This kit is the executable operating contract for
codeprobe.zero-code-access.operator.v1. A data-owner technical owner runs
CodeProbe inside an environment controlled by the data owner. Provider
Engineering and all other provider personnel remain outside that environment
and never receive repository, source, prompt, patch, trace, task-level result,
raw result, log, or diagnostic access.
The machine-readable invariants are in
kit-contract.json, and the export boundary is
EVIDENCE_BUNDLE.md.
| Owner | Required material |
|---|---|
| Data-owner technical owner | Data-owner runbook, intake and consent, sampling plan, experiment profile, and evidence request |
| Provider support | Support methodology, coordination guide, intake, scheduling, the intervention log, and bounded findings review |
| Data-owner security/platform | Optional follow-up only; no mandatory meeting or repository access |
| Provider Engineering | Internal defect repair only; any involvement in an external run disqualifies that run |
- Provider support sends the asynchronous intake. The data owner spends no more than ten minutes on it and omits repository-identifying or prohibited data.
- The data owner and provider support freeze the sampling plan and two-configuration profile before results exist.
- The data owner installs, configures, mines, validates, and dry-runs locally.
- Provider support coordinates one structured session capped at 45 minutes. The data owner operates every command and shares only coded status or sanitized, non-identifying observations.
- The data owner runs both configurations on the same task set with at least ten paired distinct scorable tasks and three repeats per task and configuration.
- The data owner aggregates and inspects all results locally, completes the local evidence request, previews the exact five artifacts, and either rejects or explicitly approves that exact preview.
- Only the approved five-file directory may leave the environment. Provider support reviews that bundle against the bounded findings template.
Stop the external proof and record a disqualifying intervention if any provider person accesses the environment or source, provides bespoke code, repairs evidence, receives prohibited data, or reinterprets raw results. A fix may be developed and verified internally, but the data owner must begin a new external run.
Return insufficient_evidence without weakening the gate when fewer than ten
paired tasks remain, repeats are incomplete, task sets differ, the sample
changed after results, representativeness is not attested, the comparison is
refused, or support was disqualifying.