jang1563/agentic-drug-discovery-system

Agentic Drug Discovery System

This card describes the unuploaded 0.3.0.dev3 update candidate for the public 0.3.0.dev2 Agentic Drug Discovery System mirror. It contains the executable control plane, tests, documentation, schemas, aggregate evidence, release metadata, safety boundaries, and the ctdbench scorer. It is not a row dataset or model release and does not contain raw source bundles, real provider review jobs, ingestion runs, raw clinical/regulatory source snapshots, hidden labels, real curator manifests, real clinical decision, cohort, outcome-evaluation, uncertainty, design/stress/sensitivity-scenario, or closed-loop policies/manifests/submissions/catalogs/batches/receipts/packages/unit results, real scenario elicitation or hidden-dependence working records, unit-to-cluster assignments, replicate- or cluster-level results, locked episodes, generated trajectories, scheduler logs, local paths, credentials, or unpublished working notes. Upload requires exact-package review and explicit approval.

The candidate also contains an independent Biohub-context presentation-readiness packet. It positions the project as a downstream translational evidence-governance bridge, binds official public alignment sources and local artifact hashes, and freezes a proposed 90-day pilot. It does not claim Biohub affiliation, endorsement, data integration, a virtual-cell model, a completed pilot, or external clinical validation.

The package also includes a generic synthetic M6 translational handoff. Its strict reader preserves perturbation, cell/tissue/model, assay/endpoint, uncertainty, sampling, QC, source lineage, and review context while compiling only contextualizes evidence. It includes no Biohub source or real experimental payload.

An explicitly synthetic ulcerative-colitis conformance slice broadens the software path to clinical-remission odds ratios with higher_is_better semantics from M6 context through non-pooled clinical synthesis. The disease identity is real; all targets, interventions, trials, measurements, sources, and reviews are synthetic.

A separate provider-only validation executes two public ClinicalTrials.gov ulcerative-colitis induction records. It preserves one uncertain interval on HOLD and advances one bounded favorable interval while retaining exact treatment-phase and safety-group identity. The package includes only the payload-free hashes and selected aggregate values, not source bytes or reviewer jobs. This does not convert the synthetic upstream/downstream UC slice into real end-to-end validation.

A second provider-only run binds induction and maintenance endpoint/safety populations within NCT02435992. It preserves separate analysis counts, refuses participant-identity inference, and counts the shared NCT/source as one non-pooled trial.

An independent NCT01458574 run adds a primary maintenance endpoint reported as a 23.2 percentage-point risk difference. It binds percent scale, candidate/comparator sign, endpoint and serious-safety roles, and source hashes; it is not same-candidate replication or pooled evidence.

A rheumatoid-arthritis extension executes public NCT00383188 ACR20 and serious-adverse-event aggregates through the same additive contract. Its uncertain 8.05 percentage-point interval remains on HOLD, demonstrating scale-specific decision precision without claiming efficacy, safety acceptability, or a real multi-trial additive tensor.

A follow-on olokizumab execution binds source-disjoint NCT02760407 and NCT02760433 records to one Week-12 ACR20 family. Source risk differences remain proportions, decision precision is normalized to percentage points, and the two distinct inadequate-response populations are not pooled or treated as exchangeable. A higher observed serious-event risk in one trial keeps the bounded plan on HOLD; no comparative safety or treatment claim is made.

The next diagnostic hash-binds the same records to reviewed methotrexate- and TNF-inhibitor- inadequate-response strata. Descriptive stratification completes, but transport estimation remains blocked because the release has no declared target population, within-stratum replication, individual-level covariates, preregistered transport model, or risk-of-bias assessment.

An independent follow-on replaces the TNFi-IR member with MTX-IR phase 3 trial NCT02760368. The exact non-pooled report now has two trials in one reviewed MTX-IR stratum and removes the distinct-strata and no-within-stratum-replication blockers. It also exercises explicit month-precision registry chronology with conservative period-end normalization. Five transport blockers and the workflow HOLD remain; no pooled, transported, safety, or treatment conclusion is made.

The next follow-on binds each Week-12 ACR20 outcome to exact registry fields and registry-labeled protocol/SAP PDF hashes, pages, sections, and excerpts. Both project-internal trial assessments are some_concerns: source-specific arm/endpoint identities and ITT analysis denominators replay, but complete observed outcomes and a standalone final pre-unblinding SAP are not established. PDF page counts, title-page dates, and cited text are executable checks. This resolves only risk_of_bias_not_assessed; four transport blockers remain.

At a Glance

• Surface: Hugging Face Dataset repository.
• Public baseline: 0.3.0.dev2 exact-source mirror, published after explicit approval.
• Candidate state: 0.3.0.dev3, not uploaded and pending exact-package approval.
• Release lineage: 0.2.0 remains the latest tagged stable release.
• Contents: Bounded planner, typed execution core, deterministic policy replanning and hash-bound checkpoint resume, cross-stage disease/target/assay/model-system/intervention/trial/design identity ledgers, atomic multi-trial portfolio extraction, reviewer-approved endpoint mapping, mapping-gated source-disjoint non-pooled benefit-risk synthesis, bounded source-preserving ClinicalTrials.gov harmonization, ten-dimension provenance-preserving clinical evidence tensor compilation and bounded VOI action planning, accepted-state-bindable cohort diagnostics with matched policy sensitivity and provenance-overlap reporting, preregistered package-bound clinical outcome forecasts with aggregate calibration and paired policy evaluation, dependence-audited CR1 uncertainty for additive outcome metrics, deterministic aggregate prospective clustered-board design simulation, informative-evaluability and residual-dependence stress comparison over population/evaluable targets and nominal/dependence-closed clustering, prediction-stratified binary log-IMOR pattern-mixture sensitivity with matched calibration/recovery/identification diagnostics, fingerprint-bound nominal/dependence-closed cluster-jackknife sampling calibration around every fixed log-IMOR model functional, unequal-cluster influence calibration comparing normal, Student-t, delete-mj, and experimental multiplier intervals, bounded selected-action execution with compact receipts, reviewer-only refresh and exact source-rejoined transition validation, source capture and payload-free manifest compiler, semantic mappings, dependency-free pinned-evidence adapter and binding, stage and multi-stage program runners, matched and sealed evaluators, preregistered held-out curation contracts, stage-stratified uncertainty, synthetic evaluation tests, aggregate external evaluation evidence, manifests, audit code, and the ctdbench scorer.
• Excludes: Raw source data, real sealed or held-out boards, private ADDS-Frontier task/oracle, detailed preflight or oracle-fragility stage/probe reports, private support-curation packets, transition-audit records, coupled-augmentation matched records or sealed candidate/control bindings, private three-arm placebo or five-arm tokenizer-placebo packets/keys and per-placebo evaluation diagnostics, independent-family per-placebo reports and frozen WordPiece/SentencePiece model assets, semantic-review arm, sealed mapping, reviewer response, workflow or resolution ledger, triage, canonical unblinding/replay, resolution receipt, independent-oracle challenge packet/key/assignment/response/comparison/ledger, and adjudication records, curator identities/attestations/votes/adjudications, curation manifests, real clinical decision policies/action catalogs/evidence tensors/packages, real clinical cohort manifests/accepted-state bindings/package diagnostics/reports, real clinical prediction submissions/outcome or dependence manifests/unit labels/source assessments/unit-to-cluster assignments/cluster-level or per-unit scores, real design/stress/sensitivity scenarios, pilot or log-IMOR elicitation, prediction-stratum working records, latent outcomes, replicate records, correction-selection deliberations, real closed-loop policies/execution batches/provider requests or outcomes/receipts/reviewer refresh records/transitions, cached episode packets, label vaults, policy submissions, per-episode evaluations, hidden labels, generated trajectories, logs, credentials, local paths, or model weights.
• Source: Exact commit and tree are recorded in upload_manifest.json.

ADDS-Frontier Research Direction

The candidate includes a preregistered, machine-validated protocol for evaluating complete evidence-authorized drug-discovery trajectories. Full success requires all 13 scientific and operational gates plus zero unauthorized commits; fixed-slice accuracy is diagnostic only. The pilot freezes five task families and 40 baseline-blind canonical tasks before any model run. Its 0.5%-5% launch band, centered near 2%, is an observation target, not a task-selection rule.

See docs/50_adds_frontier_research_protocol.md, docs/adds_frontier_research_protocol.json, and rl_env/specs/frontier_pilot_seed_manifest.example.json. These are design artifacts, not a real benchmark board or performance claim.

The package also includes five public development fixtures, their separately hash-bound public oracles, and a curation-status manifest. They exercise 30 stages and 25 preregistered mutations, but have no independent curator roster, are not board-admitted, and count as zero of the 40 private pilot tasks.

The frozen private-board sampling frame contains 40 preregistered slots across five task families and eight disease domains. Ten calibration drafts in hematology and immune-inflammatory disease are now authored outside the mirror and bound by a payload-free progress manifest: 60 stages and 50 mutations, with packet integrity and oracle sealing complete. Independent review, expert solve, replay, contamination review, and every board admission remain pending; the other 30 slots remain unassigned. The mirror contains schemas, salted commitments, and aggregate status only, with no private task, oracle, identity, nonce, curator, or canary bytes and no benchmark result.

The payload-free automated preflight replays all 60 canonical stage labels through the scorer and executes 50 frozen contract mutations. The scorer round-trips pass 60/60, and all 30/30 machine-decidable source-rename, evidence-removal, and temporal-rebind probes match their contract outcomes. The 20 evidence-reveal and identity-rebind probes are structurally accepted but remain semantic-review-required. Human gates remain pending, board admission remains zero, and no model run or benchmark performance is claimed. See rl_env/specs/frontier_calibration_preflight_summary.json.

Those 20 unresolved probes are also compiled into private blinded two-arm review drafts. All 20 pass an exact minimal-delta certificate, while canonical-arm mappings and author expectations live in a separately committed private key set. The mirror exposes only rl_env/specs/frontier_semantic_review_readiness_summary.json: reviewer assignments, responses, consensus labels, adjudications, board admissions, and model runs all remain zero.

The response workflow requires three unique, independence-attested, conflict-free responses bound to both arms and every stage. Pair-level changed components must equal the deltas derived directly from all twelve arm-stage answers. Evidence-reveal pairs must remain invariant before the first accessible-evidence divergence, blocking future evidence from affecting an earlier stage. Exact agreement yields only a consensus candidate; abstention, action, witness/blocker, and changed-component disagreement route to human adjudication in frozen priority order. These checks establish internal causal consistency, not scientific correctness. The public workflow summary records all 20 packets as unassigned and contains no reviewer or response bytes.

The resolution contract permits sealed canonical-arm opening only after recomputing an exact three-response consensus candidate. It then checks six canonical stages across disposition, next action, risk flags, witnesses, and blockers. Even 30/30 requires an independent, conflict-free human resolution receipt before admission review, and no path automatically changes a board gate. The payload-free summary reports zero unblindings, replays, receipts, and admissions.

The independent-oracle challenge protocol separately tests author-oracle agreement. Ten private canonical task packets omit the sealed task mapping, author oracle, mutation expectations, semantic-review material, and authoring sources. Two unique, conflict-free, independence-attested challengers must complete all six stages before any oracle opening. The state machine distinguishes abstention, challenger disagreement, author-oracle disagreement, and three-way convergence across 30 factorized components. Convergence remains a candidate only and cannot establish correctness or alter an expert, admission, or board gate. The public readiness summary reports ten unassigned packets and zero assignments, responses, convergence candidates, or adjudications.

A separate machine-only oracle fragility audit applies one valid mutation to each scorer component at every canonical stage. All 296/296 applicable probes localize exactly, while four next-action probes are structurally nonapplicable. The audit also exposes a stronger limitation: all 60/60 stages and 10/10 complete tasks are witness-saturated, and seven stages reuse a witness lineage. The payload-free summary therefore marks oracle support as not ready for independent challenge and requires human curation before challenge assignment. It does not claim scientific minimality or change an expert-solve, admission, or board gate. See rl_env/specs/frontier_oracle_fragility_summary.json.

The machine-only follow-on compiles the limitation into a private review workload without changing an oracle. Of 130 witness occurrences, 20 singletons are protected and 110 become leave-one-out review candidates. All 110/110 applicable counterfactuals fail only witness_valid; 14 repeated- lineage members receive lineage-review priority and 96 receive support-selectivity priority. This is scorer behavior, not removal evidence. The public summary keeps automatic edits, human decisions, challenge assignment, and gate changes at zero or false. See rl_env/specs/frontier_oracle_support_curation_summary.json.

The cross-stage follow-on decomposes 50 canonical transitions into 20 action-only, 20 access/ witness-only, ten stable, and zero coupled action/support changes. Witness deltas match evidence- access deltas 50/50, and blocker-change booleans match action-change booleans 50/50. This does not establish inconsistency or causality; it shows that the current canonical trajectories do not contain a transition that directly exercises evidence-responsive action change. The payload-free summary keeps causal establishment, oracle edits, challenge assignment, and gate changes false. See rl_env/specs/frontier_oracle_transition_audit_summary.json.

The matched augmentation follow-on converts that gap into ten controlled experiment designs. Each task contributes one exact bounded-evidence-reveal candidate and one same-slot source-ID invariance control. Candidate and control structure pass 10/10, but the sealed author expectation is not an observed semantic label. The public summary therefore reports design readiness with zero human labels, zero admitted coupled transitions, no oracle edit, and no scientific-coupling claim. See rl_env/specs/frontier_coupled_augmentation_summary.json.

The next control layer compiles ten blinded canonical/candidate/content-null triplets. Candidate and placebo reveal arms have identical field-level deltas, equal whitespace-token counts, and counterbalanced 3/3/4 role placement across three arm IDs. All 10/10 structural controls pass. Tokenizer-level matching, placebo scientific invariance, candidate coupling, and contrast identifiability remain false, with zero reviewer responses or labels. See rl_env/specs/frontier_coupled_placebo_summary.json.

The tokenizer-aware follow-on compiles ten five-arm controls with three vocabulary-separated placebos per candidate. Frozen tiktoken 0.14.0 cl100k_base and o200k_base provenance includes official source-asset hashes and normalized vocabulary fingerprints. Structural and whitespace controls pass 30/30, token counts match 60/60, and every role occupies every arm ID exactly twice. Against a frozen trial-zero baseline, deterministic search reduces aggregate token-profile L1 distance from 1908 to 1454 (23.8%): all 30/30 placebos improve strictly, all 120/120 encoding-by-profile components avoid regression, and 102/120 improve strictly. These are in-search proxy diagnostics, not held-out validation or distributional equivalence. Token byte-length and rank-decile histogram exact matches remain 0/60; distributional, lexical, scientific-invariance, coupling, and contrast claims therefore remain false with zero reviewer responses or labels. See rl_env/specs/frontier_tokenizer_placebo_summary.json.

A post-selection diagnostic evaluates frozen r50k_base and p50k_base without using them for search. Aggregate L1 improves from 2166 to 1940 (10.4%), but only 52/60 encoding comparisons and 26/30 placebos avoid regression; four placebos regress and exact held-out counts match only 4/60. The two encodings share a 50k family and the audit was not preregistered, so robust generalization and held-out token-count-control fields remain false.

A locally sealed second-stage protocol then evaluates revision-pinned BERT WordPiece and T5 SentencePiece unigram assets. This protocol was frozen before evaluation but was not publicly preregistered or externally timestamped, and its results were not used to retune the upstream selection. Aggregate profile L1 improves from 1720 to 1532 (10.9%), yet only 51/60 encoding comparisons, 94/120 profile components, and 23/30 placebos avoid regression. Seven placebos regress, token-count-gap non-regression reaches 44/60, and optimized exact token counts match only 1/60. Independent-family generalization, token-count confound control, semantic invariance, and the overall independent-family control therefore remain false. The public protocol and payload-free aggregate are in rl_env/specs/frontier_tokenizer_independent_evaluation_*.json; model assets and per-placebo diagnostics remain excluded.

Intended Use

• Review the public system architecture and release boundary.
• Read the caveats-first SCD vertical slice before citing benchmark numbers.
• Read the small-N target-identification results card and aggregate claim ledger.
• Inspect schema and verifier-contract documentation.
• Run the illustrative, non-benchmark eight-stage control-plane demo.
• Run the dependency-free adds-bounded-agent-demo planner-to-transition fixture.
• Inspect tests/test_program_runner.py for cumulative-ledger multi-stage stopping and exact replay.
• Inspect tests/test_semantic_mappings.py for the explicit unmet-need and functional-effect non-implication boundaries.
• Inspect tests/test_pinned_evidence_adapter.py for composite pinned-source gates, matched independent/same-source cases, eight-stage provider-backed execution through clinical endpoint/safety design and regulatory review, and exact replay.
• Inspect docs/14_target_identity_continuity.md and tests/test_target_identity_continuity.py for the canonical Ensembl-to-ChEMBL target ledger, namespace invariants, candidate links, and matched target-symbol success/failure pair.
• Inspect docs/15_discovery_context_identity.md and tests/test_context_identity_continuity.py for the disease, assay, and model-system ledgers, evidence links, stage requirements, rebinding/collision attacks, and fail-closed behavior.
• Inspect docs/16_clinical_intervention_identity.md and tests/test_clinical_identity_continuity.py for candidate-to-intervention-to-trial-design continuity, source identity checks, regulatory extension, and fail-closed attacks.
• Inspect docs/17_pinned_source_ingestion.md and tests/test_pinned_evidence_ingestion.py for exact source receipts, external bundle integrity, payload-free compilation, review gates, and matched bounded-stage integration.
• Inspect docs/18_cdc_mmwr_ingestion.md and tests/test_cdc_mmwr_ingestion.py for the CDC provider-specific article, section, value, unit, context, excerpt-removal, and matched independent-source/same-document controls.
• Inspect docs/19_ncbi_pubmed_ingestion.md and tests/test_ncbi_pubmed_ingestion.py for strict EFetch request and article identity, structured abstract evidence, typed treatment-gap values, excerpt removal, matched-context advance, and cross-population defer behavior.
• Inspect docs/20_preclinical_provider_ingestion.md, tests/test_chembl_activity_ingestion.py, tests/test_ncbi_pubmed_disease_model_ingestion.py, and tests/test_preclinical_provider_pair.py for release-bound ChEMBL activity, typed PubMed in-vivo evidence, candidate aliases, publication lineage, and matched advance/shared-lineage defer behavior.
• Inspect docs/preclinical_provider_validation_snapshot.json for the payload-free machine record of external source ids, typed values, hashes, matched outcomes, and limitations.
• Inspect docs/21_clinical_provider_ingestion.md, docs/clinical_provider_validation_snapshot.json, and docs/42_uc_provider_validation.md, docs/uc_clinical_provider_validation_snapshot.json, and tests/test_clinicaltrials_gov_ingestion.py for exact ClinicalTrials.gov receipt, NCT, arm, population, endpoint, treatment phase, posted serious-adverse-event aggregate, direction-aware atomic promotion, external hashes, retained uncertainty, and matched missing-safety behavior.
• Inspect docs/22_clinical_benefit_risk_synthesis.md and tests/test_clinical_benefit_risk_synthesis.py for explicit multi-trial endpoint/safety selections, retained trial values and hashes, ratio and percentage-point decision precision, non-pooling boundaries, exact replay, and tamper controls.
• Inspect docs/45_ra_acr20_risk_difference_hold_replication.md and docs/ra_acr20_risk_difference_validation_snapshot.json for the payload-free public-source RA ACR20 replication, source and artifact hashes, uncertain interval, and HOLD outcome.
• Inspect docs/46_ra_olokizumab_source_disjoint_additive_tensor.md and docs/ra_olokizumab_additive_tensor_validation_snapshot.json for the first real source-disjoint additive tensor, source-scale preservation, population non-exchangeability, and bounded HOLD.
• Inspect docs/47_ra_olokizumab_population_stratified_transport.md and docs/ra_olokizumab_population_transport_report.json for exact prior-therapy strata and the fail-closed transport estimability result.
• Inspect docs/48_ra_olokizumab_mtx_ir_same_stratum_replication.md, docs/ra_olokizumab_mtx_ir_replication_spec.json, and docs/ra_olokizumab_mtx_ir_replication_report.json for partial-date provenance, independent MTX-IR two-trial support, the exact blocker delta, and the remaining non-transport boundary.
• Inspect docs/49_ra_olokizumab_mtx_ir_outcome_risk_of_bias.md, docs/ra_olokizumab_mtx_ir_risk_of_bias_spec.json, and docs/ra_olokizumab_mtx_ir_risk_of_bias_report.json for the outcome-specific public-source judgments, denominator/chronology checks, exact PDF/registry provenance, and narrow blocker delta.
• Inspect docs/23_clinical_portfolio_endpoint_mapping.md and tests/test_clinical_portfolio.py for exact-set multi-job/bundle preflight, payload-free output, reviewer-approved ontology identity, append-only mapping continuity, and atomic failure controls.
• Inspect docs/24_policy_replanning_and_resume.md and tests/test_policy_replanning.py for typed observations, bounded rule application, append-only queue revisions, hash-bound checkpoints, and deterministic resume.
• Inspect docs/25_cutoff_safe_policy_evaluation.md, docs/retrospective_policy_evaluation_snapshot.json, and tests/test_sealed_evaluation.py for role-neutral board sealing, externally separated labels, commitment opening, strict JSON round-trip, exact submission binding, aggregate policy comparison, and leakage controls.
• Inspect docs/26_independent_heldout_evaluation.md and tests/test_heldout_evaluation.py for preregistered cohort/label/curator contracts, strict-majority and independent-adjudication validation, stage minima, Wilson intervals, action coverage, selective risk, and the explicit no-real-result boundary.
• Inspect docs/27_clinical_evidence_tensor_and_voi.md, agentic_drug_discovery/clinical_decision.py, and tests/test_clinical_benefit_risk_synthesis.py for committed-synthesis tensor compilation, provenance-linked gaps, deterministic bounded VOI ranking, budget failure, safety-signal hold, integrity checks, and the evidence-workflow-only decision boundary.
• Inspect docs/28_clinical_evidence_closed_loop.md, agentic_drug_discovery/clinical_closed_loop.py, and the adjacent transition schema/example for exact selected-action execution, compact payload-free receipts, bounded reviewer-verifier refresh, single-use actions, exact source rejoin, and two-state transition replay.
• Inspect docs/29_clinical_cohort_diagnostics.md, agentic_drug_discovery/clinical_cohort.py, and the adjacent manifest/report schemas and compiler-generated examples for accepted-state binding, evidence-unit identity, matched policy sensitivity, exact gap/action denominators, cross-unit provenance overlap, and the explicit no-outcome/no-calibration boundary.
• Inspect docs/30_preregistered_clinical_outcome_evaluation.md, agentic_drug_discovery/clinical_outcome_evaluation.py, and the adjacent protocol/submission/ manifest/report schemas for cutoff-safe package-bound forecasts, endpoint/safety provenance, aggregate calibration, paired policy comparisons, and the evaluator-only unit-label boundary.
• Inspect docs/31_cluster_robust_clinical_outcome_uncertainty.md, agentic_drug_discovery/clinical_outcome_uncertainty.py, and the adjacent dependence/protocol/ report schemas for exact assignment coverage, known-overlap closure, aggregate CR1 intervals, fixed stage-by-endpoint strata, explicit no-interval states, and the private assignment boundary.
• Inspect docs/32_prospective_clinical_outcome_design_simulation.md, agentic_drug_discovery/clinical_outcome_design_simulation.py, and the adjacent design protocol/report schemas for beta-binomial known-truth simulation, production CR1 parity, IID diagnostic comparison, Monte Carlo target checks, and the no-automatic-selection boundary.
• Inspect docs/33_informative_evaluability_and_dependence_stress.md, agentic_drug_discovery/clinical_outcome_stress_simulation.py, and the adjacent stress protocol/report schemas for analytic estimand shifts, exact dependence blocks, nominal/oracle-closure CR1 comparison, combined stress signatures, and the no-automatic-correction boundary.
• Inspect docs/34_preregistered_pattern_mixture_sensitivity.md, agentic_drug_discovery/clinical_outcome_pattern_mixture.py, and the adjacent pattern-mixture protocol/report schemas for exact stress binding, binary log-IMOR grids, prediction-stratum observability, matched recovery controls, point-envelope interpretation, and the no-automatic-range-selection boundary.
• Inspect docs/35_dependence_closed_pattern_mixture_uncertainty.md, agentic_drug_discovery/clinical_outcome_pattern_mixture_uncertainty.py, and the adjacent protocol/report schemas for fixed-assumption model functionals, delete-one-cluster variance, nominal/dependence-closed calibration, Monte Carlo bounds, explicit no-interval states, and the no-automatic-closure boundary.
• Inspect docs/36_unequal_cluster_influence_calibration.md, agentic_drug_discovery/clinical_outcome_pattern_mixture_influence_calibration.py, and the adjacent protocol/report/summary schemas for Student-t critical values, unequal delete-mj pseudovalues, experimental multiplier diagnostics, dominance hard stops, and the no-automatic-selection boundary.
• Inspect docs/37_informative_cluster_size_estimands.md, agentic_drug_discovery/clinical_outcome_informative_cluster_size.py, and the adjacent protocol/report/summary schemas for unit-weighted and cluster-balanced truths, fixed-profile conditional calibration, aggregate influence concentration, and the no-automatic-estimand- selection boundary.
• Inspect docs/38_cluster_superpopulation_sampling.md, agentic_drug_discovery/clinical_outcome_cluster_superpopulation.py, and the adjacent protocol/report/summary schemas for empirical-template cluster resampling, exact known-truth preservation, conditional-versus-superpopulation calibration, realized-design diagnostics, and the no-transportability/no-post-hoc-filtering boundary.
• Inspect rl_env/specs/pinned_evidence_manifest.schema.json and its synthetic example before constructing a source manifest.
• Inspect rl_env/specs/target_identity_record.schema.json and its synthetic example before producing or consuming serialized target records.
• Inspect rl_env/specs/discovery_context_identity.schema.json and its synthetic example before producing or consuming serialized disease, assay, or model-system records.
• Inspect rl_env/specs/clinical_intervention_identity.schema.json and its synthetic example before producing or consuming serialized clinical intervention, trial, or atomic design records.
• Inspect rl_env/specs/clinical_benefit_risk_synthesis.schema.json and its synthetic example before selecting source-ledger trials for cross-trial harmonization.
• Inspect rl_env/specs/clinical_evidence_decision_package.schema.json and its compiler-generated synthetic example before producing or consuming policy-bound evidence tensors or action plans.
• Inspect rl_env/specs/clinical_evidence_decision_config.schema.json and its synthetic example, then use adds-clinical-evidence to compile, validate, or summarize a package without importing internal dataclasses.
• Use adds-clinical-evidence cohort, validate-cohort, and summarize-cohort with the cohort manifest/report contracts to compare exact package rosters without treating policy variants as independent clinical observations.
• Use adds-clinical-evidence evaluate-outcomes, validate-outcomes, and summarize-outcomes with frozen package forecasts and an evaluator-controlled outcome manifest. The checked-in one-unit example verifies contract execution only and is not calibration evidence.
• Inspect rl_env/specs/clinical_endpoint_mapping.schema.json and rl_env/specs/clinicaltrials_gov_portfolio_job.schema.json before approving an endpoint family or assembling an exact multi-trial source bundle.
• Inspect rl_env/specs/source_receipt.schema.json and rl_env/specs/pinned_evidence_ingestion_job.schema.json before capturing or compiling a source.
• Inspect rl_env/specs/cdc_mmwr_ingestion_job.schema.json before authoring a CDC MMWR review job.
• Inspect rl_env/specs/ncbi_pubmed_ingestion_job.schema.json before authoring an NCBI PubMed treatment-gap review job.
• Inspect rl_env/specs/chembl_activity_ingestion_job.schema.json and rl_env/specs/ncbi_pubmed_disease_model_ingestion_job.schema.json before authoring preclinical provider review jobs.
• Inspect rl_env/specs/clinicaltrials_gov_ingestion_job.schema.json before authoring a reviewed registry study, selected-arm, population, endpoint, analysis, and serious-adverse-event contract.
• Inspect strict replay bundles and run the machine-readable adds-replay-bundle CLI.
• Use benchmark/ to score the separately hosted clinical-trial decision dataset.
• Track provenance for the public artifact surface.
• Inspect the mirrored GitHub release surface and source-commit provenance.

Sealed Retrospective Evaluation

The external evaluator executed four matched pairs and eight cutoff-safe episodes built from the real senicapoc continuous program and PALOMA-2/PALOMA-3 clinical portfolio. Only the payload-free aggregate and artifact hashes are included here.

Policy	Exact	Success arm	Failure arm	Both correct	Unsafe advance
Deterministic gated stage output	8/8	4/4	4/4	4/4	0/7
Always advance counterfactual	1/8	1/4	0/4	0/4	7/7
Defer-safe counterfactual	4/8	0/4	4/4	0/4	0/7

This is a small contract diagnostic. It does not establish drug-discovery performance, prospective clinical utility, policy optimality, or confidence calibration. The complete board, cached real packets, label vault, commitment nonces, submissions, and per-episode evaluations stay outside both public release surfaces.

Artifact Map

Path	Purpose
README.md	This Hugging Face Dataset card.
github/README.md	GitHub README preserved inside the Hub mirror.
release_manifest.json	Cross-surface release manifest.
release_decision_packet.json	Machine-readable launch decision packet.
huggingface/release_manifest.json	Hugging Face-specific include/exclude manifest.
upload_manifest.json	Exact uploaded file list and source commit.
docs/release_boundary.md	Public-release boundary and exclusion rules.
docs/release_trust_report.md	Trust claims, machine anchors, and interpretation warnings.
docs/12_scd_vertical_slice.md	Audited SCD vertical slice, with small-N caveats.
docs/13_target_id_governance_node.md	Upstream target-identification results card.
docs/14_target_identity_continuity.md	Executable target ledger, stage namespace requirements, and fail-closed identity rules.
docs/15_discovery_context_identity.md	Disease, assay, and model-system ledgers, evidence links, stage gates, and matched failure contract.
docs/16_clinical_intervention_identity.md	Candidate-to-intervention-to-trial-design continuity, source checks, regulatory extension, and failure contract.
docs/17_pinned_source_ingestion.md	Exact external source capture, payload-free compilation, review gates, and control-plane integration.
docs/18_cdc_mmwr_ingestion.md	CDC MMWR article binding, evidence-location checks, payload-free extraction, and matched stage behavior.
docs/19_ncbi_pubmed_ingestion.md	NCBI PubMed XML identity, structured abstract anchors, payload-free extraction, and context-mismatch behavior.
docs/20_preclinical_provider_ingestion.md	ChEMBL functional-activity and PubMed disease-model contracts, payload-free external validation snapshot, and lineage-independence failure control.
docs/preclinical_provider_validation_snapshot.json	Payload-free machine record of provider ids, typed values, hashes, matched outcomes, and limitations.
docs/21_clinical_provider_ingestion.md	ClinicalTrials.gov source receipt, endpoint/safety design identities, bounded promotion, and matched failure contract.
docs/clinical_provider_validation_snapshot.json	Payload-free NCT/design/safety identities, artifact hashes, live stage outcome, matched control, and limitations.
docs/42_uc_provider_validation.md	Public-source UC induction provider execution, direction-aware gating, treatment-phase identity, and non-pooling limits.
docs/uc_clinical_provider_validation_snapshot.json	Payload-free UC source/job/output/manifest hashes, selected aggregates, decisions, and exact-replay limits.
docs/22_clinical_benefit_risk_synthesis.md	Explicit reviewed selection, retained trial values, source-disjoint provenance, non-pooling boundary, and fail-closed synthesis behavior.
docs/23_clinical_portfolio_endpoint_mapping.md	Exact multi-bundle portfolio transaction, reviewer-approved endpoint mapping ledger, synthesis dependency, and release boundary.
docs/24_policy_replanning_and_resume.md	Typed policy observations, bounded replans, checkpoint integrity, and deterministic resume.
docs/25_cutoff_safe_policy_evaluation.md	Cutoff-safe sealing, submission, scoring, real aggregate results, and interpretation limits.
docs/26_independent_heldout_evaluation.md	Preregistered held-out protocol, evaluator-only curator manifest, stage uncertainty, and release boundary.
docs/27_clinical_evidence_tensor_and_voi.md	Exact evidence cells, typed workflow gaps, bounded VOI ranking, budget behavior, provenance replay, and interpretation boundaries.
docs/29_clinical_cohort_diagnostics.md	Exact package/state rosters, evidence-unit identity, matched policy sensitivity, provenance overlap, and calibration boundaries.
docs/30_preregistered_clinical_outcome_evaluation.md	Package-bound probability forecasts, cutoff-safe endpoint/safety outcomes, aggregate calibration, paired policy metrics, and evaluator-only boundaries.
docs/31_cluster_robust_clinical_outcome_uncertainty.md	Dependence commitments, known-overlap closure, aggregate CR1 intervals, fixed strata, fail-closed diagnostics, and interpretation boundaries.
docs/32_prospective_clinical_outcome_design_simulation.md	Beta-binomial design scenarios, analytic truths, CR1/IID coverage comparison, Monte Carlo target checks, and gate-selection boundaries.
docs/33_informative_evaluability_and_dependence_stress.md	Outcome-dependent evaluability, analytic population/evaluable shifts, residual dependence blocks, nominal/oracle-closure CR1 comparison, and correction boundaries.
docs/34_preregistered_pattern_mixture_sensitivity.md	Prediction-stratified binary log-IMOR sensitivity, observable aggregate inputs, matched recovery gates, public synthetic results, and claim boundaries.
docs/35_dependence_closed_pattern_mixture_uncertainty.md	Fixed log-IMOR model functionals, dependence-closed cluster jackknife, all-grid calibration, Monte Carlo precision, synthetic results, and claim boundaries.
docs/36_unequal_cluster_influence_calibration.md	Few, unequal, and dominant-cluster interval calibration, Student-t and delete-mj comparisons, experimental multiplier diagnostics, and operational boundaries.
docs/37_informative_cluster_size_estimands.md	Unit-weighted versus cluster-balanced functionals, informative-size direction drift, fixed-profile calibration, influence concentration, and estimand-selection boundaries.
docs/38_cluster_superpopulation_sampling.md	Empirical-template cluster-superpopulation sampling, preserved known truths, conditional calibration comparison, realized-design diagnostics, and transport boundaries.
docs/39_biohub_translational_evidence_bridge.md	Independent Biohub-context fit review, concrete 90-day pilot, ten-slide narrative, and explicit non-claims.
docs/biohub_research_readiness.json	Machine-readable official-source alignment, artifact hashes, maturity ledger, fit matrix, pilot gates, presentation claims, and open gaps.
docs/40_upstream_translational_handoff.md	Generic M6 handoff contract, contextual-only compilation semantics, validation boundary, and next external-review decision.
docs/41_ulcerative_colitis_conformance_slice.md	Synthetic immune/inflammatory disease conformance across M6, ratio-effect mapping, non-pooled synthesis, and decision-tensor direction checks.
docs/42_uc_provider_validation.md	Two public-source UC induction provider runs with direction-aware gating and payload-free artifact identity.
docs/43_uc_phase_population_alignment.md	Real-source induction/maintenance phase-population alignment within one non-pooled UC trial.
docs/uc_phase_population_validation_snapshot.json	Payload-free phase counts, effect and safety aggregates, artifact hashes, stage outcomes, and non-exchangeability boundary.
docs/44_uc_maintenance_risk_difference_replication.md	Independent primary-maintenance percentage-point replication with screened exclusion controls and explicit decision-layer boundary.
docs/uc_maintenance_risk_difference_validation_snapshot.json	Payload-free source/output hashes, additive effect, role-wise safety/population alignment, execution result, and exclusion codes.
docs/45_ra_acr20_risk_difference_hold_replication.md	Public-source RA ACR20 percentage-point replication with uncertain-interval HOLD behavior and a real multi-trial boundary.
docs/ra_acr20_risk_difference_validation_snapshot.json	Payload-free RA source and artifact hashes, additive endpoint and safety aggregates, decision result, and limitations.
docs/46_ra_olokizumab_source_disjoint_additive_tensor.md	Same-candidate RA source selection, proportion-scale risk differences, non-pooled synthesis, and safety-triggered HOLD.
docs/ra_olokizumab_additive_tensor_validation_snapshot.json	Payload-free two-source trial aggregates, hashes, normalized precision, decision result, and negative claims.
docs/47_ra_olokizumab_population_stratified_transport.md	Source-bound prior-therapy stratification and explicit transport estimability boundary.
docs/ra_olokizumab_population_transport_report.json	Integrity-bound strata, trial-level effects, support counts, blockers, and prohibited-inference flags.
docs/48_ra_olokizumab_mtx_ir_same_stratum_replication.md	Independent MTX-IR same-stratum execution, source-date precision, blocker delta, and HOLD boundary.
docs/ra_olokizumab_mtx_ir_replication_spec.json	Exact reviewed MTX-IR stratum bindings and source-field fingerprints.
docs/ra_olokizumab_mtx_ir_replication_report.json	Integrity-bound two-trial support, source effects, safety aggregates, remaining blockers, and prohibited-inference flags.
docs/49_ra_olokizumab_mtx_ir_outcome_risk_of_bias.md	Outcome-specific judgments, denominator and chronology checks, source provenance, and blocker delta.
docs/ra_olokizumab_mtx_ir_risk_of_bias_spec.json	Reviewed trial/arm/domain bindings with registry-field and protocol/SAP-page citations.
docs/ra_olokizumab_mtx_ir_risk_of_bias_report.json	Integrity-bound judgments, source hashes, diagnostics, cautions, and remaining transport blockers.
docs/retrospective_policy_evaluation_snapshot.json	Payload-free machine aggregate with policy metrics, artifact hashes, gate outcomes, and withheld-data boundary.
docs/public_evidence_summary.json	Machine-readable aggregate claims and limitations.
agentic_drug_discovery/	Bounded planning, typed tool execution, semantic promotion, stage and program orchestration, matched evaluation, replay, and fail-closed transitions.
agentic_drug_discovery/sealed_evaluation.py	Role-neutral board sealing, external label vaults, commitments, strict envelope readers, submission validation, and aggregate scoring.
agentic_drug_discovery/heldout_evaluation.py	Preregistered protocol binding, independent curation validation, Wilson intervals, action coverage, selective risk, and strict aggregate reporting.
agentic_drug_discovery/frontier*.py	ADDS-Frontier protocol, public development fixtures, frozen board allocation, private calibration validation, executable preflight probes, blinded semantic-review compilation, deterministic disagreement triage, sealed canonical replay, independent-oracle challenge comparison, oracle fragility diagnostics, support-curation workload compilation, cross-stage coupling audit, matched augmentation, three-arm placebo and five-arm tokenizer controls, and payload-free reporting.
agentic_drug_discovery/ingestion.py	Immutable source receipts, external bundle verification, payload-free manifest compilation, and review reports.
agentic_drug_discovery/cdc_mmwr.py	CDC MMWR article and reviewer-selected evidence verification with excerpt removal.
agentic_drug_discovery/ncbi_pubmed.py	NCBI PubMed EFetch article and treatment-gap evidence verification with excerpt and anchor removal.
agentic_drug_discovery/chembl_activity.py	ChEMBL release/resource reconciliation and typed functional-activity verification with assay-text removal.
agentic_drug_discovery/clinicaltrials_gov.py	ClinicalTrials.gov study, arm, population, endpoint, statistical-analysis, serious-adverse-event, and day/month/year chronology verification with payload removal.
agentic_drug_discovery/clinical_portfolio.py	Atomic exact-set multi-trial extraction with source-hash disjointness and payload-free output.
agentic_drug_discovery/clinical_endpoint_mapping.py	Strict approved-mapping parser, endpoint/safety fingerprint compiler, approval chronology, and replay validation.
agentic_drug_discovery/clinical_synthesis.py	Deterministic source-ledger compiler for supported trial-level ratio or percentage-point effects and serious-event risk differences without pooling.
agentic_drug_discovery/clinical_decision.py	Committed-synthesis tensor compiler with scale-specific ratio/additive precision, typed gaps, deterministic budget-aware bounded VOI planner, integrity envelopes, and state replay.
agentic_drug_discovery/clinical_cohort.py	Accepted-state-bindable package rosters, deterministic cohort aggregation, matched policy comparisons, strict readers, and cross-unit provenance overlap.
agentic_drug_discovery/clinical_outcome_evaluation.py	Preregistered protocol and submission binding, post-deadline outcome provenance, aggregate Brier/calibration/threshold metrics, paired policy comparisons, and full replay.
agentic_drug_discovery/clinical_outcome_uncertainty.py	Frozen dependence commitments, exact assignment coverage, known-overlap closure, CR1 policy/stratum/paired intervals, strict readers, and full replay.
agentic_drug_discovery/clinical_outcome_design_simulation.py	Bounded deterministic beta-binomial simulation, analytic truths, production CR1 parity, IID diagnostics, candidate-gate evaluation, strict readers, and replay.
agentic_drug_discovery/clinical_outcome_stress_simulation.py	Bounded block-Polya stress simulation, analytic population/evaluable truths, nominal/dependence-closed CR1 comparison, strict boundaries, readers, summaries, and replay.
agentic_drug_discovery/clinical_outcome_pattern_mixture.py	Exact stress-bound binary log-IMOR grids, prediction-stratified aggregate estimators, matched calibration/recovery/identification diagnostics, strict readers, summaries, and replay.
agentic_drug_discovery/clinical_outcome_pattern_mixture_uncertainty.py	Exact point-report binding, nominal/dependence-closed delete-one-cluster jackknife inference, model-functional coverage, Monte Carlo bounds, fail-closed statuses, strict readers, summaries, and replay.
agentic_drug_discovery/clinical_outcome_pattern_mixture_influence_calibration.py	Student-t critical values, unequal delete-mj pseudovalues, experimental multiplier intervals, production eligibility, Monte Carlo calibration, strict readers, summaries, and replay.
agentic_drug_discovery/clinical_outcome_informative_cluster_size.py	Profile-bound dual estimands, three Student-t methods, conditional calibration, aggregate max-block influence, strict readers, summaries, and exact replay.
agentic_drug_discovery/clinical_outcome_cluster_superpopulation.py	Empirical-template cluster resampling, preserved superpopulation truths, conditional comparison cells, realized-design rates, dynamic largest-cluster influence, strict readers, and replay.
agentic_drug_discovery/research_readiness.py	Duplicate-safe readiness reader, artifact-hash verification, maturity and fit checks, immutable pilot gates, presentation claim validation, and compact summary.
agentic_drug_discovery/translational_handoff.py	Strict perturbation-context reader, source/lineage/interval/QC/review validation, canonical integrity hash, summary, and contextual-only evidence compiler.
adapters/pinned_evidence_adapter.py	Dependency-free validation and lookup for source-pinned, payload-free evidence manifests.
adapters/clinical_synthesis_adapter.py	Local normalization of approved endpoint mappings and reviewed synthesis selections without supplied source measurements.
adapters/execution_registry.py	Typed contracts for the pinned adapter and caller-supplied GitHub adapter instances.
rl_env/specs/pinned_evidence_manifest.schema.json	Machine-readable pinned-record schema; the adjacent example is synthetic.
rl_env/specs/target_identity_record.schema.json	Machine-readable cross-stage target record; the adjacent example is synthetic.
rl_env/specs/discovery_context_identity.schema.json	Machine-readable disease, assay, and model-system records; the adjacent example is synthetic.
rl_env/specs/clinical_intervention_identity.schema.json	Machine-readable clinical intervention, trial, endpoint, safety, and atomic design records; the adjacent example is synthetic.
rl_env/specs/clinical_endpoint_mapping.schema.json	Machine-readable approved reviewer, ontology identity, and exact endpoint/safety binding contract; the adjacent example is synthetic.
rl_env/specs/clinical_benefit_risk_synthesis.schema.json	Machine-readable reviewed multi-trial selection contract; the adjacent example is synthetic.
rl_env/specs/clinical_evidence_decision_package.schema.json	Integrity-bound policy, exact tensor, gaps, action catalog, budget, and bounded-VOI plan contract; the adjacent example is synthetic.
rl_env/specs/clinical_evidence_decision_package.relaxed.example.json	Compiler-generated synthetic ADVANCE package over the same evidence unit for reproducible matched-policy sensitivity.
rl_env/specs/clinical_evidence_cohort_manifest.schema.json	Exact package roster and optional all-or-none accepted-state SHA-256 binding contract; the adjacent example is synthetic.
rl_env/specs/clinical_evidence_cohort_report.schema.json	Package/policy strata, matched transitions, gap/action diagnostics, provenance overlap, and explicit no-outcome calibration-status contract; the adjacent example is synthetic.
rl_env/specs/clinical_evidence_cohort_summary.schema.json	Compact cohort summary and optional validation-status contract.
rl_env/specs/clinical_outcome_evaluation_protocol.schema.json	Public cohort/cutoff/outcome/harmonization/curation/metric preregistration contract; the adjacent example is synthetic.
rl_env/specs/clinical_prediction_submission.schema.json	Exact package/evidence-unit-bound favorable-outcome probability contract; adjacent examples are synthetic.
rl_env/specs/clinical_outcome_manifest.schema.json	Evaluator-only endpoint/safety assessment and post-deadline source-provenance contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_evaluation_report.schema.json	Aggregate attrition, Wilson, Brier/calibration/threshold, paired-policy, and provenance-overlap contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_dependence_manifest.schema.json	Evaluator-only exact unit-to-cluster assignment and dependence-basis contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_uncertainty_protocol.schema.json	Public dependence-construction, confidence, cluster-floor, dominance, strata, and metric preregistration contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_uncertainty_report.schema.json	Aggregate cluster diagnostics and CR1 policy, stratum, and paired-policy interval contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_uncertainty_summary.schema.json	Compact dependence-aware uncertainty and optional validation-status contract.
rl_env/specs/clinical_outcome_design_simulation_protocol.schema.json	Seeded cluster-size, prevalence, ICC, evaluability, prediction-pattern, gate, and Monte Carlo design contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_design_simulation_report.schema.json	Aggregate analytic truth, replicate diagnostic, IID/CR1 performance, gate status, and privacy-boundary contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_design_simulation_summary.schema.json	Compact scenario/gate coverage, yield, width, error, status, and optional replay-validation contract.
rl_env/specs/clinical_outcome_stress_simulation_protocol.schema.json	Exact dependence partitions, outcome-specific evaluability, fixed estimand/mode comparison, gate, RNG, and Monte Carlo commitments; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_stress_simulation_report.schema.json	Aggregate analytic shifts, mode diagnostics, target performance, and fixed claim-boundary contract; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_stress_simulation_summary.schema.json	Compact evaluable-only target passage and dependence-closure recovery contract.
rl_env/specs/clinical_outcome_pattern_mixture_protocol.schema.json	Exact stress binding, log-IMOR grid, analyzability, mean-envelope width, bias, and method commitments; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_pattern_mixture_report.schema.json	Aggregate grid curves, estimand bias, matched recovery, point-envelope inclusion, fail-closed support, and claim boundaries; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_pattern_mixture_summary.schema.json	Compact log-IMOR, analyzability, recovery, identification, and claim-boundary contract.
rl_env/specs/clinical_outcome_pattern_mixture_uncertainty_protocol.schema.json	Exact stress/protocol/report binding, fixed analysis modes, cluster gates, calibration targets, and closure anchor; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_pattern_mixture_uncertainty_report.schema.json	Aggregate grid-level model-functional coverage, population recovery, Monte Carlo bounds, jackknife diagnostics, closure comparisons, and claim boundaries; the adjacent example is synthetic.
rl_env/specs/clinical_outcome_pattern_mixture_uncertainty_summary.schema.json	Compact all-grid calibration, truth-aligned coverage, dependence-closure response, and claim-boundary contract.
rl_env/specs/clinical_outcome_pattern_mixture_influence_protocol.schema.json	Exact stress/point-report binding, canonical method order, calibration gates, production eligibility, and multiplier commitments.
rl_env/specs/clinical_outcome_pattern_mixture_influence_report.schema.json	Aggregate method-grid-metric calibration, unequal-cluster structure, hard-stop, RNG, and claim-boundary contract.
rl_env/specs/clinical_outcome_pattern_mixture_influence_summary.schema.json	Compact method comparison, primary-grid diagnostics, production eligibility, and claim-boundary contract.
rl_env/specs/clinical_outcome_informative_cluster_size_protocol.schema.json	Exact stress binding, block-specific prevalence profiles, canonical estimand methods, calibration targets, and direction threshold.
rl_env/specs/clinical_outcome_informative_cluster_size_report.schema.json	Aggregate dual-estimand truths, own/alternate bias, conditional interval calibration, influence concentration, production hard stops, and privacy boundaries.
rl_env/specs/clinical_outcome_informative_cluster_size_summary.schema.json	Compact estimand-drift, reference calibration, influence, sampling-frame, and claim-boundary contract.
rl_env/specs/clinical_outcome_cluster_superpopulation_protocol.schema.json	Exact stress/fixed-protocol/fixed-report binding, empirical sampling rule, seed, canonical methods, and calibration targets.
rl_env/specs/clinical_outcome_cluster_superpopulation_report.schema.json	Aggregate preserved truths, conditional comparisons, realized-design rates, dynamic influence, deterministic substreams, and claim boundaries.
rl_env/specs/clinical_outcome_cluster_superpopulation_summary.schema.json	Compact sampling-frame finding, recovery counts, primary-functional comparison, design rates, and transport boundary.
rl_env/specs/sealed_evaluation_board.schema.json	Policy-visible, role-neutral cutoff episode and matched-pair board contract.
rl_env/specs/sealed_evaluation_vault.schema.json	Evaluator-only label, failure-cause, arm-role, and commitment-opening contract.
rl_env/specs/policy_evaluation_submission.schema.json	Exact board-bound policy prediction and confidence contract.
rl_env/specs/policy_evaluation_report.schema.json	Aggregate and evaluator-only per-episode scoring report contract.
rl_env/specs/heldout_evaluation_protocol.schema.json	Public preregistration contract for cohort, labels, opaque roster, stages, and metrics; the adjacent example is synthetic.
rl_env/specs/heldout_curation_manifest.schema.json	Evaluator-only opaque declarations, votes, consensus, and adjudication contract.
rl_env/specs/stage_stratified_evaluation_report.schema.json	Aggregate exact counts, Wilson intervals, sufficiency flags, action coverage, and selective-risk contract; the adjacent example is synthetic.
rl_env/specs/frontier_calibration_authoring_progress.schema.json	Payload-free ten-slot calibration progress with salted commitments, explicit gate states, and no-result boundaries; private payload schemas are adjacent.
rl_env/specs/frontier_calibration_preflight_summary.schema.json	Payload-free 60-stage scorer round-trip, machine-decidable versus semantic-review-required mutation counts, unchanged human gates, and a private-report commitment.
rl_env/specs/frontier_semantic_review_readiness_summary.schema.json	Payload-free commitments for 20 blinded semantic-review drafts, exact minimal deltas, zero reviewer/consensus state, and strict no-result boundaries.
rl_env/specs/frontier_semantic_review_workflow_summary.schema.json	Three-independent-response consensus, observed pair-delta equality, pre-access evidence invariance, deterministic disagreement precedence, 20 unassigned packets, and zero live response/adjudication state.
rl_env/specs/frontier_semantic_review_resolution_summary.schema.json	Consensus-gated sealed unblinding, 30-component canonical replay, human receipt, and frozen no-automatic-gate-transition rules with zero live resolution state.
rl_env/specs/frontier_oracle_challenge_readiness_summary.schema.json	Two-challenger blind solve, reveal-after-agreement, 30-component author-oracle comparison, four-route adjudication/convergence state, and zero live challenge state.
rl_env/specs/frontier_oracle_fragility_summary.schema.json	Payload-free 300-probe scorer-localization counts, witness saturation, lineage redundancy, support-readiness decision, frozen nonclaims, and a private-report commitment.
rl_env/specs/frontier_oracle_support_curation_summary.schema.json	Payload-free leave-one-out workload counts, role and lineage priorities, singleton protection, frozen nonclaims, and a private-packet commitment.
rl_env/specs/frontier_oracle_transition_audit_summary.schema.json	Payload-free action/witness/access/blocker transition classes, coupling coverage, causal nonclaims, and a private-report commitment.
rl_env/specs/frontier_coupled_augmentation_summary.schema.json	Payload-free matched candidate/control readiness, zero semantic labels/admissions, causal nonclaims, and a private-packet commitment.
rl_env/specs/frontier_coupled_placebo_summary.schema.json	Payload-free counterbalanced triplet readiness, structural and whitespace-token matching, zero labels, lexical limitations, and private packet/key commitments.
rl_env/specs/frontier_tokenizer_placebo_summary.schema.json	Payload-free five-arm readiness, frozen search/evaluation tokenizer provenance, baseline-constrained optimization, a post-selection 50k-family diagnostic with four retained regressions, exact-match limitations, and private packet/key commitments.
rl_env/specs/frontier_tokenizer_independent_evaluation_protocol.schema.json	Locally pre-sealed WordPiece/SentencePiece revisions, asset hashes, package versions, fixed profile metrics, gates, failure retention, and no-retuning policy; not a public preregistration.
rl_env/specs/frontier_tokenizer_independent_evaluation_summary.schema.json	Payload-free independent-family aggregates with seven retained placebo regressions, failed profile/token-count controls, and a private-report commitment.
rl_env/specs/source_receipt.schema.json	Machine-readable exact source version, locator, hash, size, retrieval time, and transport.
rl_env/specs/pinned_evidence_ingestion_job.schema.json	Machine-readable reviewer-authored summaries linked to external source receipts.
rl_env/specs/cdc_mmwr_ingestion_job.schema.json	Machine-readable CDC MMWR article, context, value, unit, and excerpt review contract.
rl_env/specs/ncbi_pubmed_ingestion_job.schema.json	Machine-readable PubMed article, METHODS/RESULTS, typed treatment-gap value, and context-anchor contract.
rl_env/specs/chembl_activity_ingestion_job.schema.json	Machine-readable ChEMBL release, linked resource, typed endpoint, candidate alias, target, and lineage contract.
rl_env/specs/ncbi_pubmed_disease_model_ingestion_job.schema.json	Machine-readable PubMed in-vivo exposure, endpoint, model, candidate, and lineage contract.
rl_env/specs/clinicaltrials_gov_ingestion_job.schema.json	Machine-readable exact study, arm, population, endpoint, measurement, analysis, and serious-adverse-event contract.
rl_env/specs/clinicaltrials_gov_portfolio_job.schema.json	Machine-readable exact set of single-trial jobs, receipts, identities, and approved mapping bindings.
rl_env/specs/clinical_evidence_decision_config.schema.json	Machine-readable accepted-synthesis, policy, action-catalog, and output-identity compiler contract; the adjacent example is synthetic.
rl_env/specs/clinical_evidence_decision_summary.schema.json	Machine-readable compact package summary and optional state-replay validation-report contract.
tests/test_target_identity_continuity.py	Namespace rebinding/collision, broken candidate link, and matched target-symbol coverage.
tests/test_context_identity_continuity.py	Disease/model rebinding, assay collision, unknown-candidate evidence, and strict example parsing.
tests/test_clinical_identity_continuity.py	Intervention rebinding, trial collision, unknown-intervention linkage, support removal, and strict example parsing.
tests/test_pinned_evidence_ingestion.py	Receipt/job parsing, source tamper checks, compiler boundaries, CLI capture, and matched source-independence coverage.
tests/test_cdc_mmwr_ingestion.py	Provider identity, location, value, unit, excerpt removal, stage transition, and matched-pair coverage.
tests/test_ncbi_pubmed_ingestion.py	PubMed identity, request, XML/retraction, section, value, anchor, stage transition, and matched-pair coverage.
tests/test_chembl_activity_ingestion.py	ChEMBL release/resource identity, endpoint, target, alias, lineage, text-removal, and CLI-hash coverage.
tests/test_ncbi_pubmed_disease_model_ingestion.py	PubMed article, exposure, endpoint, model/candidate anchor, text-removal, and CLI-hash coverage.
tests/test_preclinical_provider_pair.py	Matched independent-lineage advance and shared-lineage defer integration coverage.
tests/test_clinicaltrials_gov_ingestion.py	Strict registry extraction, payload removal, atomic promotion, continuity attacks, and matched mismatch coverage.
tests/test_clinical_benefit_risk_synthesis.py	Two-source tool-to-replay synthesis plus mismatch, overlap, pooling, forgery, unbound-support, direct-commit, and removal controls.
tests/test_clinical_portfolio.py	Exact-set portfolio extraction, source chronology/disjointness, strict schemas, payload removal, and atomic CLI failure controls.
tests/test_clinical_decision_cli.py	Exact package reproduction, accepted-packet provenance, strict config parsing, atomic CLI output, replay validation, and compact summary coverage.
tests/test_clinical_cohort.py	State-bound roster replay, matched policy sensitivity, provenance overlap, strict readers/schemas, tamper rejection, and atomic cohort CLI coverage.
tests/test_clinical_outcome_evaluation.py	Cutoff/source/package/roster controls, attrition, calibration math, paired comparison, strict schemas/readers, and atomic outcome CLI coverage.
tests/test_clinical_outcome_uncertainty.py	CR1 math, paired covariance, fixed strata, chronology, known-overlap closure, non-estimable states, privacy, strict readers, and atomic CLI coverage.
tests/test_clinical_outcome_design_simulation.py	Analytic truths, seeded replay, ICC undercoverage stress, floor/dominance/attrition states, strict bounds/readers, privacy, schemas, and atomic CLI coverage.
tests/test_clinical_outcome_stress_simulation.py	Analytic estimand shifts, informative-selection bias, hidden-linkage undercoverage, oracle-closure recovery, combined stress, exact partitions, strict readers, privacy, schemas, and atomic CLI coverage.
tests/test_clinical_outcome_pattern_mixture.py	Binary log-IMOR recovery, excluded-grid controls, MCAR alignment, sparse-stratum failure, exact replay, privacy, schemas, and atomic CLI coverage.
tests/test_clinical_outcome_pattern_mixture_uncertainty.py	All-grid jackknife calibration, hidden-linkage repair, independent-mode equivalence, model-functional/population separation, Monte Carlo bounds, fail-closed states, exact replay, privacy, schemas, and CLI coverage.
tests/test_clinical_outcome_pattern_mixture_influence_calibration.py	Student-t references, delete-mj algebra, seed isolation, dominant-cluster hard stops, strict schemas/readers, public exact replay, and CLI coverage.
tests/test_clinical_outcome_informative_cluster_size.py	Null equivalence, positive/negative direction drift, target preservation, dominant hard stop, aggregate influence, strict schemas/readers, public replay, privacy, and CLI coverage.
tests/test_clinical_outcome_cluster_superpopulation.py	Known-truth preservation, dynamic tie denominators, no post-hoc filtering, exact binding/replay, strict schemas/readers, public findings, privacy, and CLI coverage.
tests/test_sealed_evaluation.py	Synthetic board determinism, commitment, submission, confidence, schema, leakage, and baseline-policy coverage.
tests/	Dependency-free planning, multi-stage stopping, mapping, evaluation, execution, replay, and transition regression tests.
benchmark/	Installable ctdbench scorer and tests.
docs/public_launch_checklist.md	Human launch checklist before any visibility change.
scripts/audit/*.py	Local release audits and reproducible Hub package builder.

Not Included

• Raw source snapshots or full case banks.
• Raw source bundles, real provider review jobs, ingestion runs, multi-trial portfolio selections, endpoint-family reviewer approvals, ontology-authority resolutions, or reviewer working files.
• Hidden/evaluator labels or locked episode records.
• Real sealed evaluation boards, cached episode packets, label vaults, commitment nonces, policy submissions, or per-episode evaluations.
• Generated reward/verifier outputs or run logs.
• Credentials, local machine paths, or private infrastructure details.
• Model weights or a complete autonomous discovery or wet-lab system.
• Live adapter implementations, endpoint configuration, or raw execution ledgers.
• A real matched success/failure episode corpus or a claim of discovery performance.
• A pooled meta-analysis, benefit-risk score, clinical acceptability judgment, or treatment recommendation; the synthesis path is descriptive and trial-preserving only.
• Real disease-burden, treatment-gap, functional-assay, or disease-model source payloads. The included pinned manifest example is synthetic and demonstrates the contract only.
• Source-pinned clinical registry payloads or reviewer jobs. Typed synthetic design records and one payload-free external validation snapshot are included.
• Provider-specific reviewed disease/preclinical ingestion jobs or real compiled manifests.
• Croissant metadata for jang1563/clinical-trial-decision-benchmark; that metadata belongs to the separate external dataset, not this artifact mirror.

Linked External Dataset

The scorer in benchmark/ targets https://huggingface.co/datasets/jang1563/clinical-trial-decision-benchmark. That dataset has its own card, rows, and Croissant metadata. This repository's Hub package intentionally does not duplicate those data or metadata.

Validation Before Upload

Run these checks from the GitHub repository root before creating or updating the Hugging Face repository:

python3 -m pip install -e . -e ./benchmark pytest build ruff
python3 scripts/audit/github_release_file_audit.py
python3 scripts/audit/validate_hf_release_package.py
python3 scripts/audit/validate_public_launch_packet.py
python3 scripts/audit/validate_vertical_slice_doc.py
python3 scripts/audit/validate_policy_evaluation_snapshot.py
python3 scripts/audit/validate_biohub_research_readiness.py
python3 scripts/audit/validate_translational_handoff.py
python3 -m unittest discover -s tests -v
python3 -m ruff check agentic_drug_discovery tests adapters/boltz_adapter.py adapters/chembl_adapter.py adapters/opentargets_adapter.py adapters/execution_registry.py adapters/pinned_evidence_adapter.py adapters/clinical_synthesis_adapter.py scripts/audit
python3 -m pytest -q benchmark/tests
python3 -m build --wheel . --outdir /tmp/agentic-core-dist
python3 scripts/audit/smoke_test_core_wheel.py --wheel-dir /tmp/agentic-core-dist
python3 scripts/audit/build_hf_release_package.py --output /tmp/agentic-hf-release-package --force
python3 scripts/audit/validate_hf_release_package.py --package /tmp/agentic-hf-release-package
git diff --check
python3 -m compileall agentic_drug_discovery adapters chains benchmark/src scripts/audit tests

Hub Placement

• Repository type: Dataset
• Repo id: jang1563/agentic-drug-discovery-system
• Current visibility: public and ungated
• Current public update: 0.3.0.dev2, published after explicit approval
• Candidate update: 0.3.0.dev3, not uploaded and pending exact-package approval

Source

Primary source repository:

https://github.com/jang1563/agentic-drug-discovery-system