# ClinLoopAI research preview — 30 October 2026

This update publishes recovered engineering fixes and the synthetic preview. It is not approved for patient care. Independent clinical accuracy and patient benefit remain unestablished.

## Engineering changes

- Preserve report assertions, addenda and lesion identity; generic abnormal lab flags no longer establish positive infection detection.
- Reject unusable numeric measurements as completed follow-up. Conflicting imaging records remain reviewable and cannot supply conflicting completion evidence.
- Bind patient graphs to one identity, replace stale graph histories, preserve source/review flags and check anatomical follow-up compatibility.
- Stage built-in FHIR/Bulk synchronization before a combined local commit. Nested writers use savepoints; stale stages fail for retry.
- Stop completed-late tasks entering active escalation/outreach; withdraw obsolete unsent drafts on resolution or reopening.
- Keep the entire eligible evaluation denominator, distinguish abstention from negative results and withhold patient intervals without authentic grouping.

## Measured evidence

Engine revision: `9332100304846852c8d35d6b77820e5941ad566f`.

The recovered engine passed 1,510 software tests and 239 subtests, plus 26 separate Git-dependent checks. Synthetic load and engine benchmarks passed. The optional research imaging model/weights were unavailable. Release packaging and deployment checks are recorded separately in GitHub Actions.

On 753 previously exposed synthetic reports, exact rule agreement remains 635/753. Expected obligations found increase from 466/582 to 467/582; 115 misses and 37 extra obligations remain. This includes human-review alerts and is not autonomous clinical performance. The separate conservative autonomous result accepts 325 reports, refers 428 for review and matches 62/582 expected obligations without whole-report abstention.

The 60 exposed timelines retain 35 exact status sets and 53/78 matching rule statuses. No false closures were observed in those particular fixtures; that does not establish that false closure is impossible. Four transformation checks each match 753/753 cases. Test expectations were not relabelled.

## Publication and cleanup

GitHub Pages publishes an allowlisted web artifact from a dedicated publication branch rather than the repository root. Existing repository CI is preserved; the proposed workflow automation remains local because the publisher token does not have workflow scope. Source code, tests, runtime data, caches, private setup files and recovery archives are excluded. The synthetic read-only snapshot is regenerated from repository examples. Existing recovery evidence and source archives are preserved outside the public artifact.

The version and aggregate verification status are available in `version.json` and `data/engineering_release_status.json`. Historical connection, research and model results retain their dates and do not become current validation evidence.

## Remaining requirements

Named clinician/source review and independently adjudicated patient/site/time/language evaluation remain pending. Multiple-lesion/serial-task identity, FHIR deletion/merge/reassignment reconciliation, hospital-scale operation, model/OCR calibration, endpoint transport policies and production security controls require further work. Patient grouping is unavailable for legacy fixtures, so patient confidence intervals are withheld.

The public site has no clinical backend. Hospital and model connections are not activated by this publication, and real patient information must not be entered into the preview.

## Public interface update: 2026-10-30.consistency1

All thirteen main pages and the two synthetic connection screens now use the same
header, navigation order, English/Korean controls, Light/Dark controls and visible
release version. The report-model menu label reflects its saved-output and optional
local-model comparison workflow. Sandbox pages retain their separate connection code
and original connection allowlists.

Language and theme follow known page links and the mobile page picker. Only supported
display values are forwarded; authorization codes, tokens, source text and patient
identifiers are never copied into those links. Isolated pages do not persist display
preferences. The existing SMART authorization handshake is unchanged.

The header generator reads the version from `version.json` and refreshes asset cache
keys on all fifteen screens. Korean evidence notices now translate their explanatory
text without changing the counts. Source reports and model responses stay in their
original language. Historical research evidence remains dated and separate from the
current engineering figures.

Desktop, phone and tablet presentation checks cover all fifteen screens. These checks
verify display and navigation behavior; they do not establish clinical accuracy or
validate a hospital installation.