RAPTOR

RAPTOR is a vertical TSC1/TSC2 research-evidence program for building reproducible variant evidence packets, validation controls, and governed ClinVar/VCEP preparation. It is not a clinical diagnostic system and does not issue authoritative variant classifications.

Start with the engineering method: METHOD.md summarizes the model-separated plan/test/build/review loop, Definition-of-Done gates, evidence and authorization ladder, and failure modes that changed RAPTOR's implementation and conclusions.

Current status

The complete TSC evidence census, the leakage-safe masked held-out rerun (R2, ADR-0012), and the tiered v3 post-hoc re-adjudication (ADR-0013) have run. R2 and its v1/v2 interpretation are frozen and unchanged; v3 re-reports the same frozen R2 counts on independent axes and generates no new evidence.

Public progress posts: Before the First Score · After the First Rerun

Milestone	Result
TSC VUS census (current, disabled_manual)	6,618 variants; 157 candidate-LP review, 7 candidate-LB review, 6,424 unresolved, 30 annotation/manual — internal, non-authoritative review directions only
Frozen benchmark	3,681 knowns; 2,577 held out
Mask integrity (R2)	2,577/2,577 identities removed; zero survivors
R2 masked BIAS score (ADR-0012, disabled_manual)	2,577 canonical rows; zero PP3/BP4 scored calls
v1/v2 frozen interpretation	FAIL (coarse missense gate) / BLOCKED_POLICY; vus_authorized=false (immutable)
v3 tiered re-adjudication (ADR-0013, post-hoc)	missense pathogenic NO_CALLS/NOT_ESTIMABLE; missense benign UNDERPOWERED/NOT_ESTIMABLE; truncating pathogenic ADEQUATE+MET/SUPPORTED_POSTHOC; full spectrum NOT_VALIDATED/NOT_AUTHORIZED
Prospective validation	BLOCKED_DATA — the exact preregistered August archive URL returned 404; no archive bytes, labels, or scores were accessed, and the live alternate URL was not substituted
VUS / research-scope authorization	No — canonical validated research-scope flag remains false
Mechanism Atlas source catalog	Five resolver-verified grounding sources plus seven mechanically non-grounding leads; public identifiers, licences and hashes are committed
$TSC2$ $p.\mathrm{Arg611Gln}$ deterministic pass	Two exact spans passed Gates 1–7; Gate 8 blocked for missing named review; zero accepted claims
Six-variant gate-smoke cohort	6/6 technical packages passed Gates 1–7 and blocked only at Gate 8; zero accepted claims; explicitly not the formal contrast panel
Formal Atlas contrast panel	INFEASIBLE_PANEL — the reviewed selector exhaustively completed 24 attempts across L0–R7 against universe/map v4 with no budget exhaustion; no panel or candidate was selected

v3 is a post-hoc semantic correction of the frozen R2 aggregate: it separates run integrity, data sufficiency, conditional performance, policy parity and authorization instead of collapsing them into one coarse pass/fail, but it performs no new run, scoring, or evidence generation and authorizes nothing.

Source of record: data/census/tsc_masked_holdout_gate_disabled_manual_2026-07-21.json (R2, frozen) and data/census/tsc_tiered_readjudication_2026-07-21.json (v3, post-hoc). The earlier data/census/tsc_masked_holdout_gate_2026-07-13.json run remains a superseded historical comparator, not the current approved policy.

The real-source Atlas runs establish deterministic source identity, content hashes, exact spans, ontology/context structure, and classification-leakage controls. They do not establish scientific sufficiency: no claim is promoted until named human Gate 8 review. Public run manifests are under data/atlas/runs/2026-08-03/.

What is built

• deterministic TSC1/TSC2 ClinVar ingestion and canonical GRCh38 SPDI identity;
• arm's-length Nirvana/BIAS evidence ingestion;
• criterion-lineage and ClinVar-circularity controls;
• upstream held-out masking and conservation ledgers;
• exact-set canonical BIAS adapter;
• arm's-length BP4/PP3 aggregation correction;
• exact per-direction Clopper-Pearson gate, plus a tiered v3 post-hoc re-adjudication (ADR-0013) that reports run integrity, data sufficiency, conditional performance, policy parity and authorization as independent axes over the same frozen R2 aggregate;
• a packet-free census aggregation package (raptor.census) producing the current, non-authoritative candidate-direction census under the PP3/BP4-disabled/manual policy (ADR-0012);
• immutable, first-pass-blinded candidate evidence packets;
• a 30-pattern internal calibration batch;
• a condition-agnostic Mechanism Atlas core, versioned $TSC2$ disease pack, deterministic offline citation/exact-span resolver, hash-frozen post-discovery selection protocol, and candidate-free universe lock;
• public metadata-only source and gate-run manifests for the first real $TSC2$ deterministic-verification runs;
• ClinVar/VCEP schema and lifecycle preparation with submission disabled.

Boundaries

• The current census's 157 candidate-LP-review and 7 candidate-LB-review directions are internal, eval-only triage signals—not reclassifications. The historical 238/1,333 counts were computed under the PP3/BP4-active policy later superseded by ADR-0012 and remain a labeled historical comparator only.
• The production criterion-strength/candidate-direction policy remains unapproved.
• PM1 was excluded from the R2 fixed evaluation after a zero-support audit and remains unvalidated for production; this exclusion is the reason v3 reports missense pathogenic as NO_CALLS, not as a passed or failed metric.
• v3's truncating_pathogenic scope evidence is SUPPORTED_POSTHOC only; its authorization is PENDING_PROSPECTIVE and no scope is authorized until the prospective, unseen-data validation locked by ADR-0013 completes.
• No external VUS worklist or ClinVar submission is authorized before a passed prospective validation and variant-level expert sign-off.
• Atlas Gates 1–7 verify provenance and deterministic fidelity, not biological truth. The current eight spans across seven variants remain unaccepted at Gate 8; the six-variant cohort is engineering repeatability only.
• BIAS remains an external, separate-process dependency under the arm's-length boundary documented in ADR-0007.

Repository guide

• METHOD.md — engineering method, gates, evidence ladder, and failures caught
• docs/PROGRAM.md — live program status and priorities
• docs/STRATEGY.md — strategy + operating-model authority
• docs/EVALUATION.md — evaluation protocol + acceptance-rubric authority
• docs/DECISIONS.md — architecture and policy decisions
• docs/ARCHITECTURE.md — system architecture
• docs/RISK_REGISTER.md — failure modes and controls
• docs/prd/ — feature contracts
• data/census/ — committed non-identifying aggregate records
• data/atlas/runs/2026-08-03/ — public source licences/hashes, span locators, gate outcomes, and reconstruction boundary

Large ClinVar, Nirvana, BIAS, packet, and held-out artifacts remain outside the repository. Their checksums and operator contracts are recorded under docs/ops/.

Development

The package requires Python 3.11+:

python -m pip install -e ".[dev]"
pytest

On ARM64 Windows, reference-backed tests run in the documented WSL environment; Nirvana/BIAS execution remains on the isolated x64 worker.