01 — REQUIREMENT
Plain-language summary of the requirement being reviewed.
FOUR CONTROLLED CALLS · RECORDED SEPTEMBER 8, 2026
A completed call.
What does it actually prove?
Inspect the requirement, read the evidence, and decide what is established. These are recorded calls to an AI recipient we controlled.
RECORDED OBSERVATION · MANUAL REVIEW
02 — TRANSCRIPT EXCERPT
Show additional context
03 — PROVIDER OUTCOME
Human-readable summary of the retained provider outcome, not raw JSON.
WHAT THE EVIDENCE SUPPORTS
WHAT REMAINS UNPROVEN
SEPARATE EVIDENCE TRACK · STATIC MODEL
Give uncertainty a place in the contract.
The appointment example only allowed a boolean. REDLINE proposes an explicit unknown outcome. This local example is separate from the four calls above; it is not a before/after runtime experiment.
ORIGINAL SCHEMA · EXCERPT
"confirmed": {
"type": "boolean"
}Two values cannot explicitly represent an unanswered question.
PROPOSED SCHEMA · EXCERPT
"confirmed": {
"type": "string",
"enum": ["yes", "no", "unknown"]
}Review this contract change: downstream consumers must handle a string instead of a boolean.
Inspect the related proposed instruction
Require an explicit yes or no. Do not treat a maybe, an I'll see or any other non-committal answer as agreement; if the answer is unclear, record unknown.
Excerpt of the saved proposal. Read the complete proposed task diff ↗
—Baseline modelled scenarios passed
—Corrected modelled scenarios passed
—Final benign cases passed
—Real calls in this verification
The verification evaluates a proposed correction in memory. It does not edit the source YAML, authenticate a real person, or test downstream applications. These are saved results, not a new execution in your browser.
REPRODUCE THE DEVELOPER WORKFLOW
Inspect. Review. Verify.
Use the pinned Python source and judge instructions. Local static evaluation needs no account or phone credits.
Open reproduction instructions ↗Boundaries you can inspect
- Transcript review was manual; audio was not independently assessed.
- No real booking was created. The recipient and appointment were fictional.
- The static model is shared by detection and remediation.
- AI onboarding reviews are documented. Human usability and production impact remain unverified.