Q3 2026 Claim Review Bulletin: which claims moved, which held, and what the EU AI Act enforcement window did to the corpus

This is Agent Mode AI’s second Quarterly Claim Review Bulletin. The first bulletin shipped 29 April 2026 covering the publication’s pre-relaunch corpus and the immediate post-relaunch additions through end of Q2. This bulletin covers the 91 days from 1 May 2026 to 30 July 2026 and is the first cross-cycle read on the Holding-up discipline at scale: most of the AM-, OPS-, and now RES-* claims published in Q2 have cleared their first scheduled review, and the verdict shifts that survived first contact with cross-cycle scrutiny are visible on the ledger.

The bulletin ships three days before the 2 August 2026 EU AI Act deployer-obligations enforcement window opens, which is the operational deadline that determines whether several deadline-anchored claims hold or move to Partial in the Q4 bulletin. The cadence is deliberate: by pegging nextReviewDate to the milestone rather than to calendar boundaries, the publication can mark its own claims wrong against post-enforcement reality without waiting for the calendar to catch up.

What the ledger looks like at end of Q3 2026

Three observations from the aggregate.

The corpus has crossed its first cross-cycle threshold. Most AM-* claims published in April-May 2026 carry 60-day cadences; most OPS-* claims carry 30-45-day cadences. By 30 July 2026, every claim with a Q2 publication date has cleared at least one scheduled review. The Q3 bulletin is therefore the first one whose data measures cross-cycle durability rather than publish-time discipline. The honest read: the holding rate is lower than the Q2 baseline, which is what should happen as evidence accumulates and the easier claims get their first stress test.

Cadence is current across three registers. No live claim is past its scheduled next-review date without a logged review. The discipline holds at the larger corpus scale (AM-* + OPS-* + RES-*) it now covers. The next bulletin will surface any cadence drift introduced by the EU AI Act enforcement workload; the publication’s editorial budget for Q4 is built around that risk.

The verdict shifts cluster in three predictable surfaces. The corpus pieces most exposed to verdict shifts in Q3 are the stat-correction cluster (claims about specific quantitative bands published by McKinsey, MIT NANDA, Stanford DEL, Gartner, where the underlying datasets were last refreshed in Q1/Q2 2026), the vendor-product cluster (claims about specific product capabilities, pricing tiers, or roadmap positions, where vendor changes between bulletins are common), and the deadline-anchored cluster (the eleven claims tied to the 2 August 2026 enforcement window). The Q3 ledger reflects this: the verdict changes recorded in the prior 91 days came almost entirely from these three surfaces, and the Holding-rate decline maps cleanly onto them.

Methodology, briefly

A claim moves from Holding to Partial when the spine still holds but a specific quantitative band, supporting sub-claim, or cited primary source has shifted. The article body remains live with a correction banner; the verdict_history field in /facts.json carries the dated note. A claim moves from Holding (or Partial) to Not holding when the spine itself has been falsified or materially overtaken. The article remains at its URL with a correction log; the claim is not silently removed; the next-review field is set to null because no further review is scheduled. The original sentence stays visible at its permalink, the article body carries a correction banner, and the full record sits at /corrections/ and /retractions/.

The methodology has not changed since Q2. The discipline has compounded: more claims now sit in the cross-cycle window, more verdict changes are documented in public, and the editorial cost of the discipline is now visible across two bulletins rather than one. That visibility is the reason the bulletins exist.

A new register: Resources (RES-*)

In May 2026, the publication added a third register: Resources. Where AM-* is the parent reporting register and OPS-* is the operators-register, RES-* covers downloadable templates and procurement artefacts that the publication ships as editorial work in their own right. The catalogue is at /resources/; the live ledger filters with ?segment=resources.

The five tools that opened the register on 4 May 2026:

• RES-001: AI Vendor Security Questionnaire (Procurement, 90-day cadence)
• RES-002: AI DPIA Template (Compliance, deadline-anchored to 2 Aug 2026)
• RES-003: Agent Incident Runbook (Operations, 60-day cadence)
• RES-004: Works Council AI Notification Packet (Compliance, deadline-anchored to 2 Aug 2026)
• RES-005: AI MSA Red-Team Checklist (Procurement, 60-day cadence)

The methodological choice that drove the register: each tool is its own editorial artefact (1,500-2,500-word explainer plus the downloadable Markdown plus CSV), not a thin “here’s a template, click download” page. Each carries one falsifiable RES-* claim about the tool’s premise. “This 47-question questionnaire covers the actual 2026 AI-MSA failure modes.” “This 38-item checklist surfaces the seven clause families where 2025-2026 enterprise AI MSAs cluster their failure modes.” Each premise is testable. Each next-review date is on the ledger. The verdicts will move the same way AM-* and OPS-* verdicts move when evidence shifts.

The deadline-anchored next-review dates on RES-001, RES-002, and RES-004 (all carrying nextReviewDate: 2 Aug 2026) make the Q4 bulletin the first read on whether these tools still match the regulatory reality after the first wave of enforcement actions. RES-003 and RES-005 carry 60-day cadences because the agent action surface and vendor MSA language patterns evolve faster than the regulatory surface.

The deadline-anchored cluster: 11 claims, 3 days out

AM-127 (source: 90 days to EU AI Act enforcement) is the editorial spine of the deadline-anchored cluster. The claim, made on 3 May 2026 with a 60-day cadence, is that of the eleven claims the publication has published against the 2 August 2026 enforcement deadline, the four operational-evidence claims (AM-108 data residency, AM-046 audit-evidence under four hours, AM-117 AI-BOM procurement, AM-120 works council workflow) carry materially higher risk of moving from Holding to Partial in Q3 than the two governance-process claims (AM-047 Head of AI Governance role, AM-051 centralised-vs-federated). Materially higher is defined as: at least three of the four operational-evidence claims will be downgraded by 1 October 2026, while at least one of the two governance-process claims will remain Holding.

This Q3 bulletin ships three days before the enforcement window opens. The first cross-cycle review on AM-127 happened on 1 July 2026 with no verdict change; the second falls on 1 September 2026, by which point the first wave of enforcement actions will be observable. The Q4 bulletin in late October will carry the falsifiable readout: either three of the four operational claims have downgraded (AM-127 holds), or AM-127 itself moves to Not holding because the prediction was wrong.

The eleven cited claims are visible at the Holding-up index filtered for ?topic=agentic-ai-governance. Their pre-enforcement state on 30 July 2026: all eleven still Holding. Q4 will report which moved.

Cross-linkage with AM-* and OPS-*

Each RES-* tool cross-links to the AM-* and OPS-* articles whose analysis underpinned its construction. RES-001 cross-links to AM-127 (EU AI Act enforcement), AM-029 (non-human identity), and AM-121 (IT operations reality). RES-004 cross-links to OPS-049 (Mittelstand BetrVG/DSGVO deployment), OPS-047 (SMB EU AI Act Annex III hiring), and AM-127. The cross-link is bidirectional in the editorial intent (the underpinning article gets a “Resource for this analysis” footer when the next AM-* or OPS-* revision lands); for now it remains unidirectional from the tool back to the analysis. The bidirectional linking is the discoverability bridge that turns the Resources register into corpus depth rather than a separate silo.

What the Q3 numbers mean for the Q4 bulletin

The Q4 2026 bulletin shipping at end of October 2026 will be the first post-EU-AI-Act-enforcement quarterly read. The deadline-anchored cluster (11 claims), the RES-* register’s first three deadline-tied tools (RES-001, RES-002, RES-004), and the AM-127 falsifiable prediction all clear their next reviews in that window. Three readouts that the Q4 bulletin will surface explicitly:

Whether AM-127’s prediction held. The numerical threshold (3 of 4 operational claims downgraded by 1 October 2026, at least 1 of 2 governance-process claims still Holding) is auditable on the visible status of the eleven cited claims. Either the prediction held and AM-127 stays Holding, or it did not and AM-127 moves to Partial or Not holding. This is the publication’s first deliberately-falsifiable predictive claim and the first cross-quarter test of whether the prediction discipline produces useful procurement signal beyond the live-ledger discipline.

How the RES- tools held against post-enforcement reality.* RES-001, RES-002, and RES-004 will get their first cross-cycle review against actual enforcement actions, supervisory-authority guidance, and customer-deployment patterns from August-October 2026. The Q4 readout will name which sections of which tools moved.

What the operations-register cohort looks like 90 days post-2-Aug. The OPS-* register’s deadline-adjacent claims (OPS-038 collective-agreement triggers, OPS-047 SMB Annex III hiring, OPS-049 Mittelstand BetrVG/DSGVO, OPS-040 ZZP AI displacement) all carry 30-45-day cadences and will have multiple reviews logged across August, September, and October. The Q4 bulletin will carry the rolled-up read on whether the SMB / micro-SMB advisory line held against the enforcement reality.

The discipline only carries weight if it survives contact with the deadline. That contact happens on 2 August 2026. The Q4 bulletin reports the result.

What this bulletin is not

This bulletin is not a marketing document. The holding-rate number is published in part to be testable against future quarterly data; an artificially high holding rate is a marker of editorial conservatism (claims pitched too narrowly to fail) rather than editorial honesty. The Q3 readout deliberately surfaces the verdict declines that the Q2 bulletin could not yet measure, and the Q4 bulletin will surface the deadline-driven shifts.

This bulletin is not a predictive forecast. AM-127 carries one explicit predictive claim with a falsifiable threshold; the rest of the bulletin is retrospective on the prior 91 days. Procurement teams citing the bulletin should treat it as evidence about the publication’s discipline, not as forecast input for their own deployments.

This bulletin is not a substitute for the live ledger. The bulletin compresses the prior 91 days into a quarterly read; the Holding-up ledger is the canonical source for any specific claim’s current status. The bulletin’s purpose is the editorial frame around the discipline, not the discipline itself.

The discipline is the bulletins shipping on cadence, the verdicts moving with the evidence, the corrections logged in public, and the original sentences staying visible at their permalinks. Q4 reports the next 91 days.