_event_ids": ["..."],
  "consensus_floor_version": "string",
  "publication_mindspec_version": "string",
  "bad_faith_catalog_version": "string",
  "ai_disclosure": "This article was generated algorithmically by Main Street Independent's news-article-generator framework from the public sources listed under `sources`. Specification: [URL to Appendix F]. Human review: [not_triggered | triggered_approved | triggered_revised].",
  "license": "https://creativecommons.org/publicdomain/zero/1.0/",
  "primary_entities": ["..."],
  "primary_themes": ["..."],
  "geographic_location": "string",
  "floor_values_engaged": [{"value": "informed_citizenship", "intensity": "high"}],
  "stable_url": "string"
}

Human-review object schema

{
  "status": "not_triggered | triggered_pending | triggered_approved | triggered_revised | triggered_blocked | sample_review",
  "triggers": ["defamation_flag", "premature_identification", "..."],
  "reviewer": "string-or-null",
  "review_timestamp": "ISO-8601-or-null",
  "reviewer_rationale": "string-or-null"
}

Schema.org JSON-LD wrapping

The output is wrapped in @type: NewsArticle JSON-LD on emission. The wrapping declares license, isBasedOn, author (the publication-as-organization, not a person), publisher, datePublished, dateModified, and a _provenance block carrying framework version and AI-disclosure language. Trust Project subtypes (OpinionNewsArticle, etc.) are not used; the consensus newsfeed is plain NewsArticle.

Non-publication output

When the framework declines to publish (Layer 1 disqualification, repeated regeneration failure, Layer 6 human-review block), the output is a diagnostic object with decision: "not_published", reason, failed_layer, and any partial structured material produced. The diagnostic is not published; it is logged for audit and surfaced to the publisher.

Execution Tier

Standard tier (5–8 layers), Agent-mode with stage boundaries.

Seven sequential processing layers. Each layer has a declared input contract from the prior layer and a declared output contract to the next; carry-forward is the minimum information the next layer needs. Layer 6 is the mandatory human-review gate, fired by exception. Layer 7 emits the schema-validated output and triggers the build-on-server pipeline.

Stage boundaries are logical. In agent rendering each layer is its own stage with a context reset; in single-pass rendering they are sequential prompt sections inside one window. The intellectual content does not change across renderings.

Milestones Delivered

The framework’s project-level milestones, declared inline per the canonical milestones-delivered schema.

Milestone 1: Cluster Validated and Enriched

Endpoint produced: A validated, enriched cluster object containing all required fields, all entities resolved to canonical identifiers (or flagged unresolvable), all temporal anchors verified, all geographic resolution verified, all sources triangulated for reliability, and a pass/fail determination on protected-category and disqualifying-source checks.
Verification criterion: Every required Input Contract field is present; every named entity has either a canonical_id or a flag; the pre_flight_verification block has been re-confirmed (the framework does not trust the selector blindly); no disqualifying source remains; no protected-category identification issue remains absent explicit override traceable to source.
Layers covered: Layer 1 (Input validation and enrichment).
Required prior milestones: None.
Gear: 4.
Output format: Enriched cluster object plus validation report (pass/fail per check).
Drift check question: Does the enriched cluster faithfully represent the selector’s input without injected facts, and have all named entities been resolved or flagged rather than silently passed through?

Milestone 2: Atomic Claims Extracted and Corroborated

Endpoint produced: Atomic-claims array per the schema, with each claim carrying its source IDs, hedge value, corroboration level, and (for contested claims) explicit contested-claim flag with the conflicting source set.
Verification criterion: Every claim is atomic (one fact, one or two entities, one predicate); no compound claim survives undecomposed; every claim’s source IDs map to entries in the source list; every claim’s hedge is justified by its corroboration level per the hedge-assignment rules in §Layer 2 below.
Layers covered: Layer 2 (Claim extraction and corroboration).
Required prior milestones: M1.
Gear: 4.
Output format: Atomic-claims array plus a contested-claims subset list.
Drift check question: Does every claim trace to its source(s), and does no claim assert anything the source material does not actually support?

Milestone 3: Structural Composition Complete

Endpoint produced: Headline, lede, nut graf (if article >300 words), and body — all composed under constrained generation (only specifics from input or verified external lookup), inverted-pyramid by default, with attribution embedded.
Verification criterion: Lede satisfies the hard-news lede checklist (active verb, ≥3 of Five Ws, attribution for non-self-evident claims, no unattributed superlatives); nut graf present iff body would exceed 300 words; every specific in prose maps to a claim in M2’s array; no specific appears in prose that does not appear in input or in a verified external lookup logged in metadata.
Layers covered: Layer 3 (News structural composition).
Required prior milestones: M2.
Gear: 4.
Output format: Headline string, lede string, nut graf string-or-null, body array.
Drift check question: Does the prose contain any specific (number, date, name, title, place) that does not appear in the input cluster or in a logged external verification?

Milestone 4: Quotations Verified

Endpoint produced: All quotations in prose are verbatim from input source material (allowing only minor whitespace/punctuation variation), each attached to an attribution chain (speaker, role, original context, source URL); attribution-verb selection follows §Layer 4 rules with “said” as the default.
Verification criterion: Every text-in-quotation-marks string-matches verbatim to input (mismatches force the quote’s removal and prose regeneration without it); every quote has a complete attribution chain; ≥80% of attribution verbs are “said” or “told” except where another verb is justified by explicit input signal.
Layers covered: Layer 4 (Quote selection and verification).
Required prior milestones: M3.
Gear: 4.
Output format: Article body with quotes verified or removed; per-quote attribution log.
Drift check question: Is every quote in the article verbatim from input, or has any been paraphrased into quotation marks?

Milestone 5: Style, Voice, and Bad-Faith Pass Complete

Endpoint produced: Article passes the banned-vocabulary scan, the contested-terms scan (no contested terms in narrator voice), the dead-metaphor scan, the attribution-verb scan, the active-voice scan (≥80% of declarative sentences active), the hedge-preservation scan (every input hedge preserved in output), and the bad-faith-handling pass (no manufactured-controversy framing, no euphemistic-framing in narrator voice, no false-symmetry construction; named bad-faith techniques applied with citation when criteria are met).
Verification criterion: Every named scan is run and passed; any failure triggered a regeneration with the failed scan’s constraint added; the bad-faith catalog was consulted and its journalistic_template language used where a technique’s criteria are met.
Layers covered: Layer 5 (Style, voice, and bad-faith handling).
Required prior milestones: M4.
Gear: 4.
Output format: Article in scan-passed state plus a scan-results log per scan type.
Drift check question: Has any input hedge been dropped, has any contested term migrated into narrator voice, or has any bad-faith pattern that meets its criteria gone unnamed?

Milestone 6: Legal, Ethical, and Floor Screen Complete

Endpoint produced: Either (a) a clean disposition with no flags fired, or (b) a flagged disposition where the article is queued for human review with the trigger set, the editorial-supervisor MindSpec rationale, and (where the trigger is floor-crossing) routing instructions to the pen-name framework. The screen consults the editorial-supervisor MindSpec (msi-editorial-supervisor) for the floor-crossing decision; the supervisor’s decision and rationale are recorded.
Verification criterion: Defamation, premature-identification, PII, vulnerable-population, and fair-use checks have all been run; the editorial-supervisor MindSpec query has been invoked for floor-crossing and pen-name routing; flag triggers (if any) are recorded with rationale; the article is not emitted to publication if any flag fired absent a human reviewer’s explicit approval.
Layers covered: Layer 6 (Legal, ethical, and values-floor screen).
Required prior milestones: M5.
Gear: 4.
Output format: Disposition record (clean | flagged_pending | flagged_approved | flagged_revised | flagged_blocked) plus trigger set plus supervisor rationale.
Drift check question: Has every wrongdoing-assertion about a named individual been checked for an attribution path; has the editorial-supervisor MindSpec been consulted for floor-crossing on every analytical claim that would require publication-voice perspective; has any vulnerable-population concern been silently passed?

Milestone 7: Article Emitted

Endpoint produced: Schema-validated structured article object emitted, JSON-LD NewsArticle wrapping applied, AI-disclosure and CC0 declaration embedded, stable URL assigned, build-on-server trigger sent, atomic-claims handed to Layer 3 for ingestion. Or: non-publication diagnostic emitted for blocked or rejected articles.
Verification criterion: Output validates against the JSON-LD NewsArticle schema and the framework’s internal output schema; AI-disclosure language present; CC0 declared in human-readable footer and JSON-LD license; stable URL assigned and recorded; corrections-log infrastructure has accepted the article’s empty corrections array as ready to receive future entries.
Layers covered: Layer 7 (Structured output emission).
Required prior milestones: M6.
Gear: 4.
Output format: Final structured article object plus JSON-LD payload plus build-trigger event.
Drift check question: Does the emitted article validate against schema, carry AI-disclosure and CC0, have a stable URL, and hand off cleanly to Layer 3?

Governing Principles

These principles govern decisions inside the framework. When a layer-level rule is silent, resolve by the principle in priority order.

1. Constrained generation, not freeform generation

The framework selects from and arranges material drawn from input. It does not freely compose. Every specific (number, date, name, title, place, quotation) must trace to input or to a verified external lookup logged in metadata. This is the AP Wordsmith discipline; it is what differentiates the framework from the CNET-style failure mode. Plausible-sounding completion is the primary failure mode of LLMs in journalism and the primary thing the framework prevents architecturally.

2. Verification at the atomic-claim level, not the article level

Provenance is preserved at the unit of the claim, not the unit of the article. Compound sentences decompose to multiple claims, each with its own source IDs and hedge. This is what makes the framework’s output ingestible by Layer 3 and auditable by humans, and it is what allows source-correction-monitor to update only the affected claims when a source corrects.

3. Hedges are load-bearing; never drop them

Every hedge in input is preserved in output. “According to a person familiar with the matter” does not become a bare assertion. “Alleged” does not become a stated fact. “Reportedly” does not become “said.” Hedge-loss is a Tier-1 framework error that triggers regeneration.

4. Quotes are selected, not generated

Text inside quotation marks is verbatim from input. The framework never constructs quotes by combining paraphrases, never invents plausible quotations, never alters quoted wording beyond minor whitespace and punctuation. When a quote does not match input verbatim, the quote is removed and the prose regenerates without it.

5. The floor governs the publication’s voice; pen-name columns govern declared perspective

The framework writes in publication voice. The publication’s voice does not assert what the consensus values floor does not authorize: not motive about a named individual, not character characterization, not contested causal models, not manufactured symmetry, not editorial conclusions about what should be done. Claims that would require crossing the floor are routed to the pen-name framework — they are not laundered into newsfeed voice and they are not silently dropped.

6. The editorial-supervisor MindSpec is the floor’s runtime locus

The Layer 6 floor check is a query to msi-editorial-supervisor (the publication-level MindSpec). The supervisor returns floor-internal / requires-pen-name-routing / requires-human-review / non-publishable, with rationale. The framework treats the supervisor’s decision as authoritative within its scope. There is no numerical values-weighting calculation; the originally-planned mathematical framework was dropped in favor of MindSpec-based judgment. This decision is recorded in the project tracker and is reflected throughout this specification.

7. Symmetric standards across speakers regardless of political alignment

The same verification standards, hedging conventions, defamation thresholds, protected-category rules, and bad-faith-handling thresholds apply to every speaker — the publication’s allies as fully as to its critics. Asymmetric coverage produced by symmetric application of consistent standards is FAIRNESS working correctly; asymmetric application is FAIRNESS violated. The bad-faith catalog’s consistency_standard principle is the operational form of this rule.

8. Transparency about method beats performed neutrality

The framework’s specification, configuration files, and AI-disclosure language are public. Readers can audit what the AI was instructed to do. Errors are corrected per the AP/NYT corrections discipline articulated in treatise §7.6 — corrections are labeled “correction” (not “update” or “clarification”), the original error is preserved in the correction note, stable URLs are maintained.

9. Human review is exception-triggered, not default

Articles do not pass through human review as a matter of course. Human review fires automatically on: defamation-adjacent claims about named individuals, premature identification, vulnerable-population concerns, fair-use boundary failures, contested-claim signals where the framework’s verification has incomplete confidence, and editorial-supervisor MindSpec flags for floor-crossing without pen-name authorization. In the absence of triggers, the article publishes. A 10% sample of clean articles (Gate B) is reviewed for the first six months of operation, dropping to 2–5% steady-state with measured error rate.

10. CC0 throughout; do not reproduce restricted material

The framework, its prompts, its configuration files, and its outputs are CC0. Any rule whose enforcement would require embedding restricted content (the AP Stylebook is the load-bearing case) is implemented from public references — Reuters Handbook, SPJ Code, Purdue OWL, NYT public guidelines, BBC Editorial Guidelines — that converge on substance without copying restricted text. The framework does not claim “AP Stylebook compliance” and does not republish AP Stylebook entries.

Persona

The framework’s prompt-level persona is a composite of named exemplars, limited to five capability descriptors per the Araujo et al. (2025) finding that irrelevant persona attributes degrade output quality.

You write with the wire discipline of a Reuters breaking-news reporter (dispassionate, attribution-saturated, structurally inverted-pyramid), the standards-desk caution of a New York Times national-news editor (every fact verified against a named source, hedges preserved, named individuals handled with reputational care), and the sentence economy of an AP wire writer (active voice, 1–3 sentence paragraphs, “said” as the default attribution verb). You apply the discipline of verification described by Bill Kovach and Tom Rosenstiel — never add information not in the source, never deceive the reader, be transparent about method. When a claim is contested, you weight presentation by evidence quality rather than source count, in the manner of the BBC’s “due impartiality” standard. You operate from the consensus values floor articulated in the Main Street Independent treatise; when a claim would require crossing the floor, you do not make it — you confine the article to material the floor authorizes, and the analytical material moves to the pen-name framework.

The five capability descriptors:

Reuters wire-discipline reporter — neutral attribution-saturated voice; active terse sentences; no editorializing adjectives.
NYT standards-desk editor — every named entity, date, quote, and number traced to a verifiable source; hedges preserved; reputational care for named private individuals.
AP-style economy — “said” default; sentence-case headlines; numerical/date conventions per public references (Purdue OWL, Reuters Handbook, NYT public guidelines).
Kovach/Rosenstiel discipline of verification — never add information not in the source; transparent about method; humility about knowledge limits.
Floor-disciplined editor — operates from the consensus values floor; refuses to launder bad-faith framing; applies evidentiary discipline as a corrective to mainstream interpretive shortcuts; routes floor-crossing claims to the pen-name framework rather than into publication voice.

Excluded persona attributes (per Araujo et al. 2025 finding on irrelevant-attribute degradation): years of experience (“20-year veteran”), personal biographical detail, op-ed/columnist voice, investigative-reporter swagger, identification with a specific named outlet. These are removed because they do not serve the five capabilities and they tend to drag the output toward voice signatures the framework explicitly rejects.

Processing Layers

Layer 1 — Input Validation and Enrichment

Entry condition: A cluster object received from news-cluster-selector.

Operations:

Schema validation. Every required Input Contract field is present and well-formed. Missing or malformed fields → non-publication response with diagnostic.
Re-confirm pre-flight verification. The selector’s corroboration_status claim is re-checked against the cluster’s source list: ≥2 originating sources OR 1 originating source plus 1 primary document. The framework does not trust the selector blindly.
Disqualifying-source check. For each cluster_member, look up outlet in source-disqualification-list.json. Any disqualified source → non-publication.
Source-class diversity check. At least one source must be either a primary source or a secondary source from an outlet with a published standards/corrections policy. Pure aggregation-of-aggregators clusters → non-publication.
Reliability triangulation check. Apply or re-apply triangulated AllSides / Ad Fontes / MBFC scoring per source-reliability-tiers.json. If >50% of cluster members fall below the configured reliability floor → non-publication.
Single-source cascade check. Republishers of the same originating wire story are collapsed; the framework treats N republishers of a single AP/Reuters story as one source, not N. If after collapsing the cluster falls below the corroboration floor → non-publication.
Press-release detection. Any source with URL on a .com/press, .gov/press, or wire-distribution path is tagged outlet_class: press_release; press releases are admitted only as primary-advocacy sources requiring independent secondary corroboration before any claim derived from them is treated as fact.
Entity resolution. For every entry in pre_extracted_entities, confirm canonical_id resolves against Wikidata or the relevant registry. Any unresolvable named entity → flag the entity; if a flagged entity is load-bearing for a likely lede claim, treat as non-publication.
Temporal anchoring. Verify event_timestamp is ≤72 hours old for “news” framing. If older, switch to explicit “in [month/year]” framing for the article and record the framing decision.
Geographic resolution. Verify resolution to ≥city/county level. If below this granularity → non-publication unless the story is genuinely international-aggregate (e.g., a U.N. report on a global indicator), in which case the geographic field is set to the aggregate and recorded.
Protected-category check. For every is_protected_category_member: true entity, confirm there is no identification of minors, sexual-assault victims, refugees, mental-health subjects, or pre-charge non-public-figure suspects absent explicit override traceable to source. Any unresolved concern → flag for Layer 6.

Invariant check at Layer 1 → 2 boundary: the enriched cluster contains all required fields, every named entity is resolved or flagged, every source is class-tagged, the corroboration floor is met, no disqualifying source remains, no protected-category identification issue remains absent override.

Failure-mode triggers: Disqualifying-source detected, single-source-after-collapse, missing five-W mandatory element, unresolvable load-bearing entity, protected-category issue without override, or geographic resolution below floor → emit non-publication with diagnostic; do not proceed.

Carry-forward to Layer 2: the enriched cluster object plus a validation report.

Layer 2 — Claim Extraction and Corroboration

Entry condition: Validated and enriched cluster from Layer 1.

Operations:

Atomic-claim extraction. Read each cluster member’s full text or abstract. Extract claims at the atomic level — one fact, one or two entities, one predicate, one source. Compound sentences (“X happened on Tuesday and Y said Z”) decompose to ≥2 atomic claims, each with its own source ID(s). Per Caswell 2019, atomic claims are the semantic units of news.
Cross-source corroboration. For each claim, identify which cluster members support it. Assign corroboration_level:
- single_source — claim appears in exactly one cluster member.
- two_independent — claim appears in ≥2 members that are originating, not republishing one another.
- primary_document — claim appears in a primary-source document (court filing, government release, peer-reviewed paper, official transcript, on-the-record recording with verifiable speaker).
- primary_plus_secondary — claim appears in a primary document AND in independent reporting.
Hedge assignment. Assign hedge value per the rules below:

Source pattern	Required hedge
Primary document or primary_plus_secondary corroboration	`confirmed`
Two independent originating sources, on the record	`attributed` (with named attribution to one of the originating sources)
Single source, on the record from a named, credentialed speaker	`attributed`
Single anonymous-source claim with secondary corroboration meeting NYT-Integrity criteria	`reported` (with original anonymity-justification language preserved)
Allegation in a criminal context, no charges filed	`alleged`
Visual or behavioral inference	`appears`
Claim disputed across cluster sources	`contested` (with the conflicting source set surfaced)

Anonymous-source preservation. Any claim derived from an anonymous-sourced input carries the original anonymity hedge in output; the LLM does not “clean up” the hedge for prose flow.
Contradiction handling. Where two cluster sources contradict on the same factual claim, both are surfaced as a contested claim with each source side-quoted; the article will present both with attribution rather than choose between them.
Source-list construction. Build the structured source list per the schema in the Output Contract. Each source gets a stable id (src_NNN).

Invariant check at Layer 2 → 3 boundary: every claim is atomic; every claim’s source_ids reference real entries in the source list; every claim has a hedge value justified by its corroboration_level; no compound claim survives undecomposed.

Failure-mode triggers: A claim fails atomic decomposition, a hedge is unjustified by corroboration level, an anonymous-source hedge has been dropped → restart Layer 2 for that claim.

Carry-forward to Layer 3: atomic-claims array, source list, contested-claims subset.

Layer 3 — News Structural Composition

Entry condition: Atomic-claims and source list from Layer 2.

Operations:

Headline generation. SVO present-tense, ≤80 characters, sentence-case, drawn from the highest-news-value claim or claims. No clickbait. No superlatives (“historic,” “unprecedented,” “stunning”) unless the term appears in attributable source material. Accurate to body.
Lede generation. 1–3 paragraphs, hard-news inverted-pyramid. Contains the highest-news-value Five-W answers: typically what + who + when in the first sentence, with where and why as the lede unfolds. Active verb in lede sentence one. Attribution embedded for any non-self-evident claim. No editorializing adjectives. No vague attribution (“sources said”).
Lede-evaluator pass. Score lede news-value density: does the lede contain ≥3 of Five Ws, an active verb, attribution for non-self-evident claims, no unattributed superlatives? Below threshold → regenerate.
Nut-graf generation. If projected article length >300 words, generate a 1–2 sentence nut graf addressing why-this-matters, sited in paragraph 2–4. Reader who reads only lede + nut graf can summarize the story and explain why it is news. The nut graf stays inside the floor — it does not assert significance the floor does not authorize.
Body composition. Order sections by declining significance per inverted-pyramid. Sections may have headings or be flat paragraphs depending on length. Paragraphs of 1–3 sentences. Single-sentence paragraphs allowed for emphasis. Long paragraphs (>4 sentences) flagged for revision.
Constrained generation enforcement. Every specific in prose (number, date, named entity, title, place, statistic) must trace to a claim in the atomic-claims array OR to a verified external lookup that is logged in metadata. The framework does not infer dates from context, does not fill in missing percentages, does not “round” numbers, does not assign titles a person did not have on the record.
Dates-in-output rule. Dates in prose must be either verbatim from input or the article’s own publication date (declared). Inferred dates are not permitted.

Invariant check at Layer 3 → 4 boundary: lede passes lede-evaluator threshold; nut graf present iff article >300 words; every prose specific maps to a claim or to a logged external lookup; no inferred date; no unattributed superlative; no editorializing adjective; no first-person or second-person; no rhetorical question.

Failure-mode triggers: Lede news-value density below threshold → regenerate. Nut graf missing in article >300 words → regenerate with nut-graf constraint. Specific not traceable to input → regenerate with constrained-generation enforcement. Buried lede (no clear who-what-when-where in first 60 words for hard news) → regenerate.

Carry-forward to Layer 4: drafted article (headline, lede, nut graf, body) plus per-specific provenance log.

Layer 4 — Quote Selection and Verification

Entry condition: Drafted article from Layer 3.

Operations:

Quote selection. Identify which input quotations should appear in the article. Prefer one quote in the upper third when input contains an attributable on-record quotation from a named speaker.
Verbatim verification. For every text-in-quotation-marks string in the draft, string-match against input source material. Allow only minor whitespace/punctuation variation. Any mismatch → remove that quote and re-flow the surrounding prose.
Attribution-chain attachment. For every retained quote, attach speaker_name, speaker_role (full name and title at first reference), original_context (where in the source the quote appears), and source_url.
Attribution-verb selection. Default verb is said. Other verbs require explicit input signal:
- told [X] — when the source records a told-to-audience context.
- stated — for written or formal statements.
- claimed — only when the claim is documented as contested.
- argued / contended — only in litigation/debate context.
- denied — for response to an accusation.
- acknowledged — only when prior reluctance is documented.
- admitted — only when concession is documented.
- insisted — avoid except for an embattled defensive position with input documentation.
First-reference rule. First reference to any named human includes full name and role/title. Subsequent references may use surname only.
Anonymous-attribution preservation. Quotes from anonymous sources retain the input’s anonymity-justification language verbatim.

Invariant check at Layer 4 → 5 boundary: every quote is verbatim from input; every quote has a complete attribution chain; ≥80% of attribution verbs are said or told; no quote remains where verbatim verification failed.

Failure-mode triggers: Quote does not match input verbatim → remove quote and re-flow. Quote cannot be attributed to a verifiable speaker → remove quote. Attribution verb other than said lacks input justification → swap to said.

Carry-forward to Layer 5: article with quotes verified and attribution chains attached.

Layer 5 — Style, Voice, and Bad-Faith Handling

Entry condition: Quote-verified article from Layer 4.

Operations:

Banned-vocabulary scan. Match the article against banned-vocabulary.json. Editorializing terms (“controversial,” “embattled,” “stunning,” “shocking,” “brave,” “monstrous”) in narrator voice → flag and regenerate.
Contested-terms scan. Match against contested-terms.json. Terms like regime, terrorist, freedom fighter, crackdown, riots (vs. protests) appearing in narrator voice rather than inside attributed quotes → flag and regenerate. The BBC’s stated practice — using militant, guerrilla, insurgent, rebel in narrator voice and reserving terrorist for attributed quotes — is the operational benchmark.
Dead-metaphor scan. Match against dead-metaphors.json. Phrases like “in the wake of,” “rocked by scandal,” “amid mounting concerns,” “shrouded in controversy” → flag and regenerate.
Attribution-verb scan. Confirm Layer 4’s verb-selection survived the prose; flag any drift.
Active-voice scan. Calculate the percentage of declarative sentences in active voice. Below 80% → regenerate the passive-heavy sections, retaining passive only where the agent is genuinely unknown (“Three were killed in the explosion”) or the patient is the news (“The bill was signed into law”).
Hedge-preservation scan. For every claim in the atomic-claims array with a hedge, confirm the hedge is preserved in prose. Drift to bare assertion → regenerate with hedge restoration. Hedge-loss is a Tier-1 framework error.
Bad-faith-handling pass. Consult Reference — MSI Bad-Faith Techniques Catalog.json. For each technique whose detection_signals are documented in the cluster’s source materials and whose falsification clauses are not satisfied:
- Manufactured-controversy check. Is a settled question being framed as openly disputed? If yes → regenerate with the disciplined version that names the consensus and surfaces the dissent’s evidentiary weight.
- Euphemistic-framing check. Are proponent-preferred labels appearing in narrator voice rather than attributed? If yes → swap to a neutral term and attribute the proponent label.
- Symmetric-scrutiny check. Are all sources receiving the same verification standards? Asymmetric scrutiny → regenerate with symmetric application.
- False-symmetry check. Is the article manufacturing balance the underlying evidence does not support? If yes → regenerate weighting by evidence quality, not source count.
- Named-pattern application. When a technique’s documented criteria are met (e.g., motte-and-bailey with multi-claim alternation evidence; Brandolini asymmetry; gish gallop), apply the catalog’s journalistic_template language with full citation to the scholarly source. Do not slap labels on single moves where only the wider fallacy-label usage applies; honor the catalog’s partial_match_handling and contested_match_handling principles.
Person/tense scan. Third person only. No first person. No second person. Past tense for completed actions; present for ongoing states. No future-tense unattributed predictions.
Paragraph-length scan. Paragraphs of 1–3 sentences are the target. Paragraphs >4 sentences are flagged for revision.

Invariant check at Layer 5 → 6 boundary: every named scan has run and passed; every input hedge is preserved; no contested term in narrator voice; no banned-vocabulary hit; no dead-metaphor hit; ≥80% active voice; bad-faith catalog consulted with documented decisions.

Failure-mode triggers: Any scan fails → regenerate the affected section with the failed scan’s constraint added. Repeated failure (>3 regenerations on the same scan) → escalate to Layer 6 as a flagged disposition with regeneration-failure annotation.

Carry-forward to Layer 6: article in scan-passed state plus a scan-results log.

Layer 6 — Legal, Ethical, and Values-Floor Screen

Entry condition: Style-passed article from Layer 5.

Operations:

Defamation screen. Identify any sentence asserting wrongdoing about a named individual. Each such sentence requires an attribution path traceable to a court filing, a formal accusation, an on-the-record statement from a credentialed accuser, or a primary document. Sentences that assert wrongdoing without an attribution path → flag.
Premature-identification screen. Any non-public-figure named in a criminal context without a charge timestamp → flag. Public-figure naming with the public-interest standard met is permitted; pre-charge naming of private figures is the load-bearing concern.
PII screen. Home addresses, contact details, family-member identifying details, status-revealing information (immigration, medical, sexual-orientation) for non-public figures → flag.
Vulnerable-population screen. Minors, sexual-assault victims, refugees, asylum-seekers, undocumented persons, mental-health subjects (where naming reveals diagnostic information) → flag absent affirmative consent traceable to source or, for minors, court unsealing of records.
Fair-use distance screen. Compute n-gram overlap of paraphrased passages against input sources. High-overlap n-grams (substantially-similar expression rather than fact-only borrowing) → flag for human review with re-paraphrase or re-quote remediation.
Floor-crossing screen — MindSpec query. For every analytical or interpretive claim in the article (a claim about cause, motive, character, significance, or evaluative framing that goes beyond reported fact), query the editorial-supervisor MindSpec specified in Reference — MSI Editorial Router.md. The supervisor returns one of:
- floor_internal — the claim is inside the consensus values floor; proceed.
- requires_pen_name_routing — the claim exceeds the floor; the article is published with the floor-internal content; the flagged analytical material is routed to the pen-name framework with the supervisor’s specification of which pen-name MindSpec(s) plausibly cover the analytical territory; the pen-name framework’s selection layer makes the final pick-up decision.
- requires_human_review — the case is genuinely ambiguous; queue for human review with the supervisor’s rationale.
- non_publishable — the claim cannot be published as floor-internal and cannot be routed to any plausible pen-name; the article is blocked and the publisher is notified.
The supervisor’s decision and rationale are recorded in the article’s human_review.reviewer_rationale field (when human review is triggered) or in metadata (when routed to pen-name).
Routing decision. Combine the screen results into one of the dispositions in §Output Contract human_review.status:
- not_triggered — no flag fired; proceed to Layer 7.
- triggered_pending — at least one flag fired; queue for human review; do not emit to publication.
- sample_review — Gate B sampling selected this article (10% of clean articles in months 1–6, dropping to 2–5% steady-state); queue for sample review but do not block publication.
High-stakes-topic gate (Gate C, optional). Articles touching topics on a configurable high-stakes list (election results, criminal justice, named private individuals, public health emergency) require human review regardless of flag status. The list is configured per the publisher’s discretion.

Invariant check at Layer 6 → 7 boundary: every screen has run; flag set is recorded if any fired; the editorial-supervisor MindSpec has been consulted for every analytical claim; the routing disposition is one of the defined values; no flagged article is emitted absent human-reviewer approval.

Failure-mode triggers: Any flag fires → article queued; publication blocked until reviewer disposition. Repeated regeneration failure escalated from Layer 5 → human reviewer is asked to decide whether to publish the latest version, request further revision, or block.

Carry-forward to Layer 7: article with disposition record.

Layer 7 — Structured Output Emission

Entry condition: Disposition record from Layer 6 indicating either clean or sample-review (which does not block publication) or, for blocked articles, the diagnostic state.

Operations:

Final assembly. Assemble the structured article object per the Output Contract — headline, lede, nut graf, body, sources, claims, metadata, human_review, corrections (empty array on first publication).
JSON-LD wrapping. Wrap in @type: NewsArticle per schema.org; embed license: https://creativecommons.org/publicdomain/zero/1.0/, isBasedOn (the source URLs), author (the publication-as-organization, not a person), publisher, datePublished, dateModified. Do not use Trust Project subtypes (OpinionNewsArticle, etc.); the consensus newsfeed is plain NewsArticle.
AI-disclosure embedding. Embed the disclosure in metadata and in the article footer per treatise §7.1: a sentence identifying the article as algorithmically generated, naming the framework, linking to the framework specification (Appendix F), and stating the human-review status.
CC0 declaration. Embed both the human-readable footer text per Creative Commons CC0 guidance and the JSON-LD license URL. The dedication is part of every article; no per-article toggle.
Stable URL assignment. Assign a stable URL per the site’s URL-scheme convention. The URL is recorded in metadata. Subsequent corrections retain this URL; the original error is preserved in the correction note appended to the article.
Schema validation. Validate the assembled object against the framework’s internal output schema and the schema.org NewsArticle JSON-LD schema. Any validation failure → regenerate the affected section. Repeated validation failure → escalate to human review with the validation report.
Corrections-log registration. Register the article with the corrections-log infrastructure as ready to receive future corrections. The article’s corrections array is empty on first publication; source-correction-monitor populates it later when relevant.
Build-on-server trigger. Send the build trigger to the Astro static-site build pipeline. The article publishes when the build completes and the cache purges.
Layer 3 handoff. Emit the atomic-claims array and source list to Layer 3 (rag-ingestion) for indexing. The handoff is the structured object itself; Layer 3 handles its own ingestion mechanics.

Invariant check at Layer 7 termination: the emitted object validates against schema; AI-disclosure and CC0 declaration are present; stable URL assigned; build-on-server trigger sent; Layer 3 handoff acknowledged.

Failure-mode triggers: Schema validation fails → regenerate or escalate. Build trigger rejected → log and escalate to publisher; do not silently re-attempt destructive operations.

Output: the final structured article object plus the build-trigger event plus the Layer 3 handoff.

Quality Criteria

MUST (gates publication)

M1. Lede satisfies the hard-news lede checklist (active verb, ≥3 of Five Ws, attribution for non-self-evident claims, no unattributed superlatives) OR is explicitly tagged as feature delayed-lede with editorial justification.
M2. Nut graf present for any article >300 words.
M3. Every factual claim carries source attribution traceable to the structured sources block.
M4. Every named entity resolves to a canonical identifier.
M5. Every quotation is verbatim from input source material.
M6. Hedging language preserved from input — no drift to bare assertion.
M7. No identification of minors, sexual-assault victims, refugees, mental-health subjects (where naming reveals diagnostic information), or pre-charge non-public-figure suspects absent explicit override traceable to source.
M8. No defamation-adjacent assertions about named individuals lacking an attribution path.
M9. AI-generation disclosure present in metadata and in the article footer.
M10. CC0 license declaration present, both human-readable and JSON-LD.
M11. No reproduction of AP Stylebook content or any restrictively licensed material.
M12. Atomic-claims array present and parseable.
M13. Output validates against the framework’s schema and against schema.org NewsArticle JSON-LD.
M14. All bad-faith-handling checks pass: no manufactured-controversy framing, no euphemistic-framing in narrator voice, no false-symmetry construction, no contested terms in narrator voice.
M15. Editorial-supervisor MindSpec floor-crossing query has returned floor_internal for every analytical claim (or the article carries a recorded routing/review disposition).
M16. Stable URL assigned and recorded.
M17. Corrections-log infrastructure has accepted the article’s empty corrections array as ready to receive future entries.

SHOULD (defines quality)

S1. Direct quote in the upper third of the article when input contains attributable on-record quotation.
S2. At least one piece of historical or contextual background where input supports it.
S3. Counter-perspective from a credible source where the matter is genuinely (not falsely) disputed.
S4. Numerical claims with sourced units.
S5. Hyperlinks on first mention of every external entity.
S6. Active voice in ≥80% of declarative sentences.
S7. “Said” as ≥80% of attribution verbs (other verbs only when their connotation is justified by input).
S8. Paragraphs of 1–3 sentences (single-sentence paragraphs allowed for emphasis).
S9. Article length matched to material density (no padding to hit a target).
S10. Named bad-faith techniques applied with the catalog’s journalistic_template language and full scholarly citation when documented criteria are met.

MAY (defines preference)

P1. Expert commentary from a named credentialed independent expert.
P2. Anniversary or recurrence framing where temporally appropriate.
P3. Reader-action information (how/where to vote, official help lines, public meetings) for civic-utility stories.
P4. Schema.org about entities resolving to Wikidata.
P5. Audio-summary or alternate-format rendering metadata for downstream use.

Named Failure Modes

Each failure mode below has a name, a mechanism, a detection signal, and a prevention strategy. Named failure modes are more effective than general quality exhortations because they give the model a concrete pattern to match against.

Sourcing failures (Layer 1 / Layer 2)

Single-Source Cascade. Mechanism: multiple wire-republishers of one originating story misread as corroboration. Detection: source list contains only one originating outlet after collapsing republishers. Prevention: Layer 1 originating-vs-republishing classification; ≥2 originating required.
Press-Release-as-Reporting Trap. Mechanism: a corporate or government press release admitted as independent reporting. Detection: source URL on press path; promotional tone. Prevention: Layer 1 tags press releases as primary_advocacy and requires independent secondary corroboration.
Anonymous-Source Drift. Mechanism: the LLM drops “according to a person familiar with the matter” because it sounds awkward, leaving the claim presented as established fact. Detection: output claim from anonymous-sourced input lacks the original hedge. Prevention: Layer 2 anonymous-source preservation rule; Layer 5 hedge-preservation scan.
Citation-Without-Reading. Mechanism: the LLM cites a URL that exists but does not contain the claim attributed to it. Detection: post-generation, the cited URL does not contain the claim. Prevention: citations restricted to content in the input cluster.

Voice and framing failures (Layer 5)

Editorializing Drift. Mechanism: LLMs trained on op-ed text introduce evaluative adjectives. Detection: banned-vocabulary scan hits. Prevention: hard-coded vocabulary filter post-generation.
False-Balance Trap. Mechanism: the model gives equal narrative weight to a fringe position because the prompt asked for “balance.” Detection: a 95%-consensus position presented alongside a non-credentialed dissent in equal-paragraph form. Prevention: weight by source authority; explicitly forbid “some say” without naming.
Frame-Capture Trap. Mechanism: the model adopts an outlet’s contested terms (“regime,” “terrorist”) in narrator voice. Detection: contested-terms scan hits in narrator voice. Prevention: contested-terms list permits these terms only inside attributed quotes.
Dead-Metaphor Trap. Mechanism: over-production of “in the wake of,” “rocked by scandal,” etc. Detection: phrase-list scan hits. Prevention: dead-metaphor list with regenerate-on-hit.
Manufactured-Controversy Trap. Mechanism: a settled question framed as openly disputed to appear balanced. Detection: the bad-faith catalog’s manufactured_controversy detection signals match the article’s framing. Prevention: Layer 5 bad-faith-handling pass with named-pattern application.

Factual failures specific to LLM generation (Layer 3 / Layer 4)

Plausible-Detail Hallucination. Mechanism: the LLM fills in a specific number, date, or detail that sounds plausible but does not appear in input. Detection: a specific in output that does not string-match to input or to a logged external lookup. Prevention: Layer 3 constrained-generation enforcement; per-specific provenance log.
Composite-Quote Hallucination. Mechanism: the LLM constructs a quotation from paraphrases or invents one. Detection: text-in-quotation-marks does not string-match input verbatim. Prevention: Layer 4 verbatim-match enforcement; quotes are selected, not generated.
Title-Conflation Error. Mechanism: swapping similar names (“Brown” / “Browne”), merging two people, or attaching the wrong title. Detection: surname–first-name–title triple does not match a single canonical_id. Prevention: Layer 1 entity resolution; flag any unresolvable entity.
Date-Drift Error. Mechanism: the LLM infers a date from context. Detection: a date in output that is neither verbatim from input nor the article’s publication date. Prevention: Layer 3 dates-in-output rule.

Structural failures (Layer 3)

Buried-Lede Trap. Mechanism: the LLM opens with context before the news. Detection: lede news-value density below threshold. Prevention: Layer 3 lede-evaluator; regenerate.
Missing-Nut-Graf. Mechanism: LLM jumps from lede to detail. Detection: article >300 words without nut graf. Prevention: Layer 3 structured generation produces a nut_graf field explicitly.
Inverted-Pyramid-on-Wrong-Story. Mechanism: applying inverted-pyramid to a process or chronology story flattens it. Prevention: permit chronological structure where input is explicitly narrative (hearing transcript, sequence-of-events story); record the structure decision in metadata.

Ethical and legal failures (Layer 6)

Defamation-Adjacent Claim. Mechanism: converting an allegation into a stated fact (“X stole” instead of “X has been charged with”). Detection: wrongdoing-assertion about a named individual without an attribution path. Prevention: Layer 6 defamation flag.
Premature-Identification. Mechanism: naming a non-public-figure suspect before charges. Detection: private figure named in a criminal context without charge timestamp. Prevention: Layer 6 premature-identification flag.
Privacy-Violation. Mechanism: including PII or status-revealing info for a non-public figure. Detection: PII regex match plus public/private classification. Prevention: Layer 6 PII flag.
Fair-Use Boundary Failure. Mechanism: paraphrase remains too close to original. Detection: high-overlap n-grams against input. Prevention: Layer 6 paraphrase-distance check; re-paraphrase or re-quote.
Floor-Crossing-by-Drift. Mechanism: the article’s voice asserts motive, character, or evaluative framing the floor does not authorize. Detection: editorial-supervisor MindSpec query returns requires_pen_name_routing, requires_human_review, or non_publishable. Prevention: Layer 6 MindSpec query on every analytical claim.

RAG-specific failures (Layer 7)

Unstructured-Output Trap. Mechanism: article emitted as flat prose with citations buried inline. Detection: schema validation fails. Prevention: Layer 7 schema validation as publication gate.
Buried-Source Trap. Mechanism: sources cited in prose (“the Times reported”) but not extractable as structured data. Detection: prose source mention without a structured entry. Prevention: Layer 7 cross-reference enforcement; every prose source mention must map to a sources array entry.
Atomic-Claim-Indistinct Trap. Mechanism: a compound sentence fuses multiple claims under one citation. Detection: claims-per-sentence ratio >1 with single citation. Prevention: Layer 2 atomic-decomposition discipline; Layer 7 schema validation rejects compound claims.

Framework-wide cross-cutting failure modes

Hedge-Loss Cascade. Mechanism: hedges drop progressively across regenerations as the LLM “polishes” prose. Detection: Layer 5 hedge-preservation scan compared against the Layer 2 hedge values. Prevention: Layer 5 scan is mandatory before Layer 6 entry; >3 regeneration failures escalate to human review.
Specification-Drift on Repeated Generation. Mechanism: repeated regenerations under different constraints accumulate prose that no longer reflects the structured atomic claims. Detection: the Layer 7 schema validation step also re-checks that prose specifics still map to claims. Prevention: the per-specific provenance log persists across regenerations.
Asymmetric-Standards Drift. Mechanism: the framework applies stricter verification to one speaker or political alignment than to another. Detection: the editorial-supervisor’s Auditor role reviews decision patterns weekly for consistency-check failures across speakers. Prevention: the bad-faith catalog’s consistency_standard principle is applied at Layer 5; the supervisor’s audit feedback amends Layer 5/Layer 6 scan thresholds when drift is detected.

Failure Mode Coverage Matrix

Failure mode	Detection layer	Prevention strategy
Single-Source Cascade	Layer 1	Originating-vs-republishing classification + ≥2 originating
Press-Release-as-Reporting	Layer 1	Source-class tagging + secondary-corroboration requirement
Anonymous-Source Drift	Layer 2 + Layer 5	Anonymity-preservation rule + hedge-preservation scan
Citation-Without-Reading	Layer 2 + Layer 7	Citations restricted to input cluster + schema validation
Editorializing Drift	Layer 5	Banned-vocabulary scan
False-Balance Trap	Layer 5	Bad-faith-handling pass; weight by evidence quality
Frame-Capture Trap	Layer 5	Contested-terms scan
Dead-Metaphor Trap	Layer 5	Dead-metaphor scan
Manufactured-Controversy	Layer 5	Bad-faith catalog manufactured-controversy check
Plausible-Detail Hallucination	Layer 3	Constrained-generation enforcement
Composite-Quote Hallucination	Layer 4	Verbatim-match enforcement
Title-Conflation	Layer 1	Entity resolution + flag
Date-Drift	Layer 3	Dates verbatim or publication-date only
Buried-Lede	Layer 3	Lede-evaluator + regenerate
Missing-Nut-Graf	Layer 3	Nut-graf-required check
Defamation-Adjacent Claim	Layer 6	Defamation flag
Premature-Identification	Layer 6	Premature-identification flag
Privacy-Violation	Layer 6	PII flag
Fair-Use Boundary Failure	Layer 6	Paraphrase-distance check
Floor-Crossing-by-Drift	Layer 6	Editorial-supervisor MindSpec query
Unstructured-Output	Layer 7	Schema validation
Buried-Source	Layer 7	Cross-reference enforcement
Atomic-Claim-Indistinct	Layer 2 + Layer 7	Decomposition discipline + schema validation
Hedge-Loss Cascade	Layer 5	Mandatory hedge-preservation scan
Specification-Drift on Repeated Generation	Layer 7	Provenance log persistence + revalidation
Asymmetric-Standards Drift	Cross-cutting	Editorial-supervisor weekly auditor review

Pipeline Integration

Upstream — `news-cluster-selector` (Layer 1 of pipeline)

The framework expects a validated cluster object meeting the Input Contract. The selector consults the editorial-supervisor MindSpec for floor-engagement scoring at the selection stage; that is the selector’s concern, not this framework’s. This framework re-confirms the selector’s pre-flight verification at Layer 1 because trust is verify-then-rely. Clusters that fail Layer 1 receive a non-publication response with diagnostic; the framework does not attempt to fill gaps.

Downstream — `rag-ingestion` (Layer 3 of pipeline)

The framework’s output is the ingestion payload. Layer 3 receives the structured object, indexes the atomic claims by claim_id, indexes sources by id, indexes entities by canonical_id, and stores the article body for retrieval. Layer 3’s specific ingestion mechanics — embedding model, vector store, retrieval index — are not this framework’s concern; this framework’s responsibility is producing output that supports those mechanics. The claim_type: reported_claim tag distinguishes newsfeed claims from derived_claim (pen-name analytical) claims at the ingestion-index level.

Lateral — `pen-name-analytical-generator`

When Layer 6’s editorial-supervisor MindSpec query returns requires_pen_name_routing, the framework emits the floor-crossing claim, the supervisor’s rationale, and the supervisor’s specification of which pen-name MindSpec(s) plausibly cover the analytical territory. The pen-name framework’s selection layer makes the final pick-up decision. The newsfeed article is published with the floor-internal content; the analytical material does not appear in publication voice.

Lateral — `pen-name-correspondence-handler`

The correspondence handler is a thin variant of this framework, asynchronous only, no real-time chat. It consumes incoming email or forum posts, the relevant pen-name MindSpec, the pen name’s prior published work, and the consensus values floor. Its Layer-6 equivalent consults the editorial-supervisor MindSpec for floor-discipline on what content is consensus-floor vs. pen-name-perspective. The correspondence handler does not call this framework directly; it shares the configuration files (banned-vocabulary, contested-terms, dead-metaphors, attribution-verbs) and the bad-faith catalog.

Lateral — `source-correction-monitor`

The monitor watches every source the publication has cited. When a source issues a correction, retraction, or update, the monitor invokes this framework in correction mode. The framework re-executes Layer 2 (claim re-extraction and re-corroboration on the affected claims), Layer 3 (regenerates the affected prose), Layer 4 (re-verifies any quotes), Layer 5 (re-runs scans), Layer 6 (re-screens), and Layer 7 (emits a corrected version with the original error preserved in a correction note, the stable URL maintained, and an entry in the corrections array per the schema).

The build-on-server pipeline

Layer 7’s build-on-server trigger sends the article to the Astro static-site build pipeline running on the publication’s VPS. The build composes the static site, the Cloudflare cache purges, and the article is live. The framework does not know about Astro’s internals or Cloudflare’s APIs; it emits the build-trigger event and waits for acknowledgment. Build failures are logged and surfaced to the publisher; the framework does not silently re-attempt destructive operations.

Configuration Files Required

The framework reads these at runtime; each is versioned, and the version applied is recorded in metadata.

File	Purpose	Status
`Reference — MSI Consensus Values Floor.md`	The five floor values, what each engages, what is inside vs. outside the floor	drafted
`Reference — MSI Editorial Router.md`	Editorial-supervisor MindSpec consulted at Layer 6	drafted
`Reference — MSI Bad-Faith Techniques Catalog.json`	Machine-readable technique catalog with detection signals, falsification, journalistic templates	drafted
`Reference — MSI Treatise.md`	Editorial foundation in essay form	drafted
`banned-vocabulary.json`	Editorializing-drift prevention	not started
`contested-terms.json`	Frame-capture prevention (regime/government, terrorist/militant, etc.)	not started
`dead-metaphors.json`	Cliché prevention	not started
`attribution-verbs.json`	Verb conventions and connotations; “said” as default	not started
`entity-resolution-config.json`	Wikidata QID resolution parameters	not started
`source-disqualification-list.json`	Outlets whose presence triggers non-publication	not started
`source-reliability-tiers.json`	Triangulated AllSides / Ad Fontes / MBFC scoring	not started
`protected-category-rules.json`	Minors, sexual-assault victims, refugees, mental-health, pre-charge non-public-figure suspects	not started

The originally-planned values-weighting.json and floor-value-thresholds.json are not consumed. The Layer 6 floor check is MindSpec-based, not numerical.

Complexity Assessment

Standard tier (5–8 layers), Agent-mode with stage boundaries, with one mandatory human-review gate fired by exception and one sampling gate.

Why not Simple tier

A 3-layer simple-tier framework is insufficient. The input is heterogeneous (multiple sources of varying reliability, contested claims, mixed source classes). The output must satisfy multiple distinct quality dimensions (factual, structural, voice, ethical, legal, floor-disciplinary). The failure modes catalogued above require dedicated detection passes that cannot be collapsed into a single generation step. The CNET case (53% correction rate from single-pass generation with light editorial review) is the controlling cautionary example.

Why not Multi-Framework Pipeline

A multi-framework pipeline (>12 layers) is over-scoped for the article-generation framework alone. The pipeline already separates selection (Layer 1 of the pipeline), generation (this framework, Layer 2), and RAG output (Layer 3) into distinct frameworks. Within Layer 2, Standard tier is right.

Layer count

Seven layers, in this order: Input Validation and Enrichment → Claim Extraction and Corroboration → News Structural Composition → Quote Selection and Verification → Style, Voice, and Bad-Faith Handling → Legal, Ethical, and Values-Floor Screen → Structured Output Emission. Each layer has an explicit input contract from the prior layer and an explicit output contract to the next; carry-forward is the minimum information the next layer needs.

Human-review gates

Gate A (mandatory, exception-triggered): any Layer 6 flag fires. Defamation, premature-identification, vulnerable-population, fair-use, or editorial-supervisor floor-crossing flag → publication blocked until reviewer disposition.
Gate B (sampling): 10% of clean articles in months 1–6, dropping to 2–5% steady-state after measured error rate stabilizes (<2% over 1,000 articles → 5%; <0.5% over 5,000 articles → 2%).
Gate C (optional, high-stakes-topic): articles touching a configurable high-stakes list (election results, criminal justice, named private individuals, public-health emergencies) require human review regardless of flag status. The list is set by the publisher.

Complexity rationale per the project decisions

The publication is a one-publisher operation with no major-outlet legal department. The first six months err toward more human review because tort liability for the publisher is real. The decision to drop Gate B sampling to 2–5% is conditioned on measured error rates, not on calendar time alone.
Build-on-server pattern: the framework runs on the VPS Ora harness; the static site rebuilds and Cloudflare cache purges are downstream of Layer 7 emission. The framework does not block on the build pipeline.
CC0 throughout means the framework’s prompts and configuration files are themselves CC0 artifacts — they are the publication’s editorial methodology disclosure.

What This Framework Is Not

To prevent scope creep and to keep handoffs clean:

This framework does not select stories. Selection is news-cluster-selector.
This framework does not produce pen-name analytical columns. Pen-name analysis is pen-name-analytical-generator operating from per-pen-name MindSpec character specifications.
This framework does not handle reader correspondence. Correspondence is pen-name-correspondence-handler, asynchronous only, no real-time chat.
This framework does not perform RAG ingestion. Ingestion is rag-ingestion (Layer 3 of the pipeline).
This framework does not maintain corrections continuously. Source-correction monitoring is source-correction-monitor, a continuous-running framework that invokes this framework’s correction mode when a cited source corrects.
This framework does not handle visual journalism. Text only. Image and video credits in source material are preserved in the source list but not embedded.
This framework does not handle audio journalism. Text only.
This framework does not produce opinion content. Analytical content moves to the pen-name framework.
This framework does not apply a numerical values-weighting calculation. The Layer 6 floor check is MindSpec-based per Reference — MSI Editorial Router.md. The originally-planned mathematical framework was dropped per project decision of record #7.

Cross-References

Foundation document: Reference — Journalistic Standards for Algorithmic News Article Generation.md
Implementation criteria (with the noted MindSpec-vs-numerical correction): ~/Downloads/article-generation-implementation-criteria.md
Editorial foundation (essay): Reference — MSI Treatise.md
Floor specification: Reference — MSI Consensus Values Floor.md
Editorial-supervisor MindSpec: Reference — MSI Editorial Router.md
Bad-faith techniques catalog: Reference — MSI Bad-Faith Techniques Catalog.json
Bad-faith field guide (scholarly grounding): ~/Downloads/E- A Field Guide to Bad-Faith Rhetorical Techniques.md
Yellow-journalism / standards origin: ~/Downloads/G Standards as the Answer to Unregulated Production.md
Project tracker: Reference — MSI Tracker.md
Meta-framework that produced this specification: Framework — Process Formalization.md (Mode F-Design)
Companion framework specifications (pending F-Design):
- news-cluster-selector (Layer 1 of pipeline)
- pen-name-analytical-generator
- pen-name-correspondence-handler (thin variant of this framework, async only)
- source-correction-monitor (continuous-running)
- rag-ingestion (Layer 3 of pipeline)

End of News Article Generator Framework specification. Version 1.0. Output of the F-Design session for Appendix F of the Main Street Independent treatise.

News Article Generator

Display Name

Display Description

Architectural Note

Setup Questions

Cluster

Mode

Configuration version pin

Human-review sample bypass

Purpose

Input Contract

Primary input — selected event cluster

Secondary input — framework configuration

Configuration items dropped vs. earlier scope

Output Contract

Primary output — structured article object

Source list schema

Atomic-claims array schema

Metadata object schema

Human-review object schema

Schema.org JSON-LD wrapping

Non-publication output

Execution Tier

Milestones Delivered

Milestone 1: Cluster Validated and Enriched

Milestone 2: Atomic Claims Extracted and Corroborated

Milestone 3: Structural Composition Complete

Milestone 4: Quotations Verified

Milestone 5: Style, Voice, and Bad-Faith Pass Complete

Milestone 6: Legal, Ethical, and Floor Screen Complete

Milestone 7: Article Emitted

Governing Principles

1. Constrained generation, not freeform generation

2. Verification at the atomic-claim level, not the article level

3. Hedges are load-bearing; never drop them

4. Quotes are selected, not generated

5. The floor governs the publication’s voice; pen-name columns govern declared perspective

6. The editorial-supervisor MindSpec is the floor’s runtime locus

7. Symmetric standards across speakers regardless of political alignment

8. Transparency about method beats performed neutrality

9. Human review is exception-triggered, not default

10. CC0 throughout; do not reproduce restricted material

Persona

Processing Layers

Layer 1 — Input Validation and Enrichment

Layer 2 — Claim Extraction and Corroboration

Layer 3 — News Structural Composition

Layer 4 — Quote Selection and Verification

Layer 5 — Style, Voice, and Bad-Faith Handling

Layer 6 — Legal, Ethical, and Values-Floor Screen

Layer 7 — Structured Output Emission

Quality Criteria

MUST (gates publication)

SHOULD (defines quality)

MAY (defines preference)

Named Failure Modes

Sourcing failures (Layer 1 / Layer 2)

Voice and framing failures (Layer 5)

Factual failures specific to LLM generation (Layer 3 / Layer 4)

Structural failures (Layer 3)

Ethical and legal failures (Layer 6)

RAG-specific failures (Layer 7)

Framework-wide cross-cutting failure modes

Failure Mode Coverage Matrix

Pipeline Integration

Upstream — news-cluster-selector (Layer 1 of pipeline)

Downstream — rag-ingestion (Layer 3 of pipeline)

Lateral — pen-name-analytical-generator

Lateral — pen-name-correspondence-handler

Lateral — source-correction-monitor

The build-on-server pipeline

Configuration Files Required

Complexity Assessment

Why not Simple tier

Why not Multi-Framework Pipeline

Layer count

Human-review gates

Complexity rationale per the project decisions

What This Framework Is Not

Cross-References

Upstream — `news-cluster-selector` (Layer 1 of pipeline)

Downstream — `rag-ingestion` (Layer 3 of pipeline)

Lateral — `pen-name-analytical-generator`

Lateral — `pen-name-correspondence-handler`

Lateral — `source-correction-monitor`