Content Review & Escalation Framework

A single model for how much human review a piece needs before publish, what the enforcement action is, and who reviews it, driven by risk of harm if something is wrong, not by how hard the content was to produce. The goal: concentrate human editorial attention where an error could hurt a reader, expose legal or reputational risk, or erode domain authority across the network, and let low-risk content move fast on lighter checks.

Scope: all CSA content, with the Trend Hunter verticals (Mind/Body, Everyday Living, Experiences), the Bridge / Deep-Explainer / FAQ-Service-Journalism types (Google Discover Explainer, What to Know Next, FAQ), and the multi-property syndication footprint (Apple News, MSN, SmartNews, Us Weekly, Woman’s World, McClatchy metros) called out explicitly. A general-newsroom baseline (mapping to the same tiers) covers content outside those verticals.

The model has two dimensions

Every piece carries both a Review Tier and a Disposition. They answer different questions:

• Review Tier: how much human review the piece requires.
• Disposition: the enforcement action when a rule fires (block, gate, or flag).

Keeping them separate is deliberate: a piece can be high-review-tier yet only produce non-blocking flags, or low-tier yet hit an absolute block (a restricted name in a travel listicle).

Dimension A: Review Tier (one ladder, reconciled vocabulary)

The prior CSA “pipeline tier” names and the editorial “oversight” names are the same three levels of review intensity. This framework uses Tier 1 / 2 / 3 as canonical (Tier 1 = most review) and records the equivalence once so engineering and claims validation stay aligned:

Boundary note: Fully Automated (no claims-validation pass) is reserved for structured output with no editorial judgment: scores, weather, automated inbound feeds. Editorial Tier-3 prose (travel, lifestyle, non-safety how-tos) is Semi-Automated and still takes a light selective pass; it is not exempt from fact-checking.

Dimension B: Disposition (the enforcement action)

“Hard block” vs “hard stop”: a hard block is publish-prohibited (restricted list). A hard stop is a Gate that cannot clear until a specific condition is met (e.g., a flagged health claim removed or confirmed). Different mechanisms; do not conflate.

Review Tier assignment: content category → tier

Assign a tier at the Decision stage (the pre-production step where a trend/node is scoped and greenlit), before production, so the reviewer knows the bar going in. The tier is a floor, not a ceiling: when a piece feels riskier than its category, review it higher.

Tier 1: Mandatory full review

E-E-A-T-critical (Experience, Expertise, Authoritativeness, Trust), credentialed-source, or harm-bearing content. Read the source, not just the draft: if a claim, number, or safety statement can’t be traced to a credible cited source, it comes out before publish.

Reader-facing disclaimer on the health and legal categories

Verification protects the reader from a wrong claim. A disclaimer protects them from a right one applied to the wrong person, which is the residual risk on every category in this table that gives health or legal information. A Tier-1 piece in a health or legal category carries a reader-facing disclaimer, and the disclaimer is a gate, not a courtesy.

Three rules govern it:

• It is specific to the claim, not boilerplate. A generic line at the foot of every page is furniture and gets read as furniture. Name the thing that needs the caution.
• It sits where the claim sits, not only at the end. A dosage figure in paragraph three is where the caution belongs.
• It is a distinct field, not body copy. The CMS carries a dedicated medical and legal disclaimer field; that field is its home, so the requirement is checkable and the text cannot be lost in an edit. Where a piece publishes to a surface without that field, the disclaimer goes in the body and the piece does not clear Tier 1 without it.

This formalises the “keep a see-a-professional note where relevant” line already carried on the Mind/Body row. Where relevant was doing too much work: it left the most dangerous categories in this table to a reviewer’s judgement about whether a caution was warranted.

Tier 2: Spot-check

Accuracy matters, stakes are lower; verify the load-bearing facts and move on. Wrong numbers, stale prices, or a misattributed quote still cost credibility.

Tier 3: Minimal / automated

Low-stakes, low-controversy, easily corrected, factually stable. Verify the few things that would embarrass us if wrong (a venue, a date) and let them move.

Escalation flags: raise the review level

Any flag below escalates a piece one review level toward Tier 1 (Tier 3 → Tier 2 → Tier 1). Tier 1 is the ceiling: a piece already at Tier 1 that hits a flag stays Tier 1 but takes on the flag’s disposition and route. When more than one applies, escalate to the strictest (the deterministic multi-match rule is in Consolidated escalation rules).

Universal (any content)

• Names a living private person in an otherwise-lower-tier piece, or makes a claim of wrongdoing about anyone. (A victim/warning story built around a named private individual is already Tier 1; this flag catches incidental naming elsewhere.)
• Gives instructions someone could be physically harmed by following.
• States a statistic, dollar figure, or date as fact.
• Touches a protected group or a sensitive social issue.
• Is likely to be widely shared or picked up by other outlets.

Trend Hunter–specific

• Multi-property syndication. A parent trend fanning to 3+ sites means one error replicates network-wide; the parent node earns more scrutiny than a single article.
• Claim drift across variants. Any Deep-Explainer or Bridge variant (Google Discover Explainer, What to Know Next, FAQ) that adds a dosing number, safety claim, or verdict the parent didn’t: re-check claims per variant, not just per trend.
• Strong-verb health framing. “Cures,” “fixes,” “actually works,” “off switch,” “beats [X]” on a health topic triggers a verb-level claim check.
• Brand-new, untested collections. A Collection (a named cluster of related trends) with no published track record is reviewed one level stricter until the category is established.

Reviewer action rules (once a piece is in review)

Tier assignment routes the piece; these rules govern what the reviewer does with what the fact-checking module returns. Full detail: the canonical claims-validation standard.

• Verdict → action. FALSE → rewrite/remove before peer review. MISLEADING → reframe plus document. OVERGENERALIZED → narrow scope / qualify. INSUFFICIENT_EVIDENCE → add a verified source or cut. TRUE → informational, no action, except the carve-out below.
• Verdict severity is category-conditional. On health and real-estate/local-services content, OVERGENERALIZED escalates to the rewrite-or-remove (Needs Correction) track, not the milder Needs Clarification track.
• TRUE does not clear the gate for: health dosage / interaction / efficacy claims, authoritative financial figures (both require a primary-source document even on TRUE), and time-sensitive travel/scheduling (operating hours, visa/entry requirements: verify against the official source at publish even on TRUE).
• Source authority + count. Authoritative sources (government/institutional primary, peer-reviewed, established national outlets, McClatchy publications) are accepted; trade/industry publications, verified press releases, and court/police/legal filings are acceptable with verification; aggregator/unverified sources on a FALSE/MISLEADING verdict are treated as Needs Correction. Two independent authoritative sources for contested/harmful claims and for health claims without a primary source; a primary-source document (not news coverage of it) for legal outcomes, regulatory figures, health dosage/interaction/efficacy, and authoritative financial figures.
• Financial forward-looking claims (projections, forecasts) must be attributed to the originating institution, not stated as fact; one authoritative source suffices if attributed.
• Annually-updated regulatory figures (tax rates, benefit amounts, sentencing guidelines) require a vintage check: confirm the figure is current-cycle, not stale.
• Content-type hard-stops. Any flagged health claim is a hard stop until removed or confirmed. Health/legal FALSE requires senior-editor sign-off before CMS.
• Confidence. Any low-confidence verdict is treated as INSUFFICIENT_EVIDENCE and independently verified; 3+ low-confidence verdicts in one piece escalate to senior editor. (Pending module capability: confidence-level output is not yet confirmed live.)

Escalation routing: who reviews what

The disposition and tier decide whether a human is required; routing decides who. Rows marked [needs decision] are unsettled and are the piece to lock with Product before build.

Platform reality: where this gate has no mechanism yet

Everything above is written as a pre-publish gate: a piece is reviewed, then it ships. On Trend Hunter B2C that sequence does not exist today, and the framework should say so rather than describe a workflow the platform cannot run.

Confirmed at the 2026-08-03 CMS walkthrough:

• The CMS has no approval state. A piece is a draft or it is published. There is no held-for-review state a reviewer can act on and no state a gate can block at.
• Peer review is post-publication. The writer publishes, pastes the live URL into a Slack channel, and a reviewer replies. The reader can reach the piece before the reviewer does.
• The only publish gate agreed is a required headline. Nothing else prevents publication.

What this changes, and what it does not. It does not weaken any rule above: a Tier-1 piece still needs its verification and its disclaimer, and a writer who skips them has skipped them whether or not software stopped them. What it changes is the honest description of the control. On this platform the tier assignment is guidance a writer applies to their own work, and peer review is a detection step that finds an error after a reader could have seen it, not a prevention step.

Consequences worth carrying:

• A review-completeness measure cannot be built from state order. Published-then-reviewed is the normal path here, so a piece sitting in Published with no review is indistinguishable from one correctly reviewed, unless the review is recorded somewhere other than a Slack thread.
• The disclaimer requirement is the part most exposed by the gap, because it is the one thing above that a reader is harmed by the absence of and that no reviewer can add retroactively without the piece already having been live.
• The ask, if this is to become a gate: a held state the writer cannot self-clear, and a recorded reviewer against the piece. Both are CMS-side. Until they exist, do not describe Tier 1 as enforced on Trend Hunter B2C.

The tracker replacing the Slack loop is the nearest available record; it carries a peer-reviewer field and a Review state, which makes the review recordable even while it stays post-publication.

Traceability & metadata: what underpins the gate

The framework only works if the routing is captured as data on the piece.

• Tag at the Decision stage. Each node carries its Review Tier + Disposition before it enters production. Child/syndicated variants inherit the parent’s tier plus a mandatory claim-drift check.
• Firing signals (backing fields; illustrative where noted). tag_sensitive (suicide/violence/sexual-content classifier) for protected-group/sensitive-topic risk; tags_iab (Interactive Advertising Bureau content-taxonomy tier) plus a category-keyword match for content-category triggers; restricted-name match for the hard block; IS_CLUSTER_PARENT + child count for the syndication fan-out flag; variant-vs-parent claim diff (a computed check, to build) for claim drift; REUTERS_PHOTO / REUTERS_LINK for the rights spot-check (Reuters-only; a general licensed-asset signal does not yet exist). Named-individual detection requires named-entity recognition (NER); no existing field classifies it.
• Audit trail (per piece). Full module output at generation, every reviewer action taken (changed / removed / overridden plus one-line rationale), and the final publish state relative to the flags, readable by editors, senior editors, and content leads. (Stated requirement: per-piece CMS storage and reviewer role-access are pending dev/product confirmation.)

Extending the rule set

The rule set is designed to grow. A valid rule row carries all five fields: trigger, review tier, disposition (block / gate / flag), reviewer route, and firing signal. Adding a rule:

• New trigger/row (a new risk category or firing signal) is a change-controlled addition: it is versioned in the rule set’s changelog and approved by the content lead plus the standard’s owner before it goes live.
• Re-tiering an existing category (moving it between Tier 1/2/3 as performance data lands) is a lighter change, raised through the standard content-governance review.
• A new IAB tier or tag_sensitive value maps into an existing rule by extending that rule’s firing-signal set, not by adding a row.
• Retiring a rule follows the same change-controlled path as adding one: a changelog entry plus owner approval.

The systemic scaled-content gate (output-multiplier limits, no-templated-substitution, named-author, HITL-review, publication-cadence checks) is a separate decision-stage gate operating at the format/persona/cadence level, distinct from this piece-level review model; a piece can pass here and still fail there.

How to use this

Tag each node with its Review Tier + Disposition at the Decision stage (inheritance and firing signals: §Traceability; re-tiering and rule growth: §Extending the rule set). Build the check-and-route engine against §Consolidated escalation rules.

Open before build: the tier/disposition decision logic is settled. Two kinds of work remain. Routing decisions (marked [needs decision]): who the “qualified reviewer” is for Tier 1 health, and the Legal-involvement threshold for named-individual content. Firing-signal / data-layer build items (flagged across Traceability, the rule table, and Reviewer action rules): named-entity recognition for named individuals (rows 6, 27), the variant-vs-parent claim diff (row 14), a general licensed-asset signal beyond Reuters (row 20), harmful-instruction detection (row 26), an external-pickup/reach predictor (row 25), a protected-group classifier (row 24), audit-trail CMS storage, and confidence-level handling (module-gated). The restricted-individuals hard block is gated on Legal’s list handover.

Consolidated escalation rules

One row per routing rule, keyed for a check-and-route engine: trigger → review tier → disposition → route → firing signal. (Verification-modifier rules such as the TRUE carve-outs, category-conditional severity, source-count, forecast attribution, and stale-figure vintage live in Reviewer action rules, not as rows.) Escalation rules use ↑1 (ceiling: Tier 1) to mean “raise one review level toward Tier 1, clamped at Tier 1.” Firing-signal combinators: AND = all signals required, OR = any signal.

When multiple rows fire: final tier = strictest (Tier 1 wins); final disposition = most-blocking (block > gate > flag); final route = the union of all matched routes.

Named-individual centrality (rows 6, 27): an NER hit alone can’t distinguish a story built around a named private individual (Tier 1) from incidental naming (↑1); default to Tier 1 pending a human centrality judgment that may downgrade to row 27.

Human-judgment branches: the wrongdoing determination (rows 6, 27) is a human call made at the Gate, not an automated signal.

Evaluation happens at two points: category-assignment rows fire at the Decision stage, before production; rows whose firing signal is content-derived (module verdict, flag count, claim diff, and the lexical/numeric/classifier detectors: rows 11, 12, 14, 15, 23–26) fire on a post-generation pass. ↑1 is computed against the category’s base tier, then folded via strictest-wins.