Use messick agent next as the control surface. Its versioned action has an absolute cwd, complete argv including --project-dir, mutation/spending/approval flags, prerequisites, an expected transition, known artifact IDs and paths, and a structured schema for every unresolved input.
Cognitive and behavioral plans always have a non-empty respondent/model execution design. A --models ModelList.ep selected through the current ep-agent workflow is required. Supply --agents AgentList.ep to override the bounded pilot default of three varied simulated respondent profiles. The plan records both hashes and the deterministic expected call count.
The agent must use the exact Jobs.ep path for inspect, cost, approval, execution, and subsequent Results ingestion. Completion requires relevant analyses, post-analysis adjudication, both report handoffs, and successful strict validation.
Set up Messick and help me complete an auditable pretest and revision of the
structured research instrument in this repository.
Install Messick and EDSL as an isolated uv tool, verify that messick and ep
resolve from that environment, and run messick doctor. Let ep own authentication;
never display or commit API keys.
Run messick agent guide, then messick agent next. Use agent next after every
material mutation and treat its JSON envelope as authoritative. Never edit
.messick/ directly or replace registered artifacts.
Messick builds Jobs.ep but does not execute inference. Use the exact returned
artifact for ep inspect, ep jobs cost, and ep run. Show me the model, cost, and
output path and obtain my approval before paid execution.
Keep simulated and human evidence separate. Never call simulated respondents
human participants, silently pool sources, transmit private human responses to
a model without approval, or claim that an instrument is globally valid.