# API pilot run plan recorded before provider trials

The three fixture packs and expected answers are unchanged. Nine attempts are planned: three supplier-quote trials, three email-routing trials, then three meeting-action trials. Each request uses the dashboard's fixed Grok 4.3 configuration and may reserve US$0.10 within the approved US$10 monthly OpenRouter cap. Preserve unsuccessful attempts; do not silently replace them.

Protocol deviation declared before the first request: no timed human baseline is currently available. Run the model accuracy/cost condition now; keep the investigations provisional until answer review and required human work are complete. Do not claim time savings or end-to-end workflow speed. The user has been asked whether they will supply a timed baseline. If they decline, explicitly label the work as an accuracy/cost study without a human comparison; do not retrospectively invent baseline data or change the scoring criteria.

The existing quote arithmetic cross-check is recorded in the prior preparation. The assistant re-read every email and meeting case and its authored expected answer against the explicit policy before this run; no discrepancy was found. This is an assistant review, not independent human validation. Neither answer keys nor scoring code are included in model prompts.

Evidence will come from authoritative D1 model run records and spending records, exported without credentials. Keep exact prompts, settings, model identity, output, outcome, request duration, provider cost and unresolved reservations. Score with the existing frozen scorer. The dashboard's model-request duration excludes setup, review, correction and delivery time. Synthetic success cannot establish reliability on customer documents, demand, willingness to pay or profit.

## Infrastructure amendment after attempt 1, before attempt 2

Supplier attempt `150053e6-437d-4621-9fef-abacc078c14b` produced no result and retained a US$0.10 reservation. A local test in the same Cloudflare runtime compatibility profile reproduced rejection of `redirect: error` before an outbound request. The deployed provider requests now use `redirect: manual`, rejecting non-success statuses and never following redirects. The read-only OpenRouter key check subsequently succeeded. The original record is unchanged and its invoice cost remains unconfirmed; do not retroactively replace its missing diagnostic or claim an observed provider status.

Continue the originally planned nine attempts, including the failed first attempt. Prompt, model and model parameters are unchanged. Distinguish the infrastructure failure from model answer accuracy. This leaves at most two recoverable quote outputs within the original three-attempt series; report that limitation rather than silently adding a replacement trial. Any later supplementary trial must be separately labelled and retained.