INTERVENTION TIMING LAB

Not just “Can it stop?”
But “Is stopping still in time?”

Model a multi-step agent workflow, classify actions by reversibility, and reveal the point where a human correction stops being fully effective.

PLAN→ACT→INTERVENE→ROLL BACK→RECOVER

01 · WORKFLOW

Define execution steps

02 · INTERVENTION

Choose when correction arrives

Intervention point 50%

Before executionAfter completion

Human response latency 8

Immediate30 min

Checkpoint quality 78

PoorStrong

Cancel queued actionsPending stale work is invalidated Compensating actionsExternal effects can be counteracted Execution checkpointingSafe state is persisted between steps

CORRIGIBILITY HORIZON

72 /100

Wide intervention window

LAST EFFECTIVE MOMENT

76%

Late correction still matters Synthetic workflow position

03 · REVERSIBILITY MAP

Where intervention loses power

Idempotent Reversible Compensable Irreversible

Intervention

04 · INTERVENTION EFFECTIVENESS

Correction power over time

100 = fully outcome-changing

LAST FULLY REVERSIBLE 42%

REVERSIBILITY DEBT 29

LATENCY BUDGET 11 min

05 · CURRENT INTERVENTION

What happens if correction arrives now?

Effectiveness 74/100

Completed work 50%

Recoverable work 82%

Residual consequence 18%

EARLIEST CONFLICT ROLLBACK

DESIGN RECOMMENDATION

06 · HORIZON RECORD

Machine-readable JSON

How Corrigibility Horizon Mapper works Reversibility and intervention-timing heuristics ＋

Each workflow action is classified as idempotent, reversible, compensable or irreversible. Duration, external impact and checkpoint quality determine how quickly correction effectiveness declines as execution progresses.

The tool distinguishes behavioral compliance from causal effectiveness: an agent may accept a correction after an irreversible action, yet that correction can no longer fully undo the consequence.