Every AI decision, with what it cost and what it read
User, component, workflow, model, prompt version, artifact version, cost and duration — recorded for every call, because the gateway is the only way out.
Append-only record for this session
| Time | Actor | Component | Model | Prompt | Artifact version | Cost | Duration | Decision |
|---|---|---|---|---|---|---|---|---|
| 13:42:47 | Demo QA Lead | Test Intelligence | Model Router | prompt:v7 | TS-0142 v12 | ₹12.18 | 1.6s | Generated 147 test cases, each citing a source artifact |
| 13:42:31 | Demo QA Lead | Test Intelligence | Model Router | prompt:v7 | TS-0142 v12 | ₹4.12 | 0.9s | Generated a risk-weighted strategy covering 82 scenarios |
| 13:42:12 | Demo QA Lead | Product Knowledge | Model Router | prompt:v7 | KN-0842 v1 | ₹9.84 | 1.2s | Extracted 37 requirements, 12 business rules, 8 risks |
| 13:41:58 | SALC Product Knowledge | Contradiction Detection | Model Router | prompt:v7 | KN-0842 v1 | ₹2.12 | 0.5s | C-014 surfaced — conflicting daily transaction limit. Not resolved. |
| 13:41:44 | Demo QA Lead | Source Ingestion | — | — | ART-001 v17 | — | 0.7s | Normalised 4 documents into source artifacts; originals retained |
What is structurally enforced
These are properties of the architecture rather than policies someone has to remember.
- Tenant-scoped data
Every read is scoped to organisation and project
- Versioned requirements
Requirements and specifications carry versions
- Immutable approvals
An approval points at a snapshot, not a mutable row
- Source citations
Every AI output cites the artifact it came from
- AI gateway enforcement
No component can reach a provider directly
- Audit trail
User, model, prompt, cost and duration on every call
- Untrusted-input handling
Uploaded documents are treated as untrusted
- Generated code isolation
Automation lands on a branch, never on main
No security certification is claimed. These are design properties of the platform being built, not the result of an external audit.
Machine-checked, not reviewed
Eight standards are checked by the development harness itself, so a violation fails a build rather than surviving a code review.
- 01AI-gateway-only provider calls
- 02Citations on every AI output
- 03Versioning of requirements and specifications
- 04Contradictions surfaced, never auto-resolved
- 05Never on main — generated code is isolated
- 06Untrusted-input defence in ingestion
- 07Audit record for every AI decision
- 08Tenancy on every data read