# AI Risk Trajectory — manual review protocol

Updates run only when Nuvastra requests them. No autonomous crawler, scheduled AI review or public publication endpoint is enabled.

## Request
Update AI Risk Trajectory through today. Keep the scientific structure fixed. Check new and contrary primary evidence, preserve the previous release, and publish only supported record changes with a dated review log.

## Operator sequence
1. Read data/releases/index.json and the active JSON snapshot. Establish the actual review window and scope.
2. Check primary publications. Preserve publication, event and retrieval dates separately, exact source locations, authorship and limitations.
3. Assess propositions individually. Do not turn a model release, benchmark or policy announcement into an observed harmful outcome. Search contrary evidence.
4. Prepare a new dated JSON snapshot using stable IDs. Update the question answers, briefing, source catalogue, records, controls, events and correction notes consistently. Never describe not-reviewed records as unchanged after review.
5. Run `node scripts/review.cjs validate NEW.json` and `node scripts/review.cjs compare OLD.json NEW.json`.
6. Run `node scripts/review.cjs publish NEW.json`. The command refuses an existing ID, backdating, an incorrect predecessor or an undocumented methodology/topology change.
7. Run the build. It validates data, runs publication tests, regenerates current imports, produces a comparison, preserves old source data with checksums, exports dated JSON, and builds citation pages and sitemap.
8. Deploy through AppDeploy and check build/runtime status. Execute desktop/mobile workflows: exact links, sources, reading depths, exports, bounded Ask and its failure state.
9. Publish the review and briefing without implying external scientific validation. A methodological correction is not an incident alert.

## Preserving history
The 17 September 2026 source model remains archived and superseded. Later snapshots are append-only through the publication command. The manifest contains checksums; prior records must not be silently rewritten. A new methodology is explicitly non-comparable instead of a jump in a danger score.

## Limits
AI-assisted evidence preparation is not external expert review. The present form stores challenges but is not a completed institutional reviewer workflow. Search visibility, scientific endorsement and comprehension by real users require separate verification.