🚀 V3 pivot landed — 2026-05-15

Opus 4.7 direct extraction replaces 5-stage Doc AI/Vision hybrid
extractor-v2 single-shots PDFs page-by-page via Claude Opus 4.7 with tool-use schema forcing. 100% signature + dispensing capture validated on 3 real Biodeal BMRs. Tree review + cross_validating + citation_verifying STATES REMOVED from FSM. Pipeline simplified: pending → indexing → extracting → auditing → concluding → reporting → completed. document-intelligence ARCHIVED on GitHub, container stopped on Hetzner.

📦 Today's velocity

5 PRs merged · 4 PRs awaiting prtk-human
Merged: cd#49 (doc-intel rip, 11:06 IST). Awaiting prtk: api-gw#47, orch#67, orch#68, cd#50 (all 2/2 approved by 11:48 IST). Tree-kill arc total: -1410 LOC across 4 repos.
Active run: tier-2 BMR runner b2v82uobw
6/7 BMRs done at 11:53 IST. Last-known status: throughput steady on extractor-v2 path. CLI subprocess mode (Redis publisher TBD).

⚠ Open issues

extractor-v2 missing 17 fields (CIT-NEW)
Regulatory schema audit shows 17 gaps vs full BMR spec. Prioritize per L1 rule coverage.
MFR extraction TBD (extract_mfr_v3.py)
Per kushal call — separate codepath from BMR. Not yet started.
extractor-v2 Redis publisher missing
Currently CLI subprocess only. Needs DocumentExtracted Stream publisher to integrate with orchestrator FSM.
Frontend (CIT-21) still backlog
QA viewer (citral-test-frontend) covers internal review. End-user Next.js app unbuilt — MVP critical-path risk.
platform-svcs AuditTrail chain unwired
VerifyChainIntegrity returns FailedPrecondition. ~30 min fix. Compliance gap.

✅ Resolved this window

cd #49 — doc-intel rip from compose (V3 pivot, merged 11:06 IST)
api-gw #47 — costbreaker fix (CIT-387) + JWT WS query-param fix (CIT-NEW) — 2/2 approved
orch #67 — TreeBuilt event removal, FSM state-rip — 2/2 approved
orch #68 — TreeConfirmed event removal, simplified state machine — 2/2 approved
cd #50 — V3-ready compose, extractor-v2 path wiring — 2/2 approved
ae 6 PRs — CIT-465 consumer + ALCOA fix + prefetch fallback + sig-noise collapse
reg-knowledge — V3 rules seeded, defensive unmarshal
extractor-v2 — 9 IPC fields per CIT-487 USP <905>
Arc total — Tree-kill -1410 LOC, 4 repos, 5 merged + 4 pending

Completion

≥90% live
60–89% gaps remain
<60% major work left
infra (no %)

Edges

gRPC sync call
Redis Streams event
broken / unwired
drag to rotate · scroll to zoom · click a node

V3 Happy Path

Upload→ Ingest→ Extract→ Audit→ Conclude→ Report
FSM states: pending → indexing → extracting → auditing → concluding → reporting → completed. tree_review / cross_validating / citation_verifying REMOVED.

Today's velocity (2026-05-15)

PRs merged today5
PRs awaiting prtk-human4
Tree-kill LOC delta-1,410
Repos touched (arc)4
Repo archiveddocument-intelligence

Active runs

tier-2 BMR runnerb2v82uobw
progress (11:53 IST)6 / 7 BMRs
extractor pathCLI subprocess (Opus 4.7)

Tree-kill arc impact

Tree-review feature deprecated. Removed TreeBuilt + TreeConfirmed events. document-intelligence archived. Replaced with extractor-v2 — Claude Opus 4.7 direct PDF extraction (tool-use schema forcing, 300 DPI, 4-concurrent chunks). 100% signature + dispensing capture on 3 real Biodeal BMRs.