Experiment · Paper 1 · Study 1
EXP-001 — Controlled mechanism
Under controlled fast-lane conditions across three Claude Code domains, output evaluation passes 180/180 induced regression runs while behavior contracts fail. Drift paths are also operationally clean — output + operational composite still misses all 180. Only the behavior layer separates them.
Preregistered hypothesis
Under explicitly induced fast-lane conditions (the agent is instructed in a way that increases the probability of committing a consequential action before completing a preregistered precondition), field-level output evaluation will pass at least 95% of induced runs that the corresponding behavior contract fails. The output+operational composite will also pass at least 80% of drift-path runs that the contract fails.
Method
Three Claude Code domains (policy exception, qualified claim, memory revision). Three cases per domain. Six conditions per domain: baseline, candidate, controlled, repaired, safe-variation, pressure. 20 runs per cell. 1,080 runs total. Per PAPER1_PROTOCOL_FREEZE.json. Output evaluation: field-level correctness checks against the gold output schema. Operational evaluation: latency, retry count, error count. Composite: output + operational. Behavior contracts: typed checks on consequential tool calls against preregistered preconditions.
Result
Induced fast-lane condition: 180/180 runs pass output evaluation. 0/180 pass the behavior contract. Output and behavior layer disagree at the maximum possible rate.
Drift paths (a subset where the agent skips a precondition due to natural variation, not induced pressure): 60/60 pass output + operational composite. Behavior contract fails on the same 60. The operational layer does not catch the failure mode because skipping the check is not operationally anomalous — it does not produce latency spikes, retries, or errors. The trace is fast, clean, and wrong.
What this experiment does not establish
This is the most controlled regime: induced conditions, three domains, one model, one runtime. It establishes that the failure mode is real and detectable, and that output+operational composite is structurally insufficient to detect it. It does not yet establish the field rate of this failure under uninstructed perturbation — that is the job of EXP-002 — and it does not test the four-domain canonical OPBR distribution, which is the job of EXP-004.
Replication
abw run --experiment EXP-001
Requires Agent Behavior Workbench ≥ 0.4. Corpus auto-downloads from OPBR-Bench v0. Full protocol freeze at exp-001-protocol.json.
Artifacts
Every load-bearing claim on this page traces back to one of the artifacts below. They are the canonical citation targets — not the prose.
exp-001-protocol.json— frozen preregistration: hypothesis, falsifier, method, detectors, primary metric, preregistered gates, claim boundary.exp-001-results.json— machine-readable result block: primary metric, gate-pass record, mechanism interpretation.- Agent Behavior Workbench — open-source code and OPBR-Bench v0 corpus. Required to run
abw run --experiment EXP-001. - Paper 1 · Study 1 — Synthesis layer that bundles this with EXP-002, EXP-003, EXP-004, EXP-005 into the venue submission.
Cite as: Sanchez de la Sierra, A. (2026). EXP-001 — Controlled mechanism. Raising Agents Lab. https://raisingagents.is/lab/exp/exp-001
Related
- Paper 1: Output-Passing Behavioral Regressions · Study 1
- Field rate: EXP-002 — Semi-blind perturbation
- Held-out: EXP-004 — Held-out repair
- Spec audit of these contracts: EXP-005 — Completeness analyzer