Worked examples

See the controls in action.

Two small, runnable implementations connect architecture decisions to observable behavior. Each includes source, captured output, validation results, and lessons you can apply.

Permission-aware retrieval Release evidence

LAB-230-01 · Synthetic prototype

Private retrieval that respects document permissions.

The same finance question should produce different results for a permitted finance user and an HR user without access.

01Identify the user
02Filter by permissions
03Retrieve active documents
04Sanitize untrusted text
05Validate source citations
06Record versioned evidence

Permitted user · fiona.finance

Finance budget requests above 50000 credits require CFO approval and a recorded risk review.

Citation: FIN-BUDGET-2026 v3.0

User without access · harper.hr

No permitted source supports an answer to this question.

0 citations · 0 documents in context

6 tests passedPython 3.14.4Recorded 2026-09-07

Captured LAB-230-01 validation and sample output — Captured report view from the local synthetic lab run. Download the JSON below for the complete machine-readable output.

Inspect complete validation output

test_every_citation_resolves_to_an_active_exact_version (test_citations.CitationTests.test_every_citation_resolves_to_an_active_exact_version) ... ok
test_deleted_document_is_absent_from_index_context_and_citations (test_no_answer.NoAnswerAndDeletionTests.test_deleted_document_is_absent_from_index_context_and_citations) ... ok
test_unsupported_question_uses_approved_no_answer (test_no_answer.NoAnswerAndDeletionTests.test_unsupported_question_uses_approved_no_answer) ... ok
test_denied_finance_document_never_enters_hr_context (test_permissions.PermissionTests.test_denied_finance_document_never_enters_hr_context) ... ok
test_permitted_finance_user_receives_current_version (test_permissions.PermissionTests.test_permitted_finance_user_receives_current_version) ... ok
test_retrieved_instruction_is_detected_and_removed (test_prompt_injection.PromptInjectionTests.test_retrieved_instruction_is_detected_and_removed) ... ok

----------------------------------------------------------------------
Ran 6 tests in 0.005s

OK

Download runnable lab Download result JSON

Source fingerprint

SHA-256: 55c96fd2d12b02f70480cc72e044a35dd12a7d0b7a7f947ce4f8b1af216dbda4

Digest of sorted relative source paths and their contents in the downloadable lab.

Reproduce the result

Extract the source archive. With Python 3.11 or newer, open a terminal in the extracted LAB-230-01 folder. No third-party Python packages or model downloads are required.

python scripts/validate.py
python -m app.server

Validation writes evidence/latest-validation.json. The server exposes /query on port 8080; the included README shows the request format. Run times and timestamps will vary.

Implementation lesson

Filter documents before relevance scoring and context assembly. A citation should identify the exact active source version. The tests also check deleted documents, unsupported questions, and a small set of injected instruction patterns.

This lab uses lexical retrieval and deterministic answer assembly. It does not measure LLM answer quality, production latency, or comprehensive prompt-injection resistance.

Explore the private RAG architecture

LAB-250-01 · Synthetic prototype

A release gate that holds when evidence is incomplete.

A passing test suite can verify that a release gate correctly blocks a release. This example makes that distinction visible.

01Load control catalog
02Link evidence records
03Build traceability matrix
04Check required evidence
05Apply valid exceptions
06Issue release recommendation

Recorded recommendation

HOLD

1 uncovered finding: CTRL-INJECTION-01 lacks required integrity evidence. A valid exception covers the deletion-control gap, but the unresolved injection finding still blocks release.

16 tests passedPython 3.14.4Recorded 2026-09-07

Captured LAB-250-01 validation and sample output — Captured report view from the local synthetic lab run. Download the JSON below for the complete machine-readable output.

Inspect complete validation output

test_complete_evidence_control_has_no_finding (test_evidence_quality.EvidenceQualityTests.test_complete_evidence_control_has_no_finding) ... ok
test_evidence_missing_a_required_field_is_flagged (test_evidence_quality.EvidenceQualityTests.test_evidence_missing_a_required_field_is_flagged) ... ok
test_evidence_record_missing_required_field_is_not_silently_treated_as_complete (test_evidence_quality.EvidenceQualityTests.test_evidence_record_missing_required_field_is_not_silently_treated_as_complete) ... ok
test_every_evidence_record_references_a_known_control (test_passport_linkage.PassportLinkageTests.test_every_evidence_record_references_a_known_control) ... ok
test_every_exception_references_a_known_control (test_passport_linkage.PassportLinkageTests.test_every_exception_references_a_known_control) ... ok
test_every_referenced_control_exists_in_the_catalog (test_passport_linkage.PassportLinkageTests.test_every_referenced_control_exists_in_the_catalog) ... ok
test_passport_links_controls_by_id_rather_than_duplicating_their_text (test_passport_linkage.PassportLinkageTests.test_passport_links_controls_by_id_rather_than_duplicating_their_text) ... ok
test_covered_finding_yields_approve_with_conditions (test_release_gate.ReleaseGateTests.test_covered_finding_yields_approve_with_conditions) ... ok
test_expired_exception_yields_hold_and_is_reported (test_release_gate.ReleaseGateTests.test_expired_exception_yields_hold_and_is_reported) ... ok
test_missing_owner_yields_reject_even_with_complete_evidence (test_release_gate.ReleaseGateTests.test_missing_owner_yields_reject_even_with_complete_evidence) ... ok
test_no_findings_yields_approve (test_release_gate.ReleaseGateTests.test_no_findings_yields_approve) ... ok
test_shipped_dataset_holds_on_the_uncovered_injection_finding (test_release_gate.ReleaseGateTests.test_shipped_dataset_holds_on_the_uncovered_injection_finding) ... ok
test_uncovered_finding_yields_hold (test_release_gate.ReleaseGateTests.test_uncovered_finding_yields_hold) ... ok
test_control_without_evidence_is_marked_covered_not_silently_passed (test_traceability.TraceabilityTests.test_control_without_evidence_is_marked_covered_not_silently_passed) ... ok
test_every_catalog_control_appears_exactly_once_in_the_matrix (test_traceability.TraceabilityTests.test_every_catalog_control_appears_exactly_once_in_the_matrix) ... ok
test_every_control_has_a_named_owner_and_reviewer (test_traceability.TraceabilityTests.test_every_control_has_a_named_owner_and_reviewer) ... ok

----------------------------------------------------------------------
Ran 16 tests in 0.003s

OK

Download runnable lab Download result JSON

Source fingerprint

SHA-256: 881d0d1adecce814b28a03ecb5ecbf96ab3801aeb1cc96412515bee1f8b2fff3

Digest of sorted relative source paths and their contents in the downloadable lab.

Reproduce the result

Extract the source archive. With Python 3.11 or newer, open a terminal in the extracted LAB-250-01 folder. No third-party Python packages or model downloads are required.

python scripts/validate.py
python -m app.cli

Validation writes evidence/latest-validation.json. The CLI prints the traceability matrix, findings, exceptions, and release recommendation. Run times and timestamps will vary.

Implementation lesson

Keep controls, evidence, ownership, and exceptions traceable by ID. Count an exception only when it has the required compensating control and approval and has not expired. The tests cover approve, conditional approval, hold, and reject decisions.

The catalog and evidence are synthetic. These are local reproduction results, not an independent audit or a production approval. The sample uses the lab's configured assessment date; it does not assess a live system.

Start your own system passport

Your next step

Apply the pattern to your own boundary.

Use the starting-point guide to find a relevant solution, or bring your requirements to an architecture review.

Find my starting point Architecture review