Bounded acquisition closed, 9 September 2026. One new substantive primary original was acquired and read. Two final deliverables were located through the official results index but their download service returned technical errors. No dataset, model or live platform was accessed. Existing D2.1 was neither reacquired nor reread in this lane; root owns its selected annex reading.
The new primary finding
The [CORDIS final-period report](http[local research file] covering 1 March 2024–28 February 2026 and last updated 3 August 2026, reports that the final Disinformation Warning System combines outputs from six AI tools, using an ensemble of LightGBM classifiers with isotonic regression calibration, producing triage categories and feature attribution. This is a considerably more specific mechanism than an unspecified risk score. Locator: “Work performed,” second reporting period, “Disinformation Warning System (WP3)”; official-reader line 141.
The same report describes an enhanced Human Validation Dashboard and iterative user evaluation. It reports three evaluation rounds involving 206 users, including media professionals, fact-checkers, policy experts and students. These are project-reported results hosted by the Commission, not this investigation's reproduced performance or an independent Commission audit. Locators: WP5 and WP6, lines 143–144.
Its approximately 56% weighted F1 concerns the separate textual disinformation-signals tool, which detects 39 signals across six tactics and eight languages. It is not the final warning system's accuracy or a probability that a particular item is false. Locator: WP3 textual component, line 137. No model or benchmark was run here.
Where the stages remain distinct
Collection, annotation, inference and action cannot be collapsed into one decision. The report separately describes data collection and ground-truth resources, component analyses, an aggregate warning system, human validation, and intended reporting to media professionals and policymakers. Its collection counts are not counts of adjudicated misinformation. Its component metrics are not measures of subsequent journalistic decisions or advertising restrictions.
The published ensemble/calibration description does not disclose, in this read record, the six selected inputs, exact model weights, thresholds or triage-category definitions; who approved them; the training/validation split for the aggregate system; or how competing human assessments alter a score. “Human Validation Dashboard” is a named component, not an acquired override policy. Nor does a configurable interface establish authority to change model thresholds. The note therefore does not assert a reviewer veto, automatic score replacement, publisher appeal, correction propagation or final action right.
The report identifies intended users and a research evaluation population. It does not provide an actual recipient distribution list, an executed downstream-use agreement, or an advertising-system interface. The strongest supported interpretation remains a research triage system intended to assist people. A route from its score to an ad ban, account restriction or government enforcement decision is not established by these records. That limit does not prove no subsequent use exists.
Exact unavailable instruments and access boundary
The official [project results index](http[local research file] last updated 17 June 2026, exposes these two public records:
| Indexed record | Actual official module | Access result |
|---|---|---|
| Final release of disinformation detection tools and warning system | [Document 080166e529e5a48d](http[local research file] | Module HTML, followed through its own meta-refresh to technical error HDS-010; no deliverable bytes acquired. |
| Evaluation and impact assessment report | [Document 080166e529e59930](http[local research file] | Same technical error, with a distinct document identifier; no deliverable bytes acquired. |
The download HTML explicitly supplied the next URL. Following that ordinary redirect was not a bypass. Both returned-error originals and their module originals are retained. Their HTTP 200 status does not turn the error pages into the requested reports. No retry after HDS-010, alternate mirror, login or helpdesk contact followed.
The public platform About page was discovered by search at http[local research file], but the reader returned 502 and the one ordinary GET failed verification of an expired certificate. That route was closed without disabling certificate checks. Search-indexed explanatory text remains discovery, not an independently read rule or current interface observation. The main project homepage reader separately returned a non-retryable unsafe-open error and was not fetched through another method.
Precise reopening condition
The two indexed final records are the immediate expected instruments for the actual aggregate scoring and evaluation method. The known holders are the project consortium/coordinator and the Commission's public document service. A useful next record would expose the approved score/category thresholds and evaluation protocol, or a reviewer workflow showing how disputed labels and overrides are authorized and recorded. GDI's annotation codebook remains unacquired from the previous run. The public final-report narrative does not substitute for any of these instruments.
This run produces a concrete method-level finding and a bounded document-access gap. No further source search, graph update, operational experiment or follow-on branch was undertaken.
Read and custody scope
dss-reporting.html and dss-reporting.derived.txt: full substantive final-period report read, from reporting-period heading through uptake needs, plus date field. Official-reader equivalents: lines 119–159. Original-source header was inspected only for record identity; no new grant-accounting conclusions are drawn. This is a current snapshot of a historical reporting period, not a claim that a live service is now operating.
dss-index.reader.txt: retained official-reader extract of deliverable listings, lines 119–199. Discovery index only; no full publication-list audit. dss-final-module.html, dss-evaluation-module.html, and the two *.response.html files retain the actual download wrappers/errors. dss-transport.json records URLs, timestamps, status and checksums. Exact hashes, source states and closed routes are in dss-captures.json.