4 input sources — each with a trust weight
Code evidence outweighs documentation. Documentation outweighs manifest declarations. A declaration without a code signal triggers a contradiction finding.
Evidence trust weight by source
Source code
AST-level parsing across 15 languages. Detects AI library imports, human oversight patterns, logging infrastructure, kill-switch mechanisms.
Example signals
import torch / from openai importhuman_review_trigger()logger = logging.getLogger()risk_mitigation_applied()Config & build files
package.json, requirements.txt, Cargo.toml, Dockerfile, CI/CD YAML — cross-referenced against 89 tracked AI packages.
Example signals
tensorflow==2.15.0openai in dependenciesface-api.js detectedaws-rekognition presentDocumentation & policies
Markdown, PDF, DOCX files scored for substance — keyword density, word count, boilerplate detection. <40 words penalised.
Example signals
risk_assessment.mdMODEL_CARD.mdOVERSIGHT_PROTOCOL.mddata_governance_policy.mdManifest declarations
sentinel.manifest.json — lowest evidence weight. Every declared flag is cross-checked against code. Contradiction = HIGH finding.
Example signals
risk_category: HIGH_RISKhuman_oversight_enabled: truefria_completed: trueentity_role: provider18-phase forensic pipeline
Four sequential groups. Fully deterministic — the same repository state always produces the same output. Every phase writes to an evidence registry included in the final signed bundle.
Code pattern → legal obligation
Every signal detected in your repository is mapped to a specific article — defined in probing-rules.json and locked to a rule ID in the authoritative namespace.
ARTICLE_WEIGHTED_V3 — how the score is computed
Normalised against applicable weight only. Minimal-risk systems are never penalised for high-risk obligations. The final verdict is determined by four independent tracks, not just the number.
Article scores — sample audit (TalentScreen Pro v3.1)
4 verdict tracks — any can override the numeric score
Required documents must exist and exceed quality threshold
Code-level signals for critical articles verified independently
Hard floor per article — no article may fall below minimum
Strips comments — comment-only compliance is rejected
7 outputs. Every audit. Cryptographically signed.
Machine-readable, human-readable, legally formatted, DevOps-ready, and verifiable by any third party.
audit.jsonFull machine-readable report. RFC8785-Lite canonical.
report.htmlStandalone visual report with score, breakdown, dossier.
SARIF v2.1.0Native GitHub Security tab, VS Code, any CI/CD pipeline.
SBOM CycloneDXAI component inventory for supply chain compliance.
Annex IV packPre-filled Annex IV templates for high-risk placement.
checksums.txtSHA-256 hashes — enables independent bundle verification.
RSA-PSS signatureCryptographic proof verifiable without access to Sentinel.
Cryptographic reproducibility: RFC8785-Lite canonical JSON + RSA-PSS signature + SHA-256 hashes. The same repository state always produces an identical audit.json — byte-for-byte. Verifiable by any third party, including a Notified Body, without access to Sentinel.
What Sentinel covers — and what it doesn't
Maps 33+ technical signals from source code to EU AI Act article requirements
Detects contradictions between manifest declarations and actual code evidence
Records absence of expected signals as negative evidence — not neutral silence
Produces cryptographically signed, reproducible audit bundles (RFC8785 + RSA-PSS)
Generates Annex IV–compliant documentation and gap remediation templates
Tracks compliance delta over time with automatic regression detection
Applies sector multipliers and hard floors for high-risk sectors
Execute any code — static analysis only, zero side effects on your system
Evaluate runtime behaviour, model accuracy, fairness outcomes, or live API calls
Replace a Notified Body or issue CE certification under EU AI Act
Verify generic HTTP calls made without recognised AI SDK patterns
Provide legal advice — all findings carry an explicit legal disclaimer
Legal notice: Sentinel findings constitute technical evidence, not legal advice. A score of 100/100 does not constitute CE certification or replace evaluation by an accredited Notified Body under Regulation (EU) 2024/1689. Enforcement date: 2 August 2026.