How the signals are made, and how we know they’re right.
Every signal we publish is an extract from a document an NHS trust has itself published. Nothing is inferred, aggregated from rumour, or bought from a third party. This page describes exactly how the process works and what its measured accuracy is, because a data product that won’t show its method is asking for trust it hasn’t earned.
1. Collection
Boards of NHS trusts meet in public and publish their papers under the Public Bodies (Admission to Meetings) Act 1960 and their own publication schemes. We check every covered trust weekly, follow each board’s own publishing rhythm rather than a fixed calendar, and record the source URL, a SHA-256 hash and the retrieval time for every document we read. Packs routinely run 100–300 pages.
2. Extraction
Each pack is read page by page and analysed in chunks that preserve page numbers, so every signal can point at the page it came from. We extract only forward-looking, actionable items across six categories — capital, digital & EPR, estates, service change, procurement and workforce — and record the decision stage: mentioned, outline business case, full business case, approved, or tender planned. Routine performance reporting, restatements of past decisions and administrative matters are deliberately excluded.
Every signal carries a confidence score. Nothing below 0.5 reaches a subscriber. Duplicate representations of the same decision — the same scheme described in the CEO report, the finance paper and a committee summary — are resolved to a single entry that keeps the strongest evidence.
3. Accuracy — measured, not asserted
Two complete board packs were read end to end by hand, independently of the pipeline, to build a ground-truth list of every qualifying capital and digital/EPR signal. One was read by the build team; the other by an independent reviewer with no access to the pipeline’s output. Every extracted signal was then adjudicated against the source text.
| Test | Precision |
|---|---|
| At the 0.5 confidence threshold we publish at | 33/33 — 100% |
| Including items the model itself flagged as low-confidence | 41/43 — 95.3% |
Guy’s and St Thomas’ (96pp, 22 July 2026) and Newcastle upon Tyne Hospitals (300pp, 31 July 2026). Both items that failed were self-flagged below the publication threshold — which is the threshold doing its job. We re-run this test as the method changes and will publish the number whichever way it moves.
4. Copyright and licensing
We extract facts — dates, sums, programmes, decisions — which are not themselves subject to copyright, and quote at most 25 words from any document. Every signal links to the original on the trust’s own website. We do not re-host trusts’ documents. Where a trust publishes under the Open Government Licence v3.0, its content is used and attributed accordingly; where a trust’s terms are more restrictive, we hold to facts and links only. Material owned by third parties inside a pack — commissioned consultancy reports, for example — is excluded. We publish nothing from private (Part 2) sessions, which are not published in the first place, and nothing identifying patients or non-director staff.
5. What we don’t claim
Coverage is 20 trusts of roughly 215, and we say so on the coverage page rather than implying national reach. Board papers show intent and timing, not awarded contracts — for awards, the contracts register is the right source and we are complementary to it, not a substitute. And a board can change its mind: a signal is an accurate record of what was said at a meeting, not a guarantee of what will happen.