Insights Agent → CSA Push Endpoint—Pre-Activation Audit

Status: Pre-activation audit. Endpoint wired locally by Insights Agent vendor (per 2026-05-12 demo); production activation timing unknown. Audit date: 2026-05-14 Governing standards: AI Content Vetting Policy v1.9.11 · Scaled Content Policy §6/§7


What this endpoint does

Per the 2026-05-12 demo: the Insights Agent has a push to CSA / add to package button on the trend deep-dive surface. When operated, it sends the trend’s accumulated context (deep-dive research, content match, decision scoring, social proof + cluster context) to the CSA backend as a new package—bypassing the manual content-team intake path.

This makes it a new ingestion vector into CSA. Each new ingestion vector is a separate enforcement surface for AI vetting + scaled-content + restricted-individuals + substance-floor governance.


Why it triggers the AI Content Vetting Policy

The AI Content Vetting Policy v1.9.11 (csa-content-standards/docs/ai-content-vetting.md) was authored to govern upstream content sources that may produce AI-generated or low-substance material. The Insights Agent push endpoint is exactly this:

  1. Trend deep-dive content is AI-generated (Insights Agent + MAIA do the deep research + draft the content match).
  2. The trend “decision scoring” cited in the demo is black-box (heat index, audience match, trend strength, content gap—all undocumented).
  3. Some of the upstream feeds into the Insights Agent come from the B2B Trend Hunter platform (~1000 LLM-gobbledygook articles/week, human-out-of-loop). Any B2B-derived signal entering CSA via this route inherits that contamination.

The endpoint is therefore in scope for the 5-gate test. Below: the test applied to a hypothetical push from this endpoint.


5-gate test applied

Gate 1: Named credentialed author

Policy: Every CSA-bound content unit must have a named, credentialed human author of record. AI-generated drafts don’t get bylines.

Endpoint status:FAIL by default. The Insights Agent demo showed done with MAIA for both deep-dive and content match. No human author is named on the trend object. The “deep dive done by [name]” field shown in the demo refers to the human who invoked the research, not the author of the resulting content.

Required mitigation pre-activation: The push endpoint payload must either (a) inject a named credentialed editorial author at push time, or (b) tag the package as requires_editorial_byline_assignment so the CSA HITL gate forces byline assignment before any variant generation.

Gate 2: Substance floor

Policy: Content must clear a substance floor—real reporting, real expertise, new value. Pure synthesis/restatement of public signals does not clear.

Endpoint status:CONDITIONAL FAIL. Trend deep-dives in the Insights Agent demo were aggregations of social signals (X, Bluesky, cluster proof) + AI-synthesized angle text. These are summaries of what’s already trending—not original reporting or net-new analysis.

Required mitigation pre-activation: Each push must carry a substance_origin flag indicating whether the content contains net-new editorial value (interview, original analysis, named-expert reaction) vs. pure aggregation. Aggregation-only packages route to a separate review path with a stricter cadence cap (see Gate 5).

Gate 3: Plagiarism + AI-content detection

Policy: All inbound content runs through plagiarism (Copyscape per p24) + AI-content classification at intake.

Endpoint status:FAIL by default. The demo flow showed direct push from Insights Agent to CSA’s “add to package”—no plagiarism or AI-detection step in the visible UI.

Required mitigation pre-activation: The push endpoint must trigger the same Copyscape + AI-classifier hooks the partner-feed ingest uses (per the AI vetting policy’s intake_validation block). If those hooks aren’t yet wired for the Insights Agent surface, the endpoint must not activate.

Gate 4: Editorial fit

Policy: Content must pass an editorial-fit check against the destination outlet’s house style (csa-content-standards §10.3 us-weekly · §10.5 womans-world · §10.6 ap-compatible · §10.4 trend-hunter-b2c).

Endpoint status:CONDITIONAL FAIL. The Insights Agent demo did not show outlet-selection at push time. The trend object has cross-outlet relevance but no enforced destination.

Required mitigation pre-activation: Push endpoint payload must include explicit destination outlet(s); CSA backend then applies that outlet’s house-style enforcement at variant-generation time. Default destination = TH B2C if unset; do not allow undefined destinations.

Gate 5: Cadence audit

Policy: Total cadence across all CSA ingestion vectors must remain within editorial-team capacity. Each vector has a cadence cap; pushes beyond cap queue rather than ingest.

Endpoint status:FAIL by default. No cadence cap configured for this vector. An operator who clicks “push to CSA” 100 times in an hour would flood the queue.

Required mitigation pre-activation: Establish per-operator and per-day cadence cap. Recommended starting cap: 10 pushes/operator/day, queue-and-throttle beyond. Revisit after 2 weeks of activation data.


Scaled Content Policy §6/§7 check

§6 (McClatchy / CSA Application)

Check 1—output multiplier vs human baseline. A single trend push could generate N variants × M outlets. If a single editorial-pushed trend yields >10× human-baseline output (i.e., one operator-click → 10+ published articles across the portfolio), trigger §6 review.

Endpoint status:DEPENDS ON VARIANT-FANOUT CONFIG. Push endpoint by itself doesn’t multiply; CSA’s variant generation does. But the COMBINATION (Insights Agent push × CSA’s automatic format-+-persona variant generation) materially raises the multiplier ceiling.

Required mitigation: Push packages from this endpoint should be flagged requires_editorial_fanout_approval—the writer/editor consents to fanout volume per-piece, rather than the system defaulting to full multiplication.

Check 2—named credentialed author. Covered in Gate 1 above. ❌ FAIL.

Check 3—only city/name changed. A trend pushed to multiple outlets must vary by more than templated city substitution. The bridge-variant similarity probe (2026-05-14, head_blend 0.370 vs O&O 0.376 baseline) shows bridges already differentiate adequately; this endpoint must inherit that same gate.

Check 5—HITL review. Covered in Gates 1 + 4. Currently FAIL.

Check 7—publication cadence inconsistent with team size. Covered in Gate 5. Currently FAIL (no cap set).

§7 (9-check audit list)

Applying §7 audit to a hypothetical first activation:

# Check Status Notes
1 >10× human-baseline multiplier ⚠ depends Variant fanout config gates this
2 No named credentialed author ❌ FAIL Gate 1
3 Only city/name changed ⚠ depends Inherit bridge-similarity gate
4 No editorial intent ❌ FAIL Trend-pushed = aggregation, not editorial intent
5 No HITL review ❌ FAIL Push bypasses manual intake
6 No defensible substance ⚠ depends Gate 2
7 Cadence inconsistent with team size ❌ FAIL Gate 5
8 No fact-checking ⚠ depends Gary path closed; Copyscape covers plagiarism, fact-checking is separate
9 Output appears templated ⚠ depends Inherit bridge-similarity gate

Halt triggers from §7: Checks 2, 5, 7 are unconditional halt triggers per the Scaled Content Policy. Three unconditional halts active at endpoint default state. Endpoint must not activate against production CSA until all three are resolved.


Pre-activation requirements (summary)

The Insights Agent → CSA push endpoint MUST satisfy these before any production activation:

  1. Byline gate. Push payload injects a named credentialed editorial author OR tags requires_editorial_byline_assignment for HITL block. Halts §7 check 2.
  2. Substance flag. Push payload includes substance_origin (aggregation vs net-new editorial). Aggregation-only routes to stricter cadence cap.
  3. Plagiarism + AI-content hooks wired. Copyscape + AI-classifier hooks the partner-feed ingest uses must trigger at this surface too. Halts §7 check 8 (partial—AI detection covers plagiarism + slop; fact-checking remains separate, see #7).
  4. Outlet destination required. Push payload includes explicit destination outlet(s); CSA backend applies outlet’s house-style at variant-generation time. No undefined destinations.
  5. Cadence cap. Per-operator and per-day caps configured; queue-and-throttle beyond. Recommended starting: 10 pushes/operator/day. Halts §7 check 7.
  6. Variant-fanout approval. Push packages flagged requires_editorial_fanout_approval; operator consents to fanout volume per-piece. Halts §6 check 1.
  7. Editorial HITL. Push surface enforces an editorial review step before content reaches add to package for variant generation. Halts §7 check 5 (the unconditional one).
  8. Restricted-individuals scrub. Push payload runs against the restricted-individuals list (~7 legal/risk no-fly names; see csa-content-standards/docs/restricted-individuals.md, scoping in flight as of 2026-05-14). Any match → hard-block.
  9. B2B-source quarantine. If any signal in the push payload traces to the B2B Trend Hunter platform (the LLM-gobbledygook source), the package routes to a higher-scrutiny review path. Per the data-source provenance writeup owed by upstream vendor (rq-insights-agent-source-provenance), provenance flags must be queryable.

What I’m not auditing here

These are governance-adjacent but don’t gate the AI vetting policy enforcement question. They’re tracked separately in csa-dashboard.


Recommendation

Activate-blocked posture. The endpoint should not flip on against production CSA until items 1, 5, 7 from the pre-activation requirements list are resolved (the three unconditional §7 halt triggers). Items 2-4 and 6, 8, 9 can ship in a follow-up pass but should be prioritized before any meaningful operator-volume.

The Insights Agent vendor (per the demo) is the natural surface to take this to—vendor-side build closes the endpoint-side gates; CSA-side build closes the destination-side gates.

Escalation path: AI vetting policy v1.9.11 is in csa-content-standards repo; current endorsement status is “drafted, awaiting exec/leadership + content-team-lead review” per p15-partner nextActions. This audit memo is a Pierce-side artifact prepared for the next ingestion-team coordination conversation, not yet a leadership-handed deliverable.


Reference: AI Content Vetting Policy v1.9.11 · Scaled Content Policy §6/§7 · 2026-05-12 Insights Agent demo intake (ops-hub/sessions/2026-05.md “2026-05-14 late—Insights Agent intake distribution”).