STEM BIO-AI — yorkeccak/bio

Executive Summary

Tier Lock [CA-CAP]: This repository can be explored, but the report will stop at T2 until the project clearly says it is not for clinical or diagnostic use. Current ceiling: 69. Add an explicit non-diagnostic boundary to remove this cap.

Final Score

S1 Intent

S2 Repo

S3 Code/Bio

S4 Replication

TL;DR

Decision memo

This repository lands at T1 Quarantine with a final score of 48/100. The result is driven more by boundary, workflow-support, and governance weaknesses than by classic code-pattern failures.

Policy: defaultStatus: authoritative_releaseMode: mirror_only

Primary Risks

What pushed the review down

Clinical-adjacent surfaces exist without an explicit non-diagnostic/non-clinical boundary.
Self-asserted compliance or privacy-governance claim requires independent verification.
Legal, privacy, or compliance claim appears without supporting governance or security-grounding evidence in reviewed repository sources.
Core workflow appears materially dependent on named external service providers; local or self-host claims may overstate operational independence.
External operational dependency signal surfaced in code-integrity lane.

Positive Evidence

What still supports reviewability

Package metadata was available for repo-local consistency checks.

Freshness

When to re-check

Review after 45 days. Expires on 2026-07-26.

Change-triggered re-audit now: False

⚠ About This Score

How to read the number

Score reflects calculation integrity, not calibrated validity. Triage signal only.

What is verified: calculation integrity. The same input produces the same score.
What is not verified: calibrated measurement validity. Weights and detector scope remain bounded.
Use this score as a triage signal, not as certification, safety proof, or deployment approval.

Policy Boundary

How to read this artifact

Mirror-only policy surface: selected profile metadata is shown in this report, but authoritative scan scoring still follows deterministic runtime constants. Preview-only posture changes, including Stage 4 replication emphasis, do not change the formal score until a future read-through phase.

Classification applied: ca_severity=CA-INDIRECT
Score cap: 69
Tier lock: CA-CAP @69
Practical meaning: the report can cap the tier when the repository sounds clinically adjacent but does not state a clear non-clinical boundary.

Decision Path ?

Configured, Not Rewritten

Changing review posture does not require touching the score core

Use stem policy simulate with a governed profile file when you want to preview a different review posture. The authoritative score path stays deterministic; the profile is surfaced as metadata and preview-only interpretation.

If you only need the default posture, you do not need a profile file at all.

profile.json

{
  "profile_name": "strict_clinical_adjacency",
  "profile_read_mode": "mirror_only"
}

command

stem policy simulate /path/to/repo --profile-file profile.json

artifact note

Calibration Effect: mirror-only
Policy metadata surfaced
Formal score unchanged

Stage 1 — README Intent ?

Claim language, limitation posture, and clinical boundary wording.

R3_clinical_disclaimer-5
CA-INDIRECT surface lacks explicit non-clinical or non-diagnostic boundary.
S1_domain_readme+10
README exposes bio/medical domain vocabulary.
S1_domain_package+5
Package metadata exposes bio/medical domain vocabulary.
R2_regulatory_framework+5
Self-asserted privacy/compliance language detected without stronger regulatory-framework evidence.

Stage 2R — Repo Consistency ?

Internal contradictions between README, workflow claims, and support surfaces.

R2R_D2_missing_clinical_use_boundary-20
clinical_adjacent=True and explicit non-clinical boundary was not detected | tier impact: can contribute to clinical score ceilings and hard-floor review paths when stronger clinical deployment or certainty claims are also present
R2R_D4_unsupported_workflow_claim-15
workflow/demo/CLI claims detected while workflow, tests, or local support entrypoints are absent
R2R_1_readme_package_code_alignment+15
shared bio-domain terms detected across README and package metadata

Stage 3 — Code / Bio Responsibility ?

Engineering accountability, provenance, and reviewable responsibility surfaces. (raw: 20/80 -> normalized: 25)

B1_data_provenance_controls+15
S3_B1_dependency_manifest dependency or lock manifest presence plus data-source, dataset, or IRB language review
B3_coi_funding+5
S3_B3_coi_funding COI/funding/sponsor language review across README, docs, FUNDING, CITATION, and AUTHORS surfaces

Stage 4 — Replication ?

Reproducibility evidence is reported separately and does not alter the formal tier.

S4_container_environment+10
Container or compose file exists.
S4_environment_lock_evidence+10
Environment, dependency, or lock manifest detected.
S4_exact_dependency_pins_or_hashes+10
Exact dependency pin or hash evidence detected.

Code Integrity & Contract ?

Warnings First

Mapped risk lanes that fired

C2 Dependency Pinning

WARN

External operational dependency signal surfaced in code-integrity lane.

⌄

C5 Compliance Boundary Integrity

WARN

Unsupported legal/compliance claim surfaced in boundary-integrity lane.

⌄

C6 Mock Auth / Fail-Open Boundary

WARN

Mock-auth or auto-login boundary surfaced in code-integrity lane.

⌄

Clear Lanes

What stayed quiet in the current rule scope

C1 Hardcoded Credentials

PASS

No direct credential patterns detected by local CLI scan.

⌄

C3 Deprecated Patient Paths

PASS

No deprecated patient-adjacent metadata patterns detected.

⌄

C4 Fail-Open Exceptions

PASS

No executable fail-open exception handler detected.

⌄

CC1 Clinical Zero Default

PASS

count=0

⌄

CC2 API Contract

PASS

count=0

⌄

CC3 Shallow Validator

PASS

count=0

⌄

Why can Code Integrity contain PASS while the overall score is still low?

Because Code Integrity is a narrow detector family. The formal score is still driven mainly by Stage 1, Stage 2R, and Stage 3 evidence posture.

What changed in the C4 / C5 / C6 split?

C4 is now reserved for executable fail-open exception behavior, C5 for unsupported compliance or boundary integrity claims, and C6 for mock-auth or no-auth trust-boundary signals.

Regulatory Traceability ?

Regulatory basis note

Aligned to current official source classes as of June 2026: EU AI Act (Regulation (EU) 2024/1689); FDA QMSR, FDA ML-MD Transparency (2024), FDA PCCP (2025); ICH M15 MIDD (Step 4, Jan 2026); IMDRF SaMD Clinical Eval (2017), IMDRF GMLP (2025).

This is a traceability aid, not a compliance or clearance determination.

Stage 1 — README Intent

EU AI Act Art. 13 — Transparency⚠ Signal onlyeu_ai_act_2024_1689fda_mlmd_transparency_2024

Triggered by: R2_regulatory_framework

Not assessed: IFU completeness; deployer communication workflow

Boundary, intended-use, and limitation language is relevant to transparency scaffolding only.

ICH M15 §4.1 — Model Analysis Plan⚠ Signal onlyich_m15_midd_2026

Triggered by: R2_regulatory_framework

Not assessed: MAP completeness; formal MIDD planning stage documents

Regulatory framework language aligns with ICH M15 §4.1 MAP: pre-defined documentation of intended model analysis is a core MIDD planning requirement. Post-hoc alignment — not causally derived from M15.

Compliance Claim Grounding⚠ Signal onlyfda_qmsreu_ai_act_2024_1689

Triggered by: S1_R2_unsupported_legal_or_compliance_claim:README.md:262:001

Not assessed: formal compliance review; external certification status

Legal or compliance claims without supporting governance evidence are relevant to transparency and quality-system review, not compliance proof.

Stage 2R — Repo Consistency

IMDRF SaMD — Clinical Context Boundary⚠ Signal onlyimdrf_samd_clinical_eval_2017

Triggered by: R2R_D2_missing_clinical_use_boundary, R2R_D4_unsupported_workflow_claim

Not assessed: target-population performance; operational clinical workflow fit

Repository-local contradiction and boundary signals are relevant to clinical-context traceability, not clinical validation.

ICH M15 §2.1.2 — Context of Use~ Partially alignedich_m15_midd_2026

Triggered by: R2R_D2_missing_clinical_use_boundary

Not assessed: formal CoU document; regulatory submission completeness

Missing clinical-use boundary directly maps to ICH M15 §2.1.2 Context of Use: 'a concise, clear, and explicit description of the role and scope of the model' is required. Post-hoc alignment.

Stage 3 — Code / Bio Responsibility

EU AI Act Art. 10 — Data Governance⚠ Signal onlyeu_ai_act_2024_1689imdrf_samd_clinical_eval_2017

Triggered by: B1_data_provenance_controls

Not assessed: measurement correctness; dataset adequacy; regulator adequacy

Provenance and bias signals are relevant to data-governance review, but do not verify execution quality.

ICH M15 §3 — Verification⚠ Signal onlyich_m15_midd_2026

Triggered by: B1_data_provenance_controls

Not assessed: user-generated code documentation; MAR completeness; formal verification record

Data provenance signals align with ICH M15 §3 Verification: user-generated codes and data handling must be documented and available for review. Also aligns with §4.2 MAR data and methods section. Post-hoc alignment.

Stage 4 — Replication

EU AI Act Art. 12 — Record Keeping~ Partially alignedeu_ai_act_2024_1689fda_qmsrfda_pccp_2025

Triggered by: S4_environment_lock_evidence, S4_container_environment

Not assessed: deploy-time event logging; runtime event completeness

Reproducibility and trace manifests support record-keeping scaffolding, not operational logging completeness.

ICH M15 §4.3 — Code & Data Submission~ Partially alignedich_m15_midd_2026

Triggered by: S4_environment_lock_evidence, S4_container_environment

Not assessed: coding script completeness; dataset submission; formal regulatory submission

Reproducibility environment signals align with ICH M15 §4.3: 'all documents and files supporting submitted MIDD evidence, including data used in M&S analyses and relevant coding scripts, should be submitted or available for regulatory review.' Post-hoc alignment.

Structural signals partially align with traceability scaffolding. This remains a pre-audit traceability aid, not a compliance determination.

What does “signal only” mean here?

It means repository evidence is relevant to a governance framework, but this scan is not claiming alignment, compliance, or clearance.

What does this section actually help establish?

It shows where repository evidence connects to transparency, record-keeping, model-analysis-plan, or code-submission expectations before any formal audit.

MIT AI Risk Repository Risk Triggers ?V4_03 | airisk.mit.edu

Feature Explainer

What this section is doing

AIRI is used here as a bounded risk-vocabulary layer around deterministic repository findings. The report uses the curated runtime bundle, not the full upstream AIRI universe.

7 / 32 risks in detector scope

Bundle scope: curated_medical_clinical_subset

Snapshot: 2026-04-23 | License: MIT

Derived from The AI Risk Repository V4_03. Original source remains MIT-licensed and must be attributed in README, docs, runtime artifacts, and local registry metadata.

What does 7 / 32 mean?

It means seven AIRI risk IDs are currently reached by active local detector mappings, out of thirty-two AIRI risk IDs in the current detector scope.

What does “why mapped” mean?

Each covered AIRI row carries a bounded explanation built from the triggered detector, the local mapping justification, and the trigger reason surfaced by the scan.

What does AIRI not prove here?

AIRI does not independently verify harm, causality, clinical failure, or legal noncompliance. It is a risk-vocabulary layer around local findings.

Mapped, Not Guessed

AIRI rows light up through active detector mappings

The report does not infer AIRI coverage from prose alone. Coverage appears when a local detector fires and a governed mapping exists in the current AIRI runtime bundle.

trigger

C6_mock_auth_or_fail_open_boundary
status: detected

mapping

R2R_D5_single_external_service_dependency
→ 72.04.02 Market Concentration

report surface

covered_by: detector id
why: bounded mapping reason

Coverage Explorer

Covered and gap rows

AllAll Domains7/5

1Discrimination & Toxicity0/1

2Privacy & Security1/1

3Misinformation1/1

4Malicious Actors & Misuse0/0

5Human-Computer Interaction1/0

6Socioeconomic & Environmental1/1

7AI System Safety, Failures & Limitations3/1

Click a domain card to filter. Counts are shown as covered / gaps.

ID	Risk	Domain	Covered by / Note
24.01.03	Safe exploration problem with widely deployed AI assist	Lack of capability or robustness	primary=C5_compliance_boundary_integrity \| C5_compliance_boundary
24.04.01	Physical and Psychological Harms	Overreliance and unsafe use	primary=C2_dependency_pinning \| C2_dependency_pinning: External o
33.01.05	Privacy and security	Compromise of privacy / PII leak	primary=C2_dependency_pinning \| C2_dependency_pinning: External o
39.25.00	Verifiability	Lack of transparency or interpre	primary=S1_R2_unsupported_legal_or_compliance_claim \| S1_R2_unsup
60.02.01	Reliability issues	Lack of capability or robustness	primary=R2R_D5_single_external_service_dependency \| R2R_D5_single
69.01.00	False information	False or misleading information	primary=C5_compliance_boundary_integrity; also=S1_R2_unsupported_
72.04.02	Market Concentration and Infrastructure Dependencies:		primary=R2R_D5_single_external_service_dependency \| R2R_D5_single
65.03.03	Reidentification	2.1	CC-3 catches shallow validators; dedicated reidentify() API expos
70.02.02	Misinformation — hallucination of clinical knowledge	3.1	CC-1 catches threshold=0.0 default; actual output-level hallucina
39.25.00	Verifiability — black-box AI in medical healthcare	7.4	B2 detects surface language only; Model Card / interpretability a
11.02.00	Allocative Harms — withheld resources in healthcare	1.1	Subgroup performance disparities require dynamic evaluation; outs
72.04.02	Market Concentration — healthcare single-point failures	6.1	Systemic risk beyond single-repository scope.

Evidence Detail ?

All (94) FAIL WARN PASS INFO

Showing 57 compact rows from the first 94 evidence entries.

SEV	Detector	Finding	File
INFO	S1_clinical_boundary	No evidence detected for S1_clinical_boundary.	.
INFO	S1_H1_clinical_certainty_hype	No evidence detected for S1_H1_clinical_certainty_hype.	.
INFO	S1_H2_regulatory_approval_hype	No evidence detected for S1_H2_regulatory_approval_hype.	.
INFO	S1_H3_autonomous_replacement_hyp	No evidence detected for S1_H3_autonomous_replacement_hype.	.
INFO	S1_H4_breakthrough_marketing_hyp	No evidence detected for S1_H4_breakthrough_marketing_hype.	.
INFO	S1_H5_universal_generalization_h	No evidence detected for S1_H5_universal_generalization_hype.	.
INFO	S1_H6_perfect_accuracy_hype	No evidence detected for S1_H6_perfect_accuracy_hype.	.
INFO	S1_R1_limitations_section	No evidence detected for S1_R1_limitations_section.	.
INFO	S1_R2_regulatory_framework	No evidence detected for S1_R2_regulatory_framework.	.
INFO	S1_R4_demographic_bias_boundary	No evidence detected for S1_R4_demographic_bias_boundary.	.
INFO	S1_R5_reproducibility_provisions	No evidence detected for S1_R5_reproducibility_provisions.	.
INFO	S3_T1_workflow_files	No evidence detected for S3_T1_workflow_files.	.
INFO	S3_T2_domain_tests	No evidence detected for S3_T2_domain_tests.	.
INFO	S3_T3_changelog_release_hygiene	No evidence detected for S3_T3_changelog_release_hygiene.	.
INFO	S3_T3_changelog_bugfix_evidence	No evidence detected for S3_T3_changelog_bugfix_evidence.	.
INFO	S3_B2_bias_limitations	No evidence detected for S3_B2_bias_limitations.	.
INFO	S3_B2_measurement_evidence	No evidence detected for S3_B2_measurement_evidence.	.
INFO	C1_hardcoded_credentials	No hardcoded credential evidence detected for C1_hardcoded_credentials.	.
INFO	C2_dependency_pinning	No loose dependency evidence detected.	.
INFO	C3_dead_or_deprecated_patient_ad	No evidence detected for C3_dead_or_deprecated_patient_adjacent_paths.	.
INFO	C4_exception_handling_clinical_a	No fail-open exception handler detected in executable Python code.	.
INFO	BIO_smiles_surface_integrity	No malformed or suspicious SMILES-like strings detected by conservative surface checks.	.
INFO	BIO_smiles_rdkit_validation	RDKit optional validation lane not exercised because no SMILES-like candidates were detect	.
INFO	BIO_smiles_parser_guard	No missing None/invalid guards detected after SMILES parser calls.	.
INFO	BIO_silent_mock_fallback	No silent mock or simulated-data fallback patterns detected in production code paths.	.
INFO	BIO_trace_manifest	No traceability manifest or runtime audit-log schema surface detected.	.
INFO	BIO_run_trace	No risky subprocess or os.system bio-tool execution patterns detected.	.
INFO	S4_make_reproduce_target	No Makefile detected.	.
INFO	S4_readme_reproducibility_sectio	README exists but no reproducibility or replication section heading was detected.	.
INFO	S4_checksum_files	No evidence detected for S4_checksum_files.	.
INFO	S4_dataset_url	Documentation exists but no dataset URL or data source URL was detected.	.
INFO	S4_model_weight_url_or_checksum	Documentation exists but no model artifact URL/checksum evidence was detected.	.
INFO	S4_citation_cff	No evidence detected for S4_citation_cff.	.
INFO	S4_license_restriction	License/readme/docs surfaces exist but no restriction language was detected.	.
INFO	S4_cli_entrypoint	No package metadata or Python AST surface detected.	.
INFO	S4_seed_setting	No deterministic seed setting evidence detected by AST summary.	.
INFO	S4_runnable_examples	No evidence detected for S4_runnable_examples.	.
INFO	S1_readme_bio_terms	README exposes bio/medical vocabulary.	README.md
INFO	S1_R2_weak_regulatory_self_asser	Self-asserted privacy/compliance language detected without stronger regulatory-framework e	README.md
WARN	S1_R2_unsupported_legal_or_compl	Legal, privacy, or compliance claim detected without supporting governance or security-gro	README.md
INFO	S3_B1_dependency_manifest	Dependency or environment manifest exists.	package-lock.json
INFO	S3_B1_dependency_manifest	Dependency or environment manifest exists.	package.json
INFO	S3_B1_dependency_manifest	Dependency or environment manifest exists.	pnpm-lock.yaml
INFO	S3_B1_data_source_language	Data source, dataset citation, IRB, or provenance language detected.	README.md
INFO	S3_B3_coi_funding	COI, funding, sponsor, or acknowledgement language detected.	README.md
INFO	S2_package_bio_terms	Package metadata exposes bio/medical vocabulary.	package.json
WARN	R2R_D5_single_external_service_d	Required external service API key detected for a named workflow dependency. Aggregated 2 m	.env.example
WARN	R2R_D5_single_external_service_d	Named external service provider is presented as part of the core repository workflow. Aggr	README.md
WARN	R2R_D5_single_external_service_d	Local or self-host claims coexist with required external service dependencies. Aggregated	README.md
WARN	C6_mock_auth_or_fail_open_bounda	Mock-auth, auto-login, or no-auth local/self-host boundary surfaced in reviewed sources. A	README.md
WARN	C6_mock_auth_or_fail_open_bounda	Local or self-host claims coexist with mock-auth or auto-login boundary signals. Aggregate	.env.example
WARN	C6_mock_auth_or_fail_open_bounda	Local or self-host claims coexist with mock-auth or auto-login boundary signals. Aggregate	README.md
INFO	S4_container_environment	Container or compose file exists.	Dockerfile
INFO	S4_environment_lock_evidence	Environment, dependency, or lock manifest detected.	package-lock.json
INFO	S4_environment_lock_evidence	Environment, dependency, or lock manifest detected.	pnpm-lock.yaml
INFO	S4_exact_dependency_pins_or_hash	Lock manifest with exact dependency resolution detected.	package-lock.json
INFO	S4_exact_dependency_pins_or_hash	Lock manifest with exact dependency resolution detected.	pnpm-lock.yaml

Developer Follow-up

Decision Path

Which stages are holding the report down

Stage 2R: 40/100 — repo-local contradictions and missing trust boundaries.

R2R_D2_missing_clinical_use_boundary-20
clinical_adjacent=True and explicit non-clinical boundary was not detected | tier impact: can contribute to clinical score ceilings and hard-floor review paths when stronger clinical deployment or certainty claims are also present
R2R_D4_unsupported_workflow_claim-15
workflow/demo/CLI claims detected while workflow, tests, or local support entrypoints are absent
R2R_1_readme_package_code_alignment+15
shared bio-domain terms detected across README and package metadata

Stage 3: 25/100 — accountability and bio-governance evidence.

B1_data_provenance_controls+15
S3_B1_dependency_manifest dependency or lock manifest presence plus data-source, dataset, or IRB language review
B3_coi_funding+5
S3_B3_coi_funding COI/funding/sponsor language review across README, docs, FUNDING, CITATION, and AUTHORS surfaces

Top remediation actions

What to fix first

Clinical-adjacent surfaces exist without an explicit non-diagnostic/non-clinical boundary.
Add a clear research-use / non-diagnostic boundary in README and self-host surfaces first.
Self-asserted compliance or privacy-governance claim requires independent verification.
Either remove the claim or add supporting governance documentation and operational controls.
Legal, privacy, or compliance claim appears without supporting governance or security-grounding evidence in reviewed repository sources.
Separate marketing language from reviewed evidence and add concrete governance artifacts.
Core workflow appears materially dependent on named external service providers; local or self-host claims may overstate operational independence.
Document the external-service dependency explicitly and narrow self-host or privacy claims.

Expected impact

What changes the report meaning

Boundary and workflow contradictions usually matter more than adding more replication artifacts.
Removing a tier lock changes the meaning of the report first, and can unlock higher tiers later.
Use Code Integrity for detector lanes, and Evidence Detail for file-level proof before making code changes.

Where to inspect next

Use the developer-facing sections

Decision Path: use it to see which stage is holding the formal tier down.
Code Integrity details: use it for detector-specific trust boundary failures.
Evidence detail: use it when you need file-level proof before editing the repository.