Skip to content

CODEX_HANDOFF.md

Latest commit

 

History

History
444 lines (387 loc) · 56.3 KB

CODEX_HANDOFF.md

File metadata and controls

444 lines (387 loc) · 56.3 KB

Codex Handoff

Updated: 2026-04-04

This repository is no longer in early MVP scaffolding. The core research prototype is implemented and validated, Phase 6 hosted smoke evidence capture is now complete, and the codebase now also carries the research-intel branch-wrap work that adds watchtower scheduling, audited watchtower automation, multi-run council comparison, longer-horizon council calibration, contributor packets plus collaborator bundles, richer stress-test experiments, and explicit governance for paid-source or donation-funded expansion. The best next work should now move from the expanded proof set and automated research-intel watchtower toward broader de-identified or externally supplied benchmark inputs and deeper live corpus quality built on top of the existing pilot packaging, proof surfaces, public benchmark pack, external evaluation bundle writer, and cited pancreatic oncology watchtower.

Current State

  • Deterministic pancreatic triage is implemented with section-aware sentence evidence, rationale codes, persistence, exports, and evaluation.
  • Reviewer workflow is implemented end to end with worklist filters, case detail, review actions, feedback capture, hybrid prioritization, and trial matching.
  • A second primary pillar, Research Intelligence, is now implemented with a sibling /research-intel workspace, a /api/v1/research-intel/* namespace, seeded pancreatic oncology source and topic catalogs, a lightweight domain graph, persisted run audits, normalized research documents, cited evidence spans, topic watchlists, council digests, opportunity proposals, and case-level research briefs.
  • Landing, about, quickstart, README, and handoff surfaces now frame Research Intelligence as the primary entry point, with triage, imports, benchmark proof, and case briefs positioned as explainable downstream applications of the discovery engine.
  • The discovery layer now also includes watchtower scheduling and automation: broader curated literature and trial watches, due-only ingest support, schedule snapshots, consecutive-failure tracking, a make research-intel-ingest-due path, a make research-intel-watchtower path, an audited /api/v1/research-intel/runs/watchtower trigger, and a dedicated /research-intel/schedule workspace.
  • Research-intel digests now compare themselves against recent runs, preserving confidence trend, disagreement deltas, recurring open questions, recurring disagreement points, resolved items, and a longer-horizon calibration snapshot instead of treating each digest as a sealed one-off summary.
  • Research-intel opportunities now carry contributor packets for issue, benchmark, dataset, rule, trial, case-brief, and tooling follow-through, with packet artifacts written under artifacts/research-intel/packets/ and collaborator bundle artifacts under artifacts/research-intel/collaborator-bundles/.
  • Safe experiment support is stronger too: benchmark and rule opportunities now expose both readiness and stress-test modes with dimension-level scoring, while still staying proposal-only and human-gated.
  • Governance for paid-source access and donation-funded operations is now documented in docs/RESEARCH_INTELLIGENCE_GOVERNANCE.md, and autonomous payment or procurement remains explicitly out of scope.
  • The research-intel watchtower is still fixture-first and operator-triggered for reproducibility, but it now also carries selectively live-ready curated connectors with pancreas-aware filtering and source-health tracking instead of remaining purely seeded.
  • The watchtower can now run as one audited automation tick: schedule state is captured before and after each tick, digest generation is policy-gated to avoid synthetic churn, and a cache-backed GitHub Actions workflow under .github/workflows/research-watchtower.yml can preserve state across hosted runs.
  • Phase 1 of the scientific-discovery pivot is now implemented: discovery ingest supports fixture-backed connector runs, selectively live-ready Europe PMC, ClinicalTrials.gov, NCI, FDA, and GitHub-backed open-source discovery connectors, source-health tracking, and per-document provenance plus novelty metadata.
  • Phase 2 is now underway as well: the pancreatic oncology graph is richer, typed, family-aware, and exposed through graph-backed entity resolution plus a dedicated /research-intel/graph surface.
  • Phase 3 is now underway as well: the research council persists independent opinions, peer critique, explicit confidence, open questions, evidence gaps, and next experiments.
  • Phase 4 is now underway too: opportunities are now typed discovery-to-action specs with evidence bundles, measurable outcomes, downstream artifact hints, and artifact-backed promotion flows.
  • Phase 5 is now underway too: benchmark and rule opportunities support safe experiment runs with keep-or-discard ratchet outcomes, experiment artifacts, and persisted last-result summaries on each opportunity.
  • Phase 6 pilot work is materially in place: observability, readiness probes, trusted-proxy auth, site scoping, pilot Docker overlays, checked-in env bundles, de-identified research views, FHIR ingestion, HL7 ORU ingestion, structured import metadata persistence, persisted import-run audit records, env-driven field-preference overrides for upstream variability, a dedicated web import workspace with recent-run audit visibility, and live report/FHIR/HL7 success, successful shared-visibility, report-path failed shared-visibility for validation_error, structured failed shared-visibility for FHIR unsupported_payload plus HL7 parse_error, report-path and structured-adapter site-scope rejection, report-path parse/validation, adapter-specific malformed-import, and generic plus structured cross-actor audit denial smoke coverage in both the header-auth and trusted-proxy pilot paths.
  • The web app now has clearer outside-collaborator surfaces: the home and about pages frame the product as an explainable, benchmarkable workflow stack, and /proof now publishes the checked-in demo benchmark snapshot plus a registry of published external benchmark packs as a public proof surface with side-by-side cross-pack comparison, anchored navigation, dataset coverage, queue previews, and reviewer-facing casebooks.
  • The repo now includes an adoption-facing quickstart, checked-in benchmark snapshots in docs/examples/demo-benchmark-current.json and docs/examples/demo-benchmark-current.md, and a refresh path through make benchmark-demo plus make refresh-demo-proof; that published proof now carries benchmark buckets, reviewer cues, expected rationale codes, per-case rules-versus-hybrid outcomes, and an expanded 10-report casebook spanning 5 recurring benchmark buckets.
  • The repo also now includes a public benchmark pack: docs/LABELING_GUIDE.md, docs/BENCHMARK_SUBMISSIONS.md, checked-in label and submission templates, a Pydantic benchmark submission schema, and a validator entrypoint through make validate-benchmark-submission SUBMISSION=....
  • The repo now also includes a comparable external evaluation bundle writer through scripts/run_external_eval.py plus make benchmark-external, producing JSON, Markdown, and validator-compatible submission-draft artifacts from label and prediction JSONL inputs; that external path now also preserves cohort labels, benchmark buckets, reviewer focus, expected rationale cues, optional deidentified report excerpts, dataset coverage, queue previews, and a reviewer-facing casebook in its generated outputs.
  • The external benchmark helper now also accepts an optional manifest JSON so collaborator-supplied datasets can carry dataset framing, labeling-policy notes, cohort descriptions, benchmark-bucket descriptions, and default operating-point metadata into the generated JSON, Markdown, submission draft, and checked-in proof sample without ad hoc hand editing.
  • Published external proof packs are now registry-driven through docs/examples/published-external-benchmarks.json, and make validate-strict now verifies that registry entries keep valid checked-in JSON artifact paths.
  • The repo now also carries two checked-in deidentified external proof packs: the retrospective-style multi-cohort sample under docs/examples/retrospective-benchmark-sample-* and a wording-variance challenge pack under docs/examples/wording-variance-benchmark-sample-*, each with reproducible JSON, Markdown, and submission outputs refreshed by dedicated Make targets.
  • The release-facing docs now also include a concrete operator runbook in docs/RELEASE_RUNBOOK.md plus a reusable docs/RELEASE_NOTES_TEMPLATE.md so the remaining hosted smoke evidence path is explicit rather than tribal.
  • FHIR DiagnosticReport imports now preserve patient, encounter, and accession metadata from inline Reference.identifier values when upstream payloads omit fully resolved Patient, Encounter, or ServiceRequest resources.
  • FHIR DiagnosticReport imports now also decode supported text-like presentedForm attachments, including base64 XHTML narratives with explicit charsets, bundled or contained Binary-backed attachment URLs, merge multiple supported attachments in order, suppress shorter overlapping fragments when a richer attachment already contains them, expand referenced Observation.component plus grouped Observation.hasMember findings into reviewer-visible report text, preserve interpretation plus referenceRange measurement context, fall back to conclusionCode text when free-text conclusion is absent, avoid duplicate or cyclic grouped-member expansion, and keep unsectioned attachment findings merged with conclusion text when that preserves a more reviewer-usable report shape.
  • HL7 ORU imports now decode base64 ED report text, normalize repeated OBX-5 values, respect custom MSH-2 component plus repetition separators, normalize common HL7 escape sequences, and clean composite metadata fields with subcomponent-aware extraction before the existing report-text assembly flows into triage and audit persistence.
  • Top-level repo docs now reflect the implemented platform instead of the earlier scaffold framing, and the repository includes checked-in contributor, security, and code-of-conduct docs appropriate for a near-1.0 open-source handoff.
  • The repo still includes checked-in GitHub Actions workflows for strict validation plus hosted base, attachment-backed FHIR success-path, report-path site-rejection, and structured adapter site-rejection pilot smoke coverage, and the hosted pilot workflow now supports narrower manual FHIR-only plus HL7-only dispatches with per-job raw-log and structured-summary artifacts. Those hosted slices are now verified with green March 25, 2026 runs for FHIR and HL7, while HL7 remains manual-only in the default weekly matrix by explicit decision.
  • The highest-value remaining work is no longer bootstrapping or late-Phase-6 smoke capture. The repo-side Phase 6D release polish is already in place, the research-first repositioning is visible across product and docs surfaces, the watchtower and council history loop are in place, the external starter pack matches the reviewer-facing proof shape, /proof compares the checked-in packs directly, recurring watchtower automation is in place, and digests now carry longer-horizon calibration plus collaborator bundle output, so the next work should move toward broader collaborator-supplied or deidentified benchmark inputs and deeper live discovery quality instead of reopening parser or smoke-baseline work.
  • The next benchmark-proof slice should focus on either a larger collaborator-supplied benchmark drop through the registry flow or deeper cross-pack analysis features such as bucket-level comparisons or reviewer-focused filtering, rather than just adding another static pack section.

Fresh Validation Status

Confirmed on 2026-04-04:

  • make validate-strict passes
  • Summary: 12 pass, 0 warn, 0 fail
  • Research-intel validation now covers seeded catalogs plus a temp-database ingest and digest run, reporting 9 sources, 7 topics, 25 graph nodes, 7 seeded documents, and a passing artifact-backed digest pipeline
  • Discovery ingest now also validates 9 discovery fixtures and 8 live-ready sources through the same catalog check, runs the watchtower pipeline in fixture mode for reproducible validation, and verifies both pre-run and post-run schedule state
  • Opportunity generation now also emits structured JSON and Markdown action specs under artifacts/research-intel/opportunities/, plus contributor packet artifacts under artifacts/research-intel/packets/
  • Research-intel validation now also runs one safe experiment and records a ratchet outcome in the strict pipeline summary
  • pytest now passes with 155 passed
  • cd apps/api && .venv/bin/python -m pytest tests/test_research_intel.py tests/test_research_intel_connectors.py -q passed with 11 passed
  • npm run lint and npm run build in apps/web both pass with the new /research-intel routes included in the build

Previously confirmed during the broader post-Phase-6 slice:

  • make validate-strict passes
  • Summary: 10 pass, 0 warn, 0 fail
  • API tests: 144 passed
  • apps/api/.venv/bin/python -m pytest apps/api/tests/test_imports.py -q passed with 34 passed, including split, overlapping, bundled or contained Binary-backed FHIR presentedForm, grouped and cycle-safe Observation.hasMember, conclusionCode, measurement interpretation plus referenceRange, and Observation.component coverage
  • apps/api/.venv/bin/python -m pytest apps/api/tests/test_summarize_pilot_smoke.py apps/api/tests/test_smoke_proxy_auth.py -q passed with 23 passed, covering the hosted smoke summary parser plus the attachment-backed FHIR smoke fixture and its expected presentedForm payload shape
  • apps/api/.venv/bin/python -m pytest apps/api/tests/test_smoke_proxy_auth.py -q passed with 21 passed, covering the attachment-backed FHIR smoke fixture and its expected presentedForm payload shape
  • Unsandboxed full local reruns of make pilot-proxy-demo-fhir-smoke and make pilot-header-demo-fhir-smoke both passed on 2026-03-25 after the compose fix, recording persisted import runs 50 and 51 plus web, /imports, visible-case, and reviewer round-trip verification
  • Web checks: npm run lint and npm run build now both pass through make validate-strict
  • Demo evaluation compare and sweep both run through the validation script
  • make refresh-demo-proof succeeded and refreshed the checked-in benchmark snapshot that powers /proof
  • apps/api/.venv/bin/python -m pytest apps/api/tests/test_demo_benchmark_snapshot.py apps/api/tests/test_exports_metrics.py -q passed with 11 passed, covering the casebook-backed proof snapshot and richer evaluation-case payload fields
  • apps/api/.venv/bin/python -m pytest apps/api/tests/test_validate_repo.py apps/api/tests/test_external_evaluation.py -q passed with 10 passed, covering the new published-registry validator plus the retrospective and wording-variance external sample packs
  • make validate-benchmark-submission SUBMISSION=docs/examples/benchmark-submission-template.json passed
  • cd apps/api && .venv/bin/python -m pytest tests/test_benchmark_submission.py -q passed
  • make benchmark-external LABELS=docs/examples/benchmark-label-template.jsonl PREDICTIONS=docs/examples/benchmark-prediction-template.jsonl OUT_DIR=/tmp/pancreatic-signal-external-eval BASENAME=template-external TOP_K=2 passed and wrote JSON, Markdown, and submission-draft artifacts
  • apps/api/.venv/bin/python -m pytest apps/api/tests/test_external_evaluation.py -q passed with 6 passed, covering manifest-driven dataset framing plus CLI override precedence for collaborator benchmark bundles
  • make benchmark-external-sample passed and wrote a retrospective-style external bundle from the checked-in sample pack
  • make refresh-external-sample-proof passed and refreshed the checked-in retrospective-style external benchmark proof under docs/examples/
  • make benchmark-external-wording-sample passed and wrote the wording-variance challenge bundle under artifacts/benchmarks/
  • make refresh-external-wording-sample-proof passed and refreshed the checked-in wording-variance benchmark proof under docs/examples/
  • make validate-benchmark-submission SUBMISSION=docs/examples/wording-variance-benchmark-sample-current-submission.json passed
  • make validate-benchmark-submission SUBMISSION=/tmp/pancreatic-signal-external-eval/template-external-submission.json passed against the generated draft
  • apps/api/.venv/bin/python -m pytest apps/api/tests/test_imports.py apps/api/tests/test_import_runs.py -q passed with 46 passed, including inline Reference.identifier extraction coverage, attachment-backed presentedForm decoding and audit coverage, plus custom MSH-2 delimiter, escape-sequence, subcomponent, and audit-visibility coverage
  • Targeted import / de-identification / export coverage also passes for the new import metadata surface
  • Import-run audit coverage now passes for success, update counts, validation failures, unsupported payloads, site-scope rejection, and audit-route access control
  • Config-override coverage now passes for one FHIR field-preference override and one HL7 field-preference override while preserving defaults
  • The new /imports workspace is included in the validated web build and uses persisted run IDs to deep-link failed submissions into their audit detail
  • The published demo proof currently includes a checked-in JSON plus Markdown benchmark summary for outside collaborators, and the web app reads that snapshot directly into dataset coverage cards, queue previews, and reviewer-facing casebook entries; the current checked-in corpus is 10 labeled reports across 5 recurring benchmark buckets, with hybrid-only recovery of the follow-up-only cases C-005 and C-008
  • The public external benchmark starter pack now uses the same reviewer-facing fields and publishes a casebook-shaped JSON plus Markdown bundle, so collaborators can benchmark less-synthetic datasets without losing the coverage, queue, and reviewer-cue structure established by the demo proof
  • The repo now also includes a 12-case deidentified retrospective-style multi-cohort sample pack plus a 9-case wording-variance challenge pack, both with report excerpts, reviewer cues, cohort coverage, and visible misses or confounders that demonstrate the same casebook structure on less-synthetic external datasets; /proof now renders both packs alongside the demo comparison and compares them directly instead of leaving them docs-only
  • Both pilot demo overlays resolve successfully through docker compose config
  • The smoke helper now supports both trusted-identity proxy mode and field-level header-auth mode, plus live verification against /api/v1/imports/reports, /api/v1/imports/fhir/diagnostic-reports, and /api/v1/imports/hl7/oru
  • The smoke helper now also supports live site_scope_rejection verification through /api/v1/imports/reports, including persisted failed run IDs and zero run-specific visible-case assertions
  • The smoke helper now also supports live structured adapter site_scope_rejection verification through /api/v1/imports/fhir/diagnostic-reports and /api/v1/imports/hl7/oru, including persisted failed run IDs and zero run-specific visible-case assertions
  • The smoke helper now also supports live structured adapter audit-visibility verification by reusing persisted FHIR and HL7 site_scope_rejection runs, proving the owner can inspect both while a second scoped actor receives 404 on both details and does not see either in the recent-run list
  • The smoke helper now also supports live validation_error and parse_error verification through /api/v1/imports/reports, including persisted failed run IDs and zero run-specific visible-case assertions
  • The smoke helper now also supports live structured adapter failure verification through /api/v1/imports/fhir/diagnostic-reports and /api/v1/imports/hl7/oru, including persisted unsupported_payload and parse_error run IDs plus zero run-specific visible-case assertions
  • The smoke helper now also supports live audit-visibility verification by reusing a persisted site_scope_rejection run, proving the owner can inspect it while a second scoped actor receives 404 on detail and does not see it in the recent-run list
  • The smoke helper now also supports live shared-visibility verification by reusing a successful /api/v1/imports/reports run, proving a second scoped actor can inspect the same completed run detail and recent-run list entry
  • The smoke helper now also supports live structured shared-visibility verification by reusing successful FHIR and HL7 imports, proving a second scoped actor can inspect both completed structured runs and their recent-run list entries
  • The smoke helper now also supports live failed-run shared-visibility verification by reusing a persisted /api/v1/imports/reports validation_error run, proving a second same-site actor can inspect the failed detail and recent-run entry even when imported_sites is empty
  • The smoke helper now also supports live structured failed-run shared-visibility verification by reusing persisted FHIR unsupported_payload and HL7 parse_error runs, proving a second same-site actor can inspect both failed details and recent-run entries even when imported_sites is empty
  • The proxy demo overlay now defaults to an import-capable navigator identity so the built /imports workspace and the live proxy smoke path exercise the same capability class
  • The latest strict validation pass was rerun after widening the hosted pilot smoke workflow to include the attachment-backed FHIR success path and remains green
  • GitHub Actions now runs make validate-strict on pull requests, on main, and through manual workflow dispatch using a checked-in workflow under .github/workflows/validate.yml
  • The smoke helper's FHIR demo path now uses an attachment-backed DiagnosticReport.presentedForm XHTML narrative encoded as base64 UTF-16, so the existing pilot-*-fhir-smoke targets exercise the same supported text-like attachment decode path covered by API tests
  • A checked-in workflow under .github/workflows/pilot-smoke.yml now reuses make pilot-proxy-demo-smoke, make pilot-proxy-demo-fhir-smoke, make pilot-proxy-demo-site-rejection-smoke, make pilot-proxy-demo-adapter-site-rejection-smoke, make pilot-header-demo-smoke, make pilot-header-demo-fhir-smoke, make pilot-header-demo-site-rejection-smoke, and make pilot-header-demo-adapter-site-rejection-smoke on manual dispatch plus a weekly Monday schedule; it remains intentionally narrower than the full manual visibility and failure-path overlay matrix
  • That hosted pilot workflow now also supports workflow_dispatch with smoke_scope=fhir-success-only plus a manual-dispatch smoke_scope=hl7-success-only, and every hosted matrix job uploads both its pilot-smoke.log output and generated pilot-smoke-summary.json plus pilot-smoke-summary.md artifacts, including duration, absolute smoke timestamps, and exit code, to preserve exact run context once hosted execution starts
  • A bundled evidence builder now exists as make pilot-smoke-evidence SUMMARY_DIR=/path/to/downloaded/pilot-smoke-artifacts HL7_DECISION=pending|keep-manual|promote-default, producing pilot-smoke-evidence.json plus pilot-smoke-evidence.md from downloaded hosted summary artifacts so release notes and handoff updates do not depend on manual copy-paste
  • GitHub-hosted Pilot Smoke (fhir-success-only) run #23563902873 passed on 2026-03-25 across proxy and header auth, with summary artifacts pilot-smoke-summary-proxy-demo-fhir-smoke and pilot-smoke-summary-header-demo-fhir-smoke
  • GitHub-hosted Pilot Smoke (hl7-success-only) run #23564057337 passed on 2026-03-25 across proxy and header auth, with summary artifacts pilot-smoke-summary-proxy-demo-hl7-smoke and pilot-smoke-summary-header-demo-hl7-smoke
  • Bundled hosted evidence now records phase_complete: true plus hl7_hosting_mode: manual_only, with the explicit rationale that HL7 proved operable in GitHub Actions but still stays out of the default weekly matrix to limit recurring runtime and maintenance cost
  • Release-facing documentation now includes CHANGELOG.md and docs/RELEASE_READINESS.md

Last known good live deployment check:

  • The header-demo Docker stack was booted with the pilot overlay
  • make pilot-header-demo-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, report import via /api/v1/imports/reports, persisted import-run audit detail for run 41, visible cases, and a persisted reviewer-action round-trip
  • make pilot-header-demo-fhir-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, FHIR import via /api/v1/imports/fhir/diagnostic-reports, persisted import-run audit detail, visible cases, and a persisted reviewer-action round-trip
  • make pilot-header-demo-hl7-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, HL7 import via /api/v1/imports/hl7/oru, persisted import-run audit detail, visible cases, and a persisted reviewer-action round-trip
  • The proxy-demo Docker stack was booted with the pilot overlay
  • make pilot-proxy-demo-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, report import via /api/v1/imports/reports, persisted import-run audit detail for run 40, visible cases, and a persisted reviewer-action round-trip
  • make pilot-proxy-demo-fhir-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, FHIR import via /api/v1/imports/fhir/diagnostic-reports, persisted import-run audit detail, visible cases, and a persisted reviewer-action round-trip
  • make pilot-proxy-demo-hl7-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, HL7 import via /api/v1/imports/hl7/oru, persisted import-run audit detail, visible cases, and a persisted reviewer-action round-trip
  • make pilot-proxy-demo-site-rejection-smoke passed on 2026-03-22
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a persisted site_scope_rejection run via /api/v1/imports/reports, stable audit detail for run 42, and zero run-specific visible cases
  • make pilot-proxy-demo-adapter-site-rejection-smoke passed on 2026-03-22 local time
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR and HL7 site_scope_rejection runs via the structured import endpoints, stable audit detail for runs 44 and 45, and zero run-specific visible cases
  • make pilot-proxy-demo-adapter-audit-visibility-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR and HL7 site_scope_rejection runs via the structured import endpoints, owner visibility on both audit endpoints, alternate-actor 404 detail denial, alternate-actor omission from the recent-run list, and zero run-specific visible cases
  • make pilot-proxy-demo-parse-validation-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted validation_error and parse_error runs via /api/v1/imports/reports, stable audit detail, and zero run-specific visible cases
  • make pilot-proxy-demo-adapter-failure-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR unsupported_payload and HL7 parse_error runs via the structured import endpoints, stable audit detail, and zero run-specific visible cases
  • make pilot-proxy-demo-shared-visibility-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a successful /api/v1/imports/reports run, owner visibility on both audit endpoints, alternate-actor visibility on the same detail and recent-run list entry, visible run-specific cases, and a persisted reviewer-action round-trip
  • make pilot-proxy-demo-adapter-shared-visibility-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, successful FHIR plus HL7 structured imports, owner visibility on both audit endpoints, alternate-actor visibility on both run details and recent-run list entries, visible run-specific cases, and a persisted reviewer-action round-trip
  • make pilot-proxy-demo-audit-visibility-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a persisted site_scope_rejection run via /api/v1/imports/reports, owner visibility on both audit endpoints, alternate-actor 404 detail denial, alternate-actor omission from the recent-run list, and zero run-specific visible cases
  • make pilot-proxy-demo-failed-shared-visibility-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a persisted validation_error run via /api/v1/imports/reports, owner visibility on both audit endpoints, alternate-actor allow behavior on the same failed detail and recent-run list entry, and zero run-specific visible cases
  • make pilot-proxy-demo-adapter-failed-shared-visibility-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR unsupported_payload plus HL7 parse_error runs via the structured import endpoints, owner visibility on both audit endpoints, alternate-actor allow behavior on both failed details and recent-run list entries, and zero run-specific visible cases
  • make pilot-header-demo-site-rejection-smoke passed on 2026-03-22
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a persisted site_scope_rejection run via /api/v1/imports/reports, stable audit detail for run 43, and zero run-specific visible cases
  • make pilot-header-demo-adapter-site-rejection-smoke passed on 2026-03-22 local time
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR and HL7 site_scope_rejection runs via the structured import endpoints, stable audit detail for runs 46 and 47, and zero run-specific visible cases
  • make pilot-header-demo-adapter-audit-visibility-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR and HL7 site_scope_rejection runs via the structured import endpoints, owner visibility on both audit endpoints, alternate-actor 404 detail denial, alternate-actor omission from the recent-run list, and zero run-specific visible cases
  • make pilot-header-demo-parse-validation-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted validation_error and parse_error runs via /api/v1/imports/reports, stable audit detail, and zero run-specific visible cases
  • make pilot-header-demo-adapter-failure-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR unsupported_payload and HL7 parse_error runs via the structured import endpoints, stable audit detail, and zero run-specific visible cases
  • make pilot-header-demo-shared-visibility-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a successful /api/v1/imports/reports run, owner visibility on both audit endpoints, alternate-actor visibility on the same detail and recent-run list entry, visible run-specific cases, and a persisted reviewer-action round-trip
  • make pilot-header-demo-adapter-shared-visibility-smoke passed on 2026-03-21
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, successful FHIR plus HL7 structured imports, owner visibility on both audit endpoints, alternate-actor visibility on both run details and recent-run list entries, visible run-specific cases, and a persisted reviewer-action round-trip
  • make pilot-header-demo-audit-visibility-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a persisted site_scope_rejection run via /api/v1/imports/reports, owner visibility on both audit endpoints, alternate-actor 404 detail denial, alternate-actor omission from the recent-run list, and zero run-specific visible cases
  • make pilot-header-demo-failed-shared-visibility-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, a persisted validation_error run via /api/v1/imports/reports, owner visibility on both audit endpoints, alternate-actor allow behavior on the same failed detail and recent-run list entry, and zero run-specific visible cases
  • make pilot-header-demo-adapter-failed-shared-visibility-smoke passed on 2026-03-20
  • The smoke path verified API readiness, web readiness, /imports, /api/v1/auth/me, persisted FHIR unsupported_payload plus HL7 parse_error runs via the structured import endpoints, owner visibility on both audit endpoints, alternate-actor allow behavior on both failed details and recent-run list entries, and zero run-specific visible cases

Additional note from this slice:

  • make pilot-header-demo-up required host-level Docker access and completed successfully
  • The sandboxed make pilot-header-demo-fhir-smoke and make pilot-header-demo-hl7-smoke would have had the same localhost restriction as the existing smoke targets
  • Unsandboxed runs of both make pilot-header-demo-fhir-smoke and make pilot-header-demo-hl7-smoke passed end to end
  • make pilot-header-demo-down completed successfully after the live verification run
  • make validate-strict passed after adding the GitHub Actions validation workflow and the web lint check to the strict gate
  • The checked-in workflow at .github/workflows/validate.yml now runs make validate-strict on pull requests, on main, and through manual workflow dispatch
  • docker compose -f docker-compose.yml -f docker-compose.pilot.yml -f docker-compose.pilot.proxy-demo.yml config passed
  • docker compose -f docker-compose.yml -f docker-compose.pilot.yml -f docker-compose.pilot.header-demo.yml config passed
  • make pilot-proxy-demo-up required host-level Docker access and completed successfully
  • Unsandboxed runs of make pilot-proxy-demo-smoke, make pilot-proxy-demo-fhir-smoke, and make pilot-proxy-demo-hl7-smoke all passed end to end on 2026-03-20
  • make pilot-proxy-demo-down completed successfully after the live verification run
  • A new failure-path smoke mode now verifies persisted site_scope_rejection audit detail plus zero run-specific visible cases without requiring a reviewer round-trip
  • Unsandboxed runs of make pilot-proxy-demo-site-rejection-smoke and make pilot-header-demo-site-rejection-smoke both passed end to end on 2026-03-20
  • A new malformed-report smoke mode now verifies persisted validation_error and parse_error audit detail plus zero run-specific visible cases without requiring a reviewer round-trip
  • Unsandboxed runs of make pilot-proxy-demo-parse-validation-smoke and make pilot-header-demo-parse-validation-smoke both passed end to end on 2026-03-20
  • A new adapter-failure smoke mode now verifies persisted FHIR unsupported_payload and HL7 parse_error audit detail plus zero run-specific visible cases without requiring a reviewer round-trip
  • Unsandboxed runs of make pilot-proxy-demo-adapter-failure-smoke and make pilot-header-demo-adapter-failure-smoke both passed end to end on 2026-03-20
  • A new audit-visibility smoke mode now verifies owner access plus alternate-actor denial for a persisted site_scope_rejection run across both import-run audit endpoints
  • Unsandboxed runs of make pilot-proxy-demo-audit-visibility-smoke and make pilot-header-demo-audit-visibility-smoke both passed end to end on 2026-03-20
  • A new shared-visibility smoke mode now verifies owner access plus alternate-actor allow behavior for a successful persisted /api/v1/imports/reports run across both import-run audit endpoints
  • Unsandboxed runs of make pilot-proxy-demo-shared-visibility-smoke and make pilot-header-demo-shared-visibility-smoke both passed end to end on 2026-03-21
  • A new structured shared-visibility smoke mode now verifies owner access plus alternate-actor allow behavior for successful FHIR and HL7 imports across both import-run audit endpoints
  • Unsandboxed runs of make pilot-proxy-demo-adapter-shared-visibility-smoke and make pilot-header-demo-adapter-shared-visibility-smoke both passed end to end on 2026-03-21
  • A new failed-run shared-visibility smoke mode now verifies owner access plus alternate-actor allow behavior for a persisted /api/v1/imports/reports validation_error run across both import-run audit endpoints
  • Unsandboxed runs of make pilot-proxy-demo-failed-shared-visibility-smoke and make pilot-header-demo-failed-shared-visibility-smoke both passed end to end on 2026-03-20 local time, producing persisted run timestamps 2026-03-21T04:23:26Z and 2026-03-21T04:24:10Z
  • A new structured failed-run shared-visibility smoke mode now verifies owner access plus alternate-actor allow behavior for persisted FHIR unsupported_payload and HL7 parse_error runs across both import-run audit endpoints
  • Unsandboxed runs of make pilot-proxy-demo-adapter-failed-shared-visibility-smoke and make pilot-header-demo-adapter-failed-shared-visibility-smoke both passed end to end on 2026-03-20 local time, producing persisted run timestamps 2026-03-21T04:34:36Z, 2026-03-21T04:34:36Z, 2026-03-21T04:35:13Z, and 2026-03-21T04:35:13Z
  • A new structured site-rejection smoke mode now verifies persisted FHIR and HL7 site_scope_rejection audit detail plus zero run-specific visible cases without requiring a reviewer round-trip
  • Unsandboxed runs of make pilot-proxy-demo-adapter-site-rejection-smoke and make pilot-header-demo-adapter-site-rejection-smoke both passed end to end on 2026-03-20 local time, producing persisted run IDs 31, 32, 33, and 34 with UTC timestamps 2026-03-21T04:49:33Z, 2026-03-21T04:49:33Z, 2026-03-21T04:50:48Z, and 2026-03-21T04:50:48Z
  • A new structured audit-visibility smoke mode now verifies owner access plus alternate-actor denial for persisted FHIR and HL7 site_scope_rejection runs across both import-run audit endpoints
  • Unsandboxed runs of make pilot-proxy-demo-adapter-audit-visibility-smoke and make pilot-header-demo-adapter-audit-visibility-smoke both passed end to end on 2026-03-20 local time, producing persisted run IDs 36, 37, 38, and 39 with UTC timestamps 2026-03-21T05:19:26Z, 2026-03-21T05:19:26Z, 2026-03-21T05:21:57Z, and 2026-03-21T05:21:57Z
  • A new hosted pilot smoke workflow now reuses the existing base proxy and header Make targets on manual dispatch and a weekly schedule
  • Unsandboxed reruns of make pilot-proxy-demo-smoke and make pilot-header-demo-smoke both passed end to end on 2026-03-20 local time, producing persisted run IDs 40 and 41 with UTC timestamps 2026-03-21T05:46:27Z and 2026-03-21T05:47:33Z
  • The checked-in hosted workflow under .github/workflows/pilot-smoke.yml now also reuses the report-path site-rejection targets, installs API dependencies, and runs those same overlay targets in GitHub Actions with recorded green hosted FHIR and HL7 success-path evidence on 2026-03-25

Hosted Phase 6 Evidence

  • Hosted FHIR confirmation: #23563902873 succeeded on 2026-03-25 with dispatch scope fhir-success-only, total job duration 53 seconds, max job duration 27 seconds, and summary artifacts pilot-smoke-summary-header-demo-fhir-smoke plus pilot-smoke-summary-proxy-demo-fhir-smoke
  • Hosted HL7 trial: #23564057337 succeeded on 2026-03-25 with dispatch scope hl7-success-only, total job duration 52 seconds, max job duration 26 seconds, and summary artifacts pilot-smoke-summary-header-demo-hl7-smoke plus pilot-smoke-summary-proxy-demo-hl7-smoke
  • Hosted evidence bundle output is now produced by make pilot-smoke-evidence SUMMARY_DIR=/path/to/downloaded/pilot-smoke-artifacts HL7_DECISION=keep-manual, and the current copyable outcome is: hosted FHIR confirmed, hosted HL7 trial recorded, HL7 hosting mode manual_only, phase complete True
  • The compose-side root cause behind the earlier hosted failures was the bind-mounted web service losing image-installed node_modules; the fix is the anonymous /app/node_modules volume now checked into docker-compose.yml
  • Unsandboxed reruns of make pilot-proxy-demo-site-rejection-smoke and make pilot-header-demo-site-rejection-smoke both passed end to end on 2026-03-22 local time, producing persisted run IDs 42 and 43 with UTC timestamps 2026-03-23T05:05:47Z and 2026-03-23T05:06:23Z

What Is Implemented By Phase

Phase 0-3

  • API and web boot cleanly
  • Demo data import works
  • Deterministic triage persists Case, Report, Finding, and ReviewAction
  • Reviewer queue and case detail flow are live
  • Evaluation, exports, benchmark scripts, and reproducible validation are in place

Phase 4

  • Explainable PDAC trial matching is implemented
  • Case detail shows structured abstractions and eligibility traces

Phase 5

  • Hybrid explainable scoring is implemented
  • Worklist supports hybrid review priority, active-learning priority, disagreement queueing, threshold evaluation, and reviewer feedback capture

Phase 6

  • Structured request logging and readiness probes are implemented
  • Trusted proxy auth is implemented with provider presets (generic, authentik, keycloak, oauth2-proxy)
  • Role and site-scope enforcement are implemented in the API and reflected in the UI
  • Pilot Docker overlays and smoke scripts are in place
  • Research-safe de-identification views and redacted exports are implemented
  • FHIR DiagnosticReport and HL7 v2 ORU import adapters are implemented
  • Imported reports now preserve explicit provenance fields: patient identifier, encounter identifier, accession number, ordering provider, source system, source format, and import source identifier
  • The metadata is persisted on reports, exposed in case detail and exports, and pseudonymized in research-safe views
  • The generic CSV / JSON import path also accepts the same metadata surface
  • Import endpoints now persist import-run summaries with actor, timestamps, processed / created / updated / failed counts, and stable failure buckets
  • Analyst-facing audit routes are implemented at /api/v1/imports/runs and /api/v1/imports/runs/{run_id}
  • Audit visibility respects auth plus site scope, while still allowing actors to review their own failed runs
  • Selected FHIR and HL7 metadata precedence rules are now configurable through deployment settings rather than code edits
  • The Next.js app now exposes /imports for CSV / JSON / JSONL uploads, FHIR DiagnosticReport submission, HL7 ORU submission, recent run summaries, and per-run failure details
  • Failed import responses now include X-Import-Run-ID when an audit run was recorded, allowing the web workspace to load stable failure buckets immediately
  • Checked-in env bundles now exist for local mock development, the proxy demo overlay, and a new header-auth demo overlay
  • A new docker-compose.pilot.header-demo.yml overlay now demonstrates site-scoped field-level header auth with a fixed navigator identity
  • The smoke helper and Make targets now support both proxy and header auth demos, and both pilot smoke paths verify that /imports renders
  • The header-demo smoke path now imports demo data through /api/v1/imports/reports and verifies the persisted run through /api/v1/imports/runs/{run_id}
  • The header-demo pilot path now also has live FHIR and HL7 smoke targets that verify structured adapter imports plus persisted audit detail end to end
  • The header-demo pilot path now also has a live site-scope rejection smoke target that verifies a persisted failure bucket and zero run-specific visible cases
  • The header-demo pilot path now also has a live malformed-report smoke target that verifies persisted validation_error plus parse_error buckets and zero run-specific visible cases
  • The header-demo pilot path now also has a live adapter-failure smoke target that verifies persisted FHIR unsupported_payload plus HL7 parse_error buckets and zero run-specific visible cases
  • The header-demo pilot path now also has a live shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for a successful /api/v1/imports/reports run
  • The header-demo pilot path now also has a live failed-run shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for a persisted /api/v1/imports/reports validation_error run
  • The header-demo pilot path now also has a live structured failed-run shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for persisted FHIR unsupported_payload and HL7 parse_error runs
  • The header-demo pilot path now also has a live structured shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for successful FHIR and HL7 imports
  • The header-demo pilot path now also has a live audit-visibility smoke target that verifies owner visibility plus alternate-actor denial for a persisted site_scope_rejection run
  • The proxy-demo overlay now uses a fixed site-scoped navigator identity too, and its smoke path imports demo data through /api/v1/imports/reports with persisted run verification
  • The proxy-demo pilot path now also has live FHIR and HL7 smoke targets that verify structured adapter imports plus persisted audit detail end to end
  • The proxy-demo pilot path now also has a live site-scope rejection smoke target that verifies a persisted failure bucket and zero run-specific visible cases
  • The proxy-demo pilot path now also has a live malformed-report smoke target that verifies persisted validation_error plus parse_error buckets and zero run-specific visible cases
  • The proxy-demo pilot path now also has a live adapter-failure smoke target that verifies persisted FHIR unsupported_payload plus HL7 parse_error buckets and zero run-specific visible cases
  • The proxy-demo pilot path now also has a live shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for a successful /api/v1/imports/reports run
  • The proxy-demo pilot path now also has a live failed-run shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for a persisted /api/v1/imports/reports validation_error run
  • The proxy-demo pilot path now also has a live structured failed-run shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for persisted FHIR unsupported_payload and HL7 parse_error runs
  • The proxy-demo pilot path now also has a live structured shared-visibility smoke target that verifies owner visibility plus alternate-actor allow behavior for successful FHIR and HL7 imports
  • The proxy-demo pilot path now also has a live audit-visibility smoke target that verifies owner visibility plus alternate-actor denial for a persisted site_scope_rejection run
  • GitHub Actions now also has a hosted pilot smoke workflow for the base proxy and header success-path overlays plus report-path and structured adapter site rejection in both auth modes

Research Intelligence foundation

  • Shared persisted records now exist for research sources, runs, run items, documents, evidence, topics, digests, and opportunities
  • Seed catalogs now live under data/research/ for sources, topics, a lightweight pancreatic oncology graph, and seeded documents
  • Manual ingest, due-only ingest, schedule, and digest scripts now exist through make research-intel-ingest, make research-intel-ingest-due, make research-intel-schedule, make research-intel-digest, and make research-intel-refresh
  • The Next.js app now exposes /research-intel, /research-intel/schedule, /research-intel/documents, /research-intel/digests, and /research-intel/opportunities
  • Case detail now links to a generated research brief that maps rationale or trial context to current research-intel topics and cited documents
  • Promotion stays human-gated, triage scores remain untouched, and payment or crypto execution is still intentionally out of scope

Important Guardrails

  • Preserve deterministic explainability. Do not replace the rule engine with opaque behavior.
  • Keep all new ingestion paths mapped into the existing ReportInput plus triage_report() flow unless there is a strong reason to introduce a new persistence boundary.
  • Keep auth, site scoping, and de-identification behavior intact for any new API surfaces.
  • Do not regress reviewer usability in favor of backend purity. The worklist and case detail pages are central product surfaces.
  • Do not claim clinical validation. This remains a research-first pilot stack.

Most Important Files

Core triage and evaluation

  • apps/api/app/services/triage_engine.py
  • apps/api/app/services/hybrid_analysis.py
  • apps/api/app/services/evaluation.py
  • apps/api/app/store/memory_store.py
  • apps/api/app/models/entities.py
  • scripts/run_demo_eval.py
  • scripts/validate_repo.py

Benchmark proof and public comparison surfaces

  • apps/web/app/page.tsx
  • apps/web/app/about/page.tsx
  • apps/web/app/proof/page.tsx
  • apps/web/app/marketing.module.css
  • apps/web/lib/demo-proof.ts
  • apps/api/app/schemas/benchmark_submission.py
  • apps/api/app/schemas/evaluation.py
  • apps/api/tests/test_external_evaluation.py
  • scripts/run_external_eval.py
  • scripts/write_demo_benchmark.py
  • scripts/validate_benchmark_submission.py
  • docs/QUICKSTART.md
  • docs/EVALUATION.md
  • docs/BENCHMARK_SUBMISSIONS.md
  • docs/LABELING_GUIDE.md
  • docs/examples/demo-benchmark-current.json
  • docs/examples/demo-benchmark-current.md
  • docs/examples/benchmark-submission-template.json
  • docs/examples/benchmark-label-template.jsonl
  • docs/examples/benchmark-prediction-template.jsonl

Reviewer workflow and exports

  • apps/api/app/api/routes/cases.py
  • apps/api/app/api/routes/exports.py
  • apps/api/app/schemas/case.py
  • apps/api/app/schemas/feedback.py
  • apps/web/app/cases/page.tsx
  • apps/web/app/cases/[caseId]/page.tsx
  • apps/web/app/imports/page.tsx
  • apps/web/app/imports/actions.ts
  • apps/web/lib/api.ts

Deployment, auth, and pilot path

  • README.md
  • CHANGELOG.md
  • CONTRIBUTING.md
  • SECURITY.md
  • CODE_OF_CONDUCT.md
  • docs/RELEASE_READINESS.md
  • apps/api/app/main.py
  • apps/api/app/api/routes/health.py
  • apps/api/app/auth/dependencies.py
  • apps/api/app/core/config.py
  • apps/api/app/observability/logging.py
  • docker-compose.yml
  • docker-compose.pilot.yml
  • docker-compose.pilot.header-demo.yml
  • docker-compose.pilot.proxy-demo.yml
  • deploy/examples/README.md
  • scripts/smoke_proxy_auth.py
  • Makefile

De-identification and research surfaces

  • apps/api/app/services/deidentification.py
  • apps/api/app/schemas/research.py
  • apps/api/app/api/routes/cases.py
  • apps/web/app/cases/[caseId]/research/page.tsx

Integration adapters

  • apps/api/app/api/routes/imports.py
  • apps/api/app/services/imports.py
  • apps/api/app/services/fhir_imports.py
  • apps/api/app/services/hl7_imports.py
  • apps/api/app/db/session.py
  • apps/api/app/models/entities.py
  • apps/api/app/schemas/imports.py
  • apps/api/app/schemas/triage.py
  • apps/api/tests/test_imports.py
  • apps/api/tests/test_import_runs.py

Recommended Next Slice

Broaden from the checked-in retrospective-style multi-cohort sample to a larger de-identified or collaborator-supplied benchmark drop using the same reviewer-facing proof shape.

Why this is next

  • Phase 6 exit criteria are now met: the hosted/manual smoke boundary is verified, the hosted evidence is recorded, and the release-facing docs reflect the real product state.
  • The proof surface now carries benchmark buckets, queue previews, reviewer cues, expected rationale codes, and a 10-report casebook across 5 recurring categories, while the public external benchmark helper now emits the same shape for collaborator datasets and the repo now includes both a 12-case retrospective-style multi-cohort sample pack and a 9-case wording-variance challenge pack.
  • More parser or smoke-baseline work would now deliver less value than bringing a broader real-world benchmark drop into that established structure.

Target outcome

Add one post-Phase-6 proof slice that:

  • introduces a broader de-identified or collaborator-supplied benchmark input set using the same published casebook fields
  • keeps the existing reviewer-facing proof shape while improving data realism, provenance, or collaborator handoff quality
  • refreshes /proof, release notes inputs, and benchmark-facing docs from that stronger evidence base
  • leaves the now-recorded hosted smoke baseline intact rather than reopening it unnecessarily

Suggested implementation shape

  1. Start from the existing benchmark proof surfaces under docs/examples/, /proof, and the external evaluation tooling rather than adding new infrastructure.

  2. Use the current expanded casebook fields, optional report excerpts, optional cohort labels, and checked-in retrospective sample as the compatibility target for any broader de-identified retrospective sample or externally supplied benchmark set.

  3. Refresh the published proof artifacts and make sure the updated benchmark narrative still matches the current release-facing docs and the now-explicit hosted/manual smoke boundary.

Acceptance criteria

  • The next slice improves public proof quality rather than just re-stating existing smoke evidence.
  • Any refreshed benchmark or reviewer-facing artifact is reproducible from checked-in commands.
  • make validate-strict stays green.
  • The hosted Phase 6 evidence recorded above remains accurate and unchanged unless a newer deliberate rerun supersedes it.

Good First Commands For The Next Agent

make validate-strict
sed -n '1,260p' docs/PHASES.md
sed -n '1,260p' docs/API_SPEC.md
sed -n '1,260p' docs/DEPLOYMENT.md
sed -n '1,260p' docs/OPEN_SOURCE_STRATEGY.md
sed -n '1,260p' README.md
sed -n '1,260p' docs/RELEASE_READINESS.md
sed -n '1,260p' apps/api/app/services/hl7_imports.py
sed -n '1,260p' apps/api/tests/test_imports.py
sed -n '1,260p' apps/api/tests/test_import_runs.py
sed -n '1,260p' scripts/smoke_proxy_auth.py
sed -n '1,220p' .github/workflows/pilot-smoke.yml
make benchmark-external LABELS=docs/examples/benchmark-label-template.jsonl PREDICTIONS=docs/examples/benchmark-prediction-template.jsonl
make benchmark-external-sample
make refresh-external-sample-proof
make validate-benchmark-submission SUBMISSION=docs/examples/benchmark-submission-template.json

Handoff Summary

This is a clean post-Phase-6 checkpoint. The repo is runnable, validated, and already beyond MVP scaffolding. Import metadata preservation, import-run audit trails, config overrides, the web import workspace, concrete proxy plus header-auth pilot packaging, automatic PR and main validation, checked-in hosted base plus attachment-backed FHIR success-path plus report-path and structured adapter site-rejection smoke automation, a sharper public landing experience, a checked-in benchmark proof surface, a machine-validated public benchmark submission pack, a reproducible external evaluation bundle writer, FHIR inline Reference.identifier fallback coverage, FHIR presentedForm attachment-backed narrative decoding, and HL7 ED, repeated-OBX-5, custom-MSH-2 delimiter, escape-sequence, plus subcomponent-aware metadata support are complete. The benchmark proof surface now also publishes dataset coverage, top-k queue previews, benchmark buckets, expected rationale cues, reviewer-facing casebook notes from an expanded 10-report demo snapshot that covers 5 recurring benchmark buckets, a side-by-side comparison of the published external packs, and anchored links into each checked-in external casebook; the repo also includes both a checked-in 12-case retrospective-style multi-cohort sample pack and a 9-case wording-variance challenge pack, and make validate-strict verifies the published external benchmark registry that feeds /proof. The next agent should focus on bringing in a broader de-identified or collaborator-supplied benchmark drop, or deeper bucket-level cross-pack analysis, rather than reopening hosted smoke capture or parser work that is already covered in tests.