Reliability Engineering

What if we had not done that step,
one action removed, replay the rest

Counterfactual Replay re-runs a closed incident with one action edited out. It rebuilds the timeline against the digital twin and predicts the outcome without that action. Use it during postmortems to test "did that step actually help?" Most teams find one or two recurring runbook steps that contribute nothing measurable.

Get Started Talk to Sales
app.novaaiops.com / counterfactual-replay
● LIVE
Per-action
counterfactual
Same twin
as Simulation Engine
Postmortem
evidence
Auto-suggest
runbook trims
How a Counterfactual Runs

Twin replays without one action

The replay starts with the captured incident state at start time (twin snapshot from the live archive). The engine plays back the timeline, skipping the action you marked for removal, and predicts the outcome at each subsequent step. The output is a paired comparison: actual outcome vs counterfactual outcome.

  • Captured incident state: starts with the digital-twin snapshot from the moment the incident opened
  • Skip one action: the replay plays every other step in order, predicting state without the skipped action
  • Paired comparison: output is "actual vs counterfactual", same metric, two values, decision is one click
app.novaaiops.com / counterfactual-replay · how
Postmortem Tooling

A "did this step matter?" check

During the postmortem, click any timeline step and run a counterfactual on it. The results land in the postmortem document directly so the conversation has data, not opinions. Most retros want to know whether one disputed step actually helped, counterfactual gives a measurable answer.

  • In-postmortem trigger: one click on any step opens the counterfactual; result lands in the postmortem doc
  • Measurable answer: paired SLI movement, not "I think it helped", anchored in twin replay
  • Stored on the postmortem: the counterfactual lives with the postmortem so future readers see the evidence
app.novaaiops.com / counterfactual-replay · postmortem
Runbook Trim Suggestions

When the same step is consistently neutral

When a runbook step shows neutral or negative impact across many incidents, the system flags it for trimming. The trim suggestion includes the evidence (the counterfactuals that motivated it) so the runbook owner can review with data. Trimming a step shrinks MTTR by the step's duration without losing recovery quality.

  • Cross-incident pattern: looks for steps that are consistently neutral across many incidents
  • Evidence packaged: the suggestion includes the counterfactual results that drove it
  • Reviewed by owner: final decision sits with the runbook's owner; the system never auto-trims
app.novaaiops.com / counterfactual-replay · trim
Caveats

A counterfactual is a prediction, not a proof

Counterfactuals are predictions from a twin model, not ground truth. They are best treated as evidence to weigh, not absolute answers. The page shows confidence intervals on each prediction so users see when the model is unsure. For high-stakes runbook trims, run the counterfactual in production canary instead, replace prediction with measurement.

  • Confidence intervals: every counterfactual reports an interval, not a point estimate
  • Weigh, do not blindly act: the page recommends treating counterfactuals as one input, not the verdict
  • Canary alternative: for high-stakes trims, run the new runbook in canary on the next incident as a follow-up
app.novaaiops.com / counterfactual-replay · caveats
Video walkthrough coming soon

Subscribe to Nova AI Ops on YouTube for demos, tutorials, and feature deep-dives.

Stop running runbook steps that do nothing

Counterfactual Replay is the evidence base for "we should remove that step." Anchored in data, not preference.

Get Started Request a Demo