Evidence maintenance: current process and planned automation
This page distinguishes the evidence-maintenance process used so far from the automation still to be implemented. Reports and proposed changes are open to inspection. An automated search is not a judgement about an instrument, and finding no new evidence does not establish that none exists.
Current process
The September 2026 sweep used web search because its execution environment blocked direct access to PubMed, Europe PMC, Crossref and publisher pages. It produced a report and a proposal that was reviewed and merged. The registry pages were then regenerated by hand. The September report records the published results. The proposal itself was not in the public repository, and this cycle does not establish a running public end-to-end update pipeline.
Citations and review flags may be proposed for human review. Grades and statuses remain frozen, and licence classifications require source verification and a recorded decision. The public changelog records accepted changes. Regeneration of the registry pages on merge is not yet automatic.
What the automation may and may not do
| Permitted proposals | Prohibited automated changes |
|---|---|
| Add a citation with its source and flag the relevant cell for review | Change a grade, status, evidence form or population judgement |
| Record the scope and date of a completed search | Treat a search run as human confirmation of a grade |
| Flag a change to a licence source | Change a licence classification |
| Suggest an instrument for consideration | Admit an instrument without the published review process |
A search run does not authorise a grade change. New evidence is reviewed through the public corrections process. Findings that could affect a grade, status or licence classification are referred for human review through the public corrections process; automated sweeps do not apply these changes.
When it fails
The failure mode this design guards hardest against is silence: an early version of the sweep failed silently for a month in 2026, and nobody noticed until the next run. The session is now instructed to notify its operator on success and on failure. The planned failure monitor will check that a monthly sweep produced the expected report or proposal and will raise a visible alert if it did not. The monitor and a replay fixture using saved source responses are not yet running. A missing report must be treated as an incomplete maintenance cycle, not as a month with no new evidence.
Planned, and stated as planned
Planned work comprises direct scholarly-index retrieval, a reproducible query file, a measured historical retrieval check, automatic proposal creation, registry regeneration on merge, a monthly failure monitor, a saved-response regression fixture, and licence and retraction watchers. Each capability will be described as running only after its workflow and run log are available here. Search coverage will remain limited by the sources, queries and access available.
The blind spots, stated
Web search and scholarly indexes have different coverage. Relevant work may be missed because of indexing, terminology, language, access restrictions or unavailable manuals and technical reports. Source failures and uncompleted searches must be reported separately from completed searches that found no candidates.