Alex Open Research Wiki

Counterfactual PoNR Auditing

2 min read

Counterfactual PoNR Auditing

Summary

Counterfactual PoNR Auditing (CPA) is an offline simulator-auditing method that estimates the Point of No Return: the latest replayable snapshot from which a specified intervention class can still keep a constrained trajectory feasible over a finite horizon.

replayable snapshot
  + audited intervention class
  + constraint contract
  + finite horizon
  -> counterfactual rollout
  -> recoverable / irrecoverable
  -> latest recoverable snapshot (PoNR)

The feasible-but-irrecoverable interval between PoNR and observed violation is called the Zombie Zone in the source paper. Durable cross-domain use should prefer the descriptive term feasible-but-irrecoverable interval.

Interface

For belief/state snapshot , intervention protocol , horizon , and risk tolerance , CPA evaluates whether future constraints remain satisfied under . It then searches the snapshot timeline for the latest positive recoverability label.

Each branch becomes an Evidence Object: a versioned record of the restart snapshot, intervention, counterfactual trajectory, slack path, action/event stream, recoverability outcome, failure time, audit cost, simulator/policy identity, and seed.

Evidence Boundary

The introducing KDD 2026 paper reports strong within-protocol monotonicity and bisection agreement on a hydrodynamic testbed, simulator-mined early-warning labels, snapshot-matched failure motifs, and a small damped-bistable-system validation. The result is credible tier-1 evidence for the audit interface, but not a universal reachability guarantee or production-ready monitor.

PoNR is conditional on the intervention family, actuator capacity and delay, constraint definition, audit horizon, tolerance, disturbance model, simulator fidelity, observation resolution, and available state. It MUST NOT be described as an intrinsic scalar property of an observation alone.

Role In The Wiki

CPA adds a recoverability-centric evaluation object for action-conditioned time-series world models:

  • forecast models ask whether violation is likely;
  • world models compare futures under candidate actions or control inputs;
  • CPA asks whether any admissible recovery branch remains and when that option disappears;
  • distilled monitors approximate the expensive offline labels at deployment time.

This is directly relevant to Counterfactual Safe Thermal Control, Kubernetes OTEL Control Gym, and the Foundation Time-Series Model Research Agenda.

Graph View

Backlinks