The Delivery Record
Evidence, not a comment thread.
Anyone can hand you a document and anyone can hand you a repository. What neither category produces is the unbroken line between them — and a record is only worth something if the system writing it is unable to flatter itself.
The chain, end to end
Rendered as a traceability workbook with live formulas, exportable the day someone asks. A test fails the build if a ticket cites a requirement ID that no document defines, because a fabricated link is worse than an honest gap.
- Business requirementwith an ID every later artifact cites
- Architecture decisionand the critique it survived
- Technical requirementtraced back to the business need
- Ticketin the approved implementation plan
- Change setthe actual files, in your repository
- Test resulta real exit code from a real run
- Security findingseverity, and whether it blocked
- Deployment stepwhat was promoted, and where
- Approvalwho said yes, to what, and when
Approvals
A gate that suspends the run
Approvals are signals to a durable workflow, so the run is genuinely paused rather than prompted at while it carries on. Rejections are structured rather than free text — a verdict is pass, changes requested or blocked, and a finding is a blocker, major, minor or nit — because the gates branch on severity. That is what makes the approval log a dataset instead of a conversation.
Five approvals can never be switched off
Risk is classified in one policy rather than hardcoded per call site, so these stay mandatory on every plan and in every project.
- Production deploy
- Production schema change
- Secret access
- External communication
- Infrastructure deletion
Why it holds
A system that can flatter itself is not evidence.
These are enforced in code rather than promised in marketing, which is the only form of this claim worth making.
- A run that ran out of money is never green
Work is marked complete only when the verification commands actually executed and actually exited zero. Hitting a budget ceiling or a turn limit is recorded as a failure, because that is what it is. The loop re-runs any check an agent skipped rather than taking its word for it.
- The documents are real files, and they come back
Every artifact renders as Word, PDF or Excel — cover page, table of contents, page numbers, live formulas in the traceability workbook. Edit the Word document by hand and the agents read your version back and continue from it. Prose you rewrote is marked so it is used instead of theirs.
- Agents cannot reach past the project
Coding runs inside a per-project workspace where a path that resolves outside it is refused and only nine executables can start at all. There is no shell to inject into and no network tools. That workspace sits inside a locked-down container with a hard spend ceiling enforced from within.
- When we substitute, we say so
If the requested coding runtime is unavailable and the work falls back to another, the reason travels into the change set the reviewer reads. Nobody is ever handed template output believing an agent wrote it.
- Retrieval will not invent a number
Company context is retrieved lexically, or with real embeddings when they are configured, and there is no third mode. No placeholder vector is ever generated, because downstream it would be indistinguishable from a real one while ranking on noise.
- You set the ceiling, and you can see the meter
Agent spend runs against your own key or is passed through at cost, so we earn nothing when a run gets longer. The cost is live on screen while the work happens and the limit is yours to move mid-run.