Know which delivered changes were actually checked
Months later the question is not what the agent produced. It is which of it was checked, by what, and how thoroughly. A request count is not a record.
A pass flag is a number, not an answer
A pass flag says a suite went green. It does not say how much of the required behaviour that suite exercised. Two changes with identical flags can carry very different evidence.
That gap shows up when someone has to decide which changes to trust and has nothing to sort them by.
Evidence is graded, and provenance is recorded separately
An ordered scale
What each candidate demonstrated is recorded from none through syntax and runtime to partial and complete behavioral evidence.
Coverage kept apart
How much of the required behaviour was exercised is recorded separately from the grade.
Written content marked
Content that did not earn authorization carries no verification metadata, so it is never reported as checked.
What runs between the request and the answer
Proposal
The change your agent proposes enters the write path as candidate zero, the baseline.
Alternatives
Several candidates are generated instead of the first answer being accepted.
Execution
Each runs in an isolated sandbox against the project's available checks and the runtime's own oracles.
Guarded write
An alternative must earn authorization to replace the baseline. What did not earn it carries no verification metadata.
Failures return to step 02 as repair input rather than starting over
Proposal, alternatives, execution, guarded write, with repair feeding back.
Where the record is the thing being bought
Review triage
- The work
- Deciding where reading effort should go.
- What repeats
- Every change gets the same attention because nothing distinguishes them in advance.
- What changes
- Effort concentrates where evidence is weak or absent, because evidence strength is recorded.
After the fact questions
- The work
- Answering what was checked, months later.
- What repeats
- A request count survives. What was executed against the answer does not.
- What changes
- A per-stage record of what ran, how long it took, and what it produced, written locally.
Mixed provenance in one codebase
- The work
- Living with verified and unverified change side by side.
- What repeats
- Once they are indistinguishable, the safe assumption is that none of it was checked.
- What changes
- Delivered content is separable into what earned execution evidence and what did not.
Other solutions
Get more out of the models you are allowed to run
Built to raise the reliability of whichever model your policy permits, rather than asking you to change it.
Read the solutionRun AI coding where the data is not allowed to leave
Every part of the system inside your jurisdiction, with no vendor-operated component in the path.
Read the solutionTalk through your deployment constraints
The repository and the docs need no form. This one scopes a deployment.