[ThomsenDrake]

Investigators can now compare cases through the existing isolated ChatGPT subscription integration with exactly gpt-6-astra, inspect cited findings, challenge their support, and preserve local work products. The comparison result is the visual focal point: a plain-language summary and numbered findings appear before setup details, with source passages available beside each finding.

Delivered behavior

• Find shared reviewed identities and recurring relationship/event structures involving different actors, using bounded accepted knowledge and source passages. Findings explain differences, ordinary explanations, uncertainty, source independence and coverage limits; recurrence does not establish coordination or causation.
• Retain exact preview/approval/run controls, current source governance and revision checks. Save, annotate, dismiss and reopen hypotheses; supporting corrections mark findings stale and permit deliberate scoped reruns. Excluded sources are removed from dependent views and counts.
• Ask cited cross-case questions, inspect sourced timelines, and edit local notes, briefs and unresolved questions. Record requests, correspondence and responses with supporting/contradicting evidence links; drafting sends nothing.
• Preserve the existing manual comparison experience. The corrected main roadmap requires ambient whole-ontology pattern and single-investigation gap discovery next, with a concise implementation handoff. Standing authorization is a future capability; no background execution is enabled here. Phase 5 evidence pursuit remains separate.

Validation and limits

• Real public procurement records plus live Astra demonstrated an unnamed shared actor, grounded roles/time, working citations and restart persistence across two cases about the same procurement. This is not evidence of independent unrelated recurrence.
• Clearly labelled synthetic controls plus live Astra demonstrated a different-actor recurring relationship, an unrelated negative case, and duplicate/shared-address/same-name distinctions. Controlled runtime and regression evidence also cover corrections, exclusion, saved work products and recovery. Certain multi-event sequences have automated fixture evidence only.
• Final unchanged product code at 450cc424 passed npm run verify: 277 test files passed / 3 skipped; 4,221 tests passed / 5 skipped; typecheck, assurance checks and production build passed. Both presentation changes have focused review and browser evidence. Publication commits change documentation only; handoff links/paths and git diff --check pass. PR CI passed full npm run verify on final head 6417cd5739b3bc49b25cc59fbb32752cfa4cac46 before merge.
• The six-attempt live acceptance cap is exhausted; this publication made no provider calls. New prompt wording has not had another live run, and historical provider output is preserved. No cross-provider review routing was enabled. The installed Devin PR check reports a pass but explicitly skipped its full review because its trial expired and it has no credits; this is not completed review. Recorded local code reviews remain the review evidence.
• Current “all” comparison examines the first 12 eligible cases, at most 100 assertions and 24 passages, with no persisted whole-ontology progress cursor. There is no ambient analysis or structured single-investigation gap detector. The roadmap records the user's future “Automatic within configured limits” preference without granting execution today.

Evidence: Phase 4 acceptance, plain language, result hierarchy.

Operator workflow: personal-use runbook. On the retained local demo, open Investigate, inspect the latest comparison, follow its sources and challenge supporting knowledge; deliberate new analysis requires its own applicable authorization. Private records, credentials, runtime workspaces and browser evidence remain outside Git.

Continuation: single product roadmap and ambient discovery handoff. Built on Phase 3 PR #45; base is neo.