Full client-mapped plan_type pipeline: complete header alias set (bare
'plan'/'plan_name' stay payer aliases), value canonicalization map (Medicare
Part B / FFS / MA / Part C / private / employer etc), unrecognized values
grade Plan Type Needed with a hashed once-per-value warning. Once-per-batch
deprecation fence when records lack plan_type (timing guesses from payer,
legacy; readiness never does). CSVImport mapping review gains Plan Type.
All 66 demo/drift fixture CSVs authored with plan_type columns (variant
headers + variant values on the drift corpus); corpus sweep: 66/66 map, 100%
typed. Echo paths canonicalize through the same value map (audit Low fixed).
_normalize_payer and the default config entry remain untouched (Phase 2c,
Kisa-gated). Independent audit: SHIP. 169 tests green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ends the override leak: an SWO/PA save now applies to one device's line, not
every device the patient has. doc_status gains device_type ('' = legacy
patient-scoped row, applies until re-saved); the old unique key is dropped by
introspected name and replaced with the 4-col scope key. Staff overrides now
FEED THE VERDICT (device-scoped, staff wins over the CSV cell) instead of
display-only; visit overrides keep their patient-scoped fan-out via
confirmed_visits. PUT /api/doc-status and /api/confirm-visit accept an
optional echo of the patient's lines and return {lines, rollup_status}
recomputed through the same engine (stateless, no batch rebuild). Frontend
minimal fix: doc-status state and writes keyed patient:device. Independent
audit: SHIP (migration reproduced empirically on both old-key shapes; guard
scoped per its nit). 162 tests green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
core/rollup.py implements the locked truth table: worst gradeable line wins;
a Plan Type Needed line caps the patient at Action Needed; non-gradeable
lines are displayed but never touch the light; no-graded-lines patients carry
None (no label invented). /api/upload gains additive patients[] (one entry
per patient, nested device lines with doc checklists, citations, timing
flags, dedup display fields) and patient_stats (counts patients, not rows).
Flat records payload untouched until P6. Independent audit: SHIP; its LOW
findings fixed (date coercion on timing keys, empty-rollup guard, severity
divergence documented). 151 tests green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
core/rules.py format-tolerant readers + stable registry (R001-R004 plan-scoped,
R100-R102 universal per LCD L33822). payer_rules.json wrapped in the same
commit as both engine readers (coupling hazard closed). Every DocItem now
cites the rule that decided it; NOT_EVALUATED cites nothing. Independent
audit: SHIP (registry-drift test added per its L2 finding). 138 tests green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Readiness verdicts now ride additively on /api/upload and /api/confirm-visit:
- core/worklist_readiness.py: ReadinessIndex — row-to-line membership resolved
through the dedup group key (patient_id, device_type, DOS), never re-derived
from a row's own plan_type (kills a reviewer-reproduced false green)
- RecordOut gains plan_type / readiness_status / readiness_items (additive,
default None); UploadResponse gains readiness_stats (reconciles with rows)
- Minimal plan_type CSV mapping (client-supplied, lowercased+trimmed, never
guessed) per the umbrella reactivation plan; full enforcement lands in P5
- Fixes pre-existing record_lookup miss for order-numbered CSVs; fallback
fields now come from the dedup MergedShipment (latest-non-null, collected
order numbers) so doc_state and the verdict grade from the same values
- confirm-visit normalizes plan_type identically to the CSV path
Verification: 132 tests green (108 baseline + 24 wiring/regression);
adversarial 3-lens review + refutation pass (7 findings confirmed, all fixed,
regression-locked); independent code-auditor verdict SHIP; Pi E2E 33/33.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Grade one coverage line's five documentation items on two separated axes the
old code conflated: required_state (from plan_type + payer_rules.json) and
input_state (SUPPLIED | ABSENT). A line is "Clear to Ship" only when EVERY
required item is satisfied by positive evidence — a required item with no data
is NOT_EVALUATED and blocks green, never granted by absence of contradiction.
- plan_type never guessed: None / unmapped / casing-variant -> "Plan Type Needed",
structurally cannot be green; PECOS/PA sit NOT_EVALUATED while universal items
(SWO/visit/diagnosis per LCD L33822) still grade.
- Required-ness reads from payer_rules.json (single source of truth), removing
the latent Python-set-vs-JSON divergence.
- Config-surface fail-safe: a known plan missing/partial in config cannot grade
green (NOT_EVALUATED), and the green gate checks every item.
- Diagnosis accepts a real ICD-10 code as on-file; only explicit negatives gap.
Additive only: not wired into the live pipeline. 27 tests green (full suite 73),
code-auditor reviewed (H1/H2/M1/M2/M3/M4/L1 fixed).
PROVISIONAL, flagged for Pi/Kisa: gap->label mapping (At Risk/Action Needed/
On Track) and the visit-recency scope boundary (timing layer, not readiness).
Spec: docs/readiness-model-brief-2026-06-23.md section 4
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Group normalized shipment rows into deduped shipments and gradeable
coverage lines (patient -> device -> shipment), the structural fix for the
row-keyed worklist bug (six months of orders no longer = six worklist lines).
- Locked dedup key (patient_id, device_type, date_of_service); order_number
is display-only. Monthly resupplies stay separate, not duplicates.
- Client-mapped plan_type carried through, never guessed from payer name;
conflicting plan types in a group route to "plan type needed".
- Quantity merge: sum across distinct component configs, max within a config
(duplicate guard) with a warning breadcrumb on disagreement.
- CGM gradeable this phase; pumps preserved/displayed but not graded.
Additive only: not yet wired into the live pipeline, grading unchanged.
30 tests green (full suite 46), code-auditor reviewed (M1/M2/L1/L2/L3 fixed).
Design: docs/dedup-design-2026-06-24.md, docs/readiness-model-brief-2026-06-23.md
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- supabase_client.py: lazy singleton client (no-ops when env vars absent)
- persistence.py: persist_upload writes batch, source_files, normalized_records,
mapping_decisions, report_runs; persist_export records export_files
- schema.sql: 11-table schema with RLS + WORM rules for audit/raw tables
- main.py: wire persist_upload/persist_export; add ExportRequest body model
so export accepts {records, batch_id}; batch_id returned on upload response
- api.js: add exportFromBackend helper passing batch_id through
- requirements.txt: add supabase>=2.0.0
- smoke_test.py: update export call to new body format
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- lock CORS to Vercel domain via ALLOWED_ORIGINS env var (removes allow_origins=*)
- add X-API-Key header auth on /api/upload and /api/export
- normalizer: add mapping confidence (high/inferred), new aliases for Acct #,
Member ID, External Patient Ref, DME Description, dispensedate; 63/63 CSV files pass
- coverage_calculator: add RULE_VERSION = "v0.1", rule_version on every CoverageResult
- main.py: audit logging wired on upload + export, rule_version + mapping_summary in response
- generate_samples.py: 25 CSV files now use 25 different real-world header formats
- add generate_10k.py for 10,000-patient synthetic dataset
- add tests/smoke_test.py (passes against local backend)
- add docs/pilot-guide-v1.md for Robert Robinson pilot onboarding
- add docs/daniel-pilot-readiness-whitepaper.md and .pdf
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
