91.4% of behavioral regressions produce passing outputs. 1,437 preregistered runs. Wilson CI [0.894, 0.932].">

Experiment · Paper 1 · Study 2

EXP-002 — Semi-blind perturbation

Under non-instructed perturbations across four domains, 91.4% of behavioral regressions produce passing outputs. Output-only evaluation does not make a marginal miss — it misses nearly all behavioral failures.

ID
EXP-002
Type
experiment
Status
CONFIRMED
Date
2026-04-28
Paper
Paper 1 · Study 2
Domain
refund · claim · memory · code
Evidence
1,437 runs · 4 domains · 6 conditions
Protocol
PAPER1_PROTOCOL_FREEZE.json#study-2
Corpus
OPBR-Bench v0
Result
θ_OPBR = 0.914 · Wilson 95% CI [0.894, 0.932] · n = 960 regressions across 1,437 runs.
Replication
abw run --experiment EXP-002
License
CC BY 4.0
Author
Adrian Sanchez de la Sierra, Genesis Rojas, and Radu Simonescu

Preregistered hypothesis

Under perturbations whose names do not appear in the agent's prompt, instructions, or tool descriptions (latency pressure, prompt compression, tool-semantics ambiguity, evidence noise), the proportion of behavioral regressions that produce field-level passing outputs (θ_OPBR) will be ≥ 0.80 with lower Wilson CI bound ≥ 0.75.

Method

Four domains (policy, claim, memory, code). Six conditions per domain (reference, latency-pressure, prompt-compression, tool-semantics-ambiguity, evidence-noise, safe-extra-check). Three cases per domain. 20 runs per cell. 1,437 runs total. The agent-visible prompts contain none of the forbidden patterns from the locked preregistration ("commit before verify", "fast lane", "behavior contract"). The agent is never told what failure mode is being tested. Behavior contracts and baselines are evaluated on the same traces.

Result

θ_OPBR = 0.914. Wilson 95% CI [0.894, 0.932]. Lower bound exceeds the preregistered 0.75 floor by 14.4 percentage points. Hypothesis confirmed.

Baseline ordering inverts relative to canonical OPBR (EXP-004): under semi-blind perturbation, rich-output provenance achieves the highest F1 = 0.890. The mechanism is correlated degradation: under latency pressure, an agent that skips the policy check also fills fewer provenance fields and produces sparser outputs. The rich-output detector catches the correlated output degradation, not the behavioral failure directly. This is a real signal under this regime, with the wrong mechanism. The inversion is the load-bearing finding for the baseline-selection discussion in the paper.

What this experiment does not establish

91.4% is the rate at which behavioral regressions pass output evaluation in this corpus under this perturbation set with this model and runtime. It is not a claim that all output-passing failures in production agents occur at this rate, nor that the four domains generalize to every consequential-action surface. The baseline ranking under this perturbation regime is regime-specific — see EXP-004 for the inverted regime on canonical OPBR.

Replication

abw run --experiment EXP-002 Requires Agent Behavior Workbench ≥ 0.4. Corpus auto-downloads from OPBR-Bench v0. Full protocol freeze at exp-002-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.

Cite as: Sanchez de la Sierra, A. (2026). EXP-002 — Semi-blind perturbation. Raising Agents Lab. https://raisingagents.is/lab/exp/exp-002

Related

Back to Lab