# Dispatch reconciliation study: fixed plan

Declared before execution on 2026-09-08 (Europe/London). The manifest SHA-256 is 53542810ac5f8f127fc3518b2e64f83dff0d6c94bd4c8a70be1309af9e194823.

Run four pairs (eight trials), within-pair randomized order with seed 20260908. Request gpt-6-astra, low reasoning, through installed Codex CLI 0.153.4 with isolated configuration. Each agent has 100 seconds; the complete run has 900 seconds. No adaptive extension or retry selection. Retain failed, timed-out and unrun trials in the result. The task and synthetic sources are in ../examples/dispatch-reconciliation. The only condition difference is AGENTS.md: generic instruction versus a canonical-source/join-key index.

Primary outcome: complete independently verified reconciliation across the ledger, exception list, and dispatch plan. Frozen hand-reconciled expected.json and the verifier are outside both surfaces. One known-pass and three known-fail controls must match before trials begin. Per-deliverable checks are diagnostic. Tokens, command exits and elapsed time are descriptive; there is no preregistered efficiency significance test.

Use the framework's paired success effect and conservative fixed-sample 95% bound, with its assumptions and small-sample limitations. Four pairs cannot establish a broad beneficial effect with this bound; this is an early workflow feasibility study. A ceiling, error or inconclusive result will be published as such. No comparison with the earlier single-lookup study is planned.

This run uses ordinary local execution on macOS and existing host authentication. The read-only workspace and isolated CLI configuration are not an externally attested access boundary. Complete client/provider context, reads outside the workspace, and agent awareness are unobserved; the agent is not guaranteed unable to access the host-side oracle. No causal generalization or external-observer qualification is claimed. Raw traces stay private under .evx; inspect before publishing any derived record.