Mission

Objective

Run each validated pilot at least three times with one frozen agent-system configuration in disposable blind fixtures.

Why this matters

The repository has no direct agent capability observations, variance estimate, or complete run-cost record.

Scope

Included: validated pilots, frozen configuration, random run order, trajectories, final state, retries, runtime, cost, interventions, and failures. Excluded: comparative rankings and stable success-rate claims.

Existing context

research/frontier/records/RFR-002.md, NX-002, and Workstream A2 in the evaluation roadmap.

Initial hypotheses

Contained repeated runs will produce reconstructable task-level outcomes and variance useful for the next sample design. Infrastructure or task defects may instead invalidate the observations.

Required evidence

Frozen configuration, fixture hashes, run records, action telemetry, final-state diffs, grader results, cost/intervention logs, and invalid-run reasons.

Constraints

Only pilots promoted by MS-AIRES-2026-0001 may run. No production credentials, gold access, or network access unless separately authorized and contained.

Execution instructions

Predeclare stop rules and run order. Preserve every valid and invalid attempt.

Deliverables

Immutable run dataset, exploratory variance summary, failure updates, and a baseline decision.

Success criteria

At least three valid repetitions per promoted pilot with complete metadata and reconstructable outcomes.

Stop conditions

Stop at six valid trials or immediately on containment, telemetry, task, or grader failure that invalidates further runs.

Handoff requirements

Provide the natural outcomes needed by MS-AIRES-2026-0003 and the observed variance needed for later sampling.