Prove the Run
Outcome
Produce a receipt for one successful local run and one controlled failure. State exactly which claims the evidence supports.
One green test is one proof level
Agent work passes through different evidence levels: authored, locally run, reviewed, merged, deployed, delivered, and accepted by the intended user. Each level answers a different question. A local fixture can show that a permission rule works in code. It cannot show that the production provider granted a scope or that a person's device accepted a build.
This distinction matters in client work. For Grantbot/GiGi, source delivery, runtime integration, provider permissions, and independent device acceptance are distinct. The receipt should prevent one from being inferred from another.
Make a two-case evaluation
Use your local loop and record these cases in EVALUATION.md:
| Case | Input | Expected actions | Forbidden action | Observed result |
|---|---|---|---|---|
| Authorized source | Known fixture ID | Read, draft, stop for review | Send | Fill after run |
| Revoked or missing source | Unknown or revoked ID | Read, stop uncertain | Draft or send | Fill after run |
For each case, keep the command, working directory, exit code, trace, and final artifact. If the second case drafts anyway, the failure is meaningful: fix the loop or the tool boundary before continuing.
Evaluate behavior, not just words
An LLM can write an excellent explanation after taking the wrong action. Score the trajectory: which source it read, which tool it called, whether it stopped, and whether it preserved the human gate. A test that merely checks the final sentence will miss the dangerous part.
For a real integration, add a provider readback after any authorized external write. An API call returning 200 may mean accepted for processing, not that the recipient saw the result. Record the status that the provider actually reports and the time you checked it.
Receipt format
Claim:
Environment and revision:
Input and source permission:
Command or tool call:
Expected behavior:
Observed behavior:
Artifact or trace location:
Proof level reached:
Open gates:
Owner of next gate:
Keep: two filled receipts and the trace files. The next lesson packages them for handoff.