Skip to content

real-world-examples.md

Latest commit

 

History

History
174 lines (133 loc) · 7.56 KB

real-world-examples.md

File metadata and controls

174 lines (133 loc) · 7.56 KB
Raw

Real-world examples

Five realistic industry scenarios. Each is generic by design: every regulated organisation sees their own situation in one of them. No named customers, no proprietary details. Every example maps to runnable code in examples/ or examples/policies/.

1. Defence and aerospace — autonomous go / no-go

The situation. An autonomous system evaluates mission parameters and produces a go / no-go decision. A human operator must be able to halt all decisions at any time (EU AI Act Art. 14). Every decision must be reproducible years after the fact for accident investigation and formal audit. The deployment is air-gapped.

What Sentinel provides.

  • FilesystemStorage writes decision records to NDJSON on local disk. No network. Works under VS-NfD / SECRET classification constraints.
  • SimpleRuleEvaluator (or LocalRegoEvaluator for Rego policies) runs safety checks in-process before the mission logic executes.
  • sentinel.engage_kill_switch("operator halted — see incident log") halts all decisions without restart. Every blocked attempt is recorded as a DENY trace with a HumanOverride entry.
  • RuntimeScanner proves there are no US-owned packages in the critical path — CI-enforced on every PR.

Runnable example. examples/05_kill_switch.py and examples/06_filesystem_storage.py.

Reference policy. examples/policies/mission_safety.rego.

2. Healthcare — treatment recommendation with escalation

The situation. An AI system recommends a treatment plan. For low-risk cases the recommendation goes to the treating clinician directly; for high-risk cases a second opinion must be obtained before the recommendation is presented. Patient data must remain on-premise (GDPR + national healthcare data laws). Every recommendation must be auditable and linked to the clinician who actioned it.

What Sentinel provides.

  • DataResidency.EU_DE (or equivalent) on every trace — machine evidence of where the data was processed.
  • Rego policy evaluates risk_score, patient_age, and previous_adverse_events and routes low-risk to ALLOW, medium-risk to EXCEPTION_REQUIRED (escalate), high-risk to DENY (never bypass the human).
  • Clinician sign-off is recorded as a HumanOverride linked to the original trace by parent_trace_id — the decision record is append-only.
  • inputs_hash is SHA-256 by default; raw PII never appears in the trace unless explicitly opted in. Regulator asks "was patient X's data processed?" and the answer is reproducible without exposing the data itself.

Runnable example. examples/03_policy_simple_rule.py demonstrates the escalation pattern.

Reference policy. examples/policies/medical_escalation.rego.

3. Financial services — transaction approval with velocity checks

The situation. An automated system approves (or rejects) individual financial transactions against fraud rules, sanctions lists, and per-customer velocity limits. Regulatory regimes such as DORA require a tamper-resistant incident and decision record retained for several years. Any third-party processor with US jurisdiction creates CLOUD Act exposure on the evidence record.

What Sentinel provides.

  • PostgresStorage (append-only) for the decision record. Deployed on-premise or in an EU-sovereign database service.
  • Policy evaluates amount, velocity_last_hour, sanctions_score, and customer_risk_tier. Every DENY is recorded with the name of the triggering rule.
  • OTelExporter streams span attributes sentinel.sovereign_scope and sentinel.data_residency to the existing APM stack — the same dashboards the SRE team already uses now show sovereignty posture alongside latency.
  • EUAIActChecker runs in CI and prevents regressions: if anyone removes the kill switch or disables the policy evaluator, the compliance check job fails the PR.

Runnable example. examples/13_full_pipeline.py.

Reference policy. examples/policies/financial_transaction.rego.

4. Public administration — decisions with transparency obligations

The situation. A public body uses an AI system to triage citizen requests. Law requires that any individual can ask for the reasoning behind a decision about them, and that the system operator can identify the policy version in force at the time. Procurement rules forbid single-vendor lock-in.

What Sentinel provides.

  • Every trace records agent, agent_version, model, model_version, policy, policy_version, policy_result, policy_rule. A human-readable reconstruction is a single query.
  • StorageBackend is an interface with three ready-made implementations (SQLite, PostgreSQL, filesystem NDJSON) and any new backend is 4 methods. No lock-in.
  • NDJSON export gives a portable archive format that survives any future migration.
  • Apache 2.0, permanently. No CLA, no enterprise edition, no relicensing path.

Runnable example. examples/07_postgresql_storage.py (on-premise deployment) and examples/06_filesystem_storage.py (NDJSON archive).

Reference policy. examples/policies/access_control.rego.

5. Enterprise procurement — high-volume approval with delegation

The situation. A large organisation runs thousands of procurement approvals per day. Approval authority is delegated by amount, cost centre, and approver level. Every override by a human must be recorded with the justification. The finance team needs a monthly report showing who approved what and under which policy.

What Sentinel provides.

  • @sentinel.trace wraps the approval function. Zero changes to the existing approval logic.
  • SimpleRuleEvaluator encodes the delegation matrix as Python.
  • Human overrides of a DENY become new linked traces — the original DENY is never mutated. An auditor can see both the automated decision and the documented human reasoning.
  • sentinel.query(policy_result=PolicyResult.DENY, limit=10000) is the monthly report. Exported to CSV or NDJSON for finance tooling.

Runnable example. examples/03_policy_simple_rule.py shows the pattern with sample data. Scale up by wiring to your real approval service.

Reference policy. examples/policies/procurement_approval.rego.

How to adapt these

Each scenario maps to real Sentinel primitives that already exist:

Primitive Where it lives
@sentinel.trace sentinel.core.tracer.Sentinel.trace
SimpleRuleEvaluator sentinel.policy.evaluator
LocalRegoEvaluator sentinel.policy.evaluator (needs OPA binary)
FilesystemStorage sentinel.storage.filesystem
PostgresStorage sentinel.storage.postgres (optional extra)
SentinelCallbackHandler sentinel.integrations.langchain (optional extra)
OTelExporter sentinel.integrations.otel (optional extra)
EUAIActChecker sentinel.compliance.euaiact
SentinelManifesto sentinel.manifesto

There is no code in any of these scenarios that isn't shipped in the public repository. Clone, read, adapt.