Seven invariants that hold across every stage
These are the substance of the claim that SALC is not a wrapper. They cost real engineering time, and each one is visible somewhere in this demonstrator.
Context survives the workflow.
The value is not in any single generation step. It is in the connected, versioned context that persists between them — which is what makes a requirement change traceable forward and a generated test traceable back.
See it in the productAI output must be traceable.
Every AI output cites the source artifact it came from. Citations are stored as relationships to an artifact version and a locator, never as text copied into the output.
See it in the productContradictions are surfaced, not hidden.
A system that quietly picks an interpretation when the specification conflicts with itself produces output that looks authoritative and cannot be falsified. The contradiction is the finding.
See it in the productQuality is independently validated.
The Quality Principles Engine is deterministic and separate from generation. If a rule needed a model to decide, it would not be a quality principle — it would be more generation.
See it in the productHumans remain accountable.
Approval freezes an immutable version. Later edits create successors; they cannot retroactively change what was approved. An approval record that can be altered is not an approval record.
See it in the productGenerated code never goes directly to main.
Automation lands on an isolated branch and opens a pull request for an engineer. Generated code is a proposal, and nothing here claims it is production-safe.
See it in the productEvery AI decision is auditable.
User, component, version, model, prompt version, cost and duration — recorded for every call, because all model access passes through one gateway. Cost is treated as a correctness property.
See it in the productOne pass, seven moves
The sequence is the same whether the trigger is a document upload or a changed ticket.
- SOURCE
Documents arrive as untrusted input and are normalised
- CONTEXT
An immutable, cited knowledge snapshot is built
- REASON
Risk and coverage strategy derived from that snapshot
- GENERATE
Scenarios, cases and automation produced by models
- VERIFY
Deterministic rules accept or reject the output
- APPROVE
A person freezes a specific version
- ACT
Publish, generate automation, raise a pull request
Fast, cheap, and unaccountable on its own
The connective tissue that makes the output usable
Approves, reviews, and owns the decision
What is built, what is designed, what is vision
The full lifecycle is the long-term vision. The first release is deliberately narrower.
Converts changing product knowledge into traceable, reviewable and executable test coverage, with human-governed trust. Design-to-implementation conformance sits inside this as a component, not as a separate product.
Azure DevOps publishing, Playwright TypeScript generation and GitHub pull requests. Additive: each one registers against contracts that are already frozen.
Execution runners, defect intelligence, release readiness and the wider lifecycle — business discussion through to monitoring and continuous improvement.
“Autonomous” is aspirational and staged. Every generation step in the first release terminates at a human approval gate. Autonomy means the chain runs without hand-carrying artifacts between tools — not that a person is absent.