A planned action is classified before execution, and the runtime may permit it, verify it, sandbox it, request approval or abstain. This moves enforcement from answer wording to concrete proposed effects. The paper simulates 20 seeds with 120,000 test actions per seed and reports harmful trajectories falling from 14.75% to 7.02% versus an…
Inspect the ClaimsCase · DiggingBeagle record
Consequence-aware tool gating reduced simulated harmful side-effect trajectories
CASEL selects different enforcement routes based on a tool call’s reversibility, expected harm, uncertainty and freshness rather than generic confidence alone.
Scope
These are simulated trajectories and assumed consequence models, not established savings or eliminated harms in deployed autonomous agents.
Assessment profile
Assessment method- Setting
- simulation
- Exploitation
- not established
- Evidence
- Grade C
- Remediation
- not applicable
- AI role
- AGAINST AI
Basis and provenance
The comparison reuses identical simulated action menus and compute costs across 20 seeds with 120,000 actions per seed. The action-risk model and intervention effectiveness are assumed, so the results rank policies rather than validate a production security boundary. Impact is left ungraded because a calibrated 0-to-4 impact basis for actual affected targets is unavailable from these Claims.
Assessed Oct 28, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-CP22-025-A · CLM-CP22-025-B · CLM-ENRICH-20261028-002 · CASEL: Consequence-Aware Selective Execution for Tool-Using Agents
At a glance
Calibration to real failure probabilities and independent enforcement of high-impact actions are still unverified.
Limits and uncertaintyFull account
A planned action is classified before execution, and the runtime may permit it, verify it, sandbox it, request approval or abstain. This moves enforcement from answer wording to concrete proposed effects.
The paper simulates 20 seeds with 120,000 test actions per seed and reports harmful trajectories falling from 14.75% to 7.02% versus an uncertainty-ranked gate. The authors report normalized mean harm moving from 0.184 to 0.123 under matched simulated action budgets. Primary evidence.
These are simulated trajectories and assumed consequence models, not established savings or eliminated harms in deployed autonomous agents.
Implement real tool-level gates with logged effect costs and separately validate the model’s risk scores against observed outcomes.
Claims & evidence
CLM-CP22-025-AThe paper simulates 20 seeds with 120,000 test actions per seed and reports harmful trajectories falling from 14.75% to 7.02% versus an uncertainty-ranked gate.supported
Basis: reported finding
- supportsCASEL: Consequence-Aware Selective Execution for Tool-Using Agentsprimary disclosure
Abstract; controlled intervention-menu comparison
CLM-CP22-025-BThe authors report normalized mean harm moving from 0.184 to 0.123 under matched simulated action budgets.supported
Basis: reported finding
- supportsCASEL: Consequence-Aware Selective Execution for Tool-Using Agentsprimary disclosure
Abstract; consequence and cost measurements
CLM-ENRICH-20261028-002A gate using estimated failure probability multiplied by consequence reached 5.82% harmful trajectories in the same simulation, but routed 45% more actions to human approval than CASEL.supported
Basis: reported finding
- supportsCASEL: Consequence-Aware Selective Execution for Tool-Using Agentsprimary disclosure
Abstract, cost-matched risk-product comparator
Controls and mitigations
- Implement real tool-level gates with logged effect costs and separately validate the model’s risk scores against observed outcomes.
Unknowns and contradictions
- Calibration to real failure probabilities and independent enforcement of high-impact actions are still unverified.
Sources and citation
Material revision history
- Oct 28, 2026 · Canonical change recorded · new in release · revision 199