Actionability-Filter After-Action Review¶
Retrospective protocol — instantiates Regret-Signal Calibration
Runs a post-outcome review that turns a regret into a durable rule change only when the lesson is both controllable and recurring — otherwise it routes the regret to closure.
Most regrets should change nothing. The Actionability-Filter After-Action Review exists to say so out loud: it takes the pile of "we should have..." that any disappointing outcome throws off and passes each one through a single gate — is this lesson controllable and would it recur across a class of future decisions? Only regrets that clear both earn a codified rule change. Everything else is named and closed: bad luck, a genuine one-off, or noise. Its defining discipline is subtraction, not extraction. Where a naive retrospective tries to squeeze a lesson out of every wound, this one is built to prevent a team from over-learning from single vivid failures and scarring its process with rules that only ever mattered once.
Example¶
An SRE team has had a brutal month — five painful production incidents. The blameless review has already run, so each regret arrives with a controllability verdict attached. The filter sorts them. A config push that went out with no canary stage caused two of the outages, was clearly within the team's control, and will recur every deploy — both boxes checked, so it earns a durable runbook change (mandatory canary). An upstream cloud-region failure took down a third incident; not controllable, so no rule, logged and accepted. A bizarre data-corruption bug, controllable in principle but seen once in three years, is a genuine one-off — recorded, but the runbook is left alone. The month that generated eleven candidate "lessons" yields exactly two rule changes, and the other nine are closed explicitly rather than left to haunt the next planning meeting.
How it works¶
The review assumes two upstream inputs are already in hand — the outcome-reference pair and, for each candidate regret, a controllability verdict from the blameless review. It does not re-derive them. Its own work is a two-question gate applied to each regret in turn: would the same lesson have improved a whole class of past-and-future decisions (the generalization / recurrence test) and is it in the class already marked controllable. Regrets that pass both are written up as a rule, threshold, checklist item, or option-generation change. Regrets that fail are classified — acceptance, one-off log, or noise — and formally closed. The overfitting check is the heart of it: a lesson that would only ever have improved this exact outcome, after the fact, is treated as a false positive.
Tuning parameters¶
- Recurrence bar — how broadly a lesson must generalize to earn a rule. Set it high and you avoid process bloat but risk dismissing rare-yet-important lessons; set it low and every stubbed toe becomes policy.
- Rule budget — a cap on how many durable changes one review may emit. Forces prioritization and starves scar-tissue accumulation, at the cost of deferring real fixes.
- Severity override — whether a single catastrophic-but-non-recurring regret can bypass the recurrence bar. Protects against the filter's own blind spot.
- Closure explicitness — whether non-actionable regrets are merely dropped or named-and-closed. Naming them is what actually quiets the replay.
When it helps, and when it misleads¶
Its strength is discipline against regret overfitting — the tendency to rewrite the playbook after every bad day until the process is a museum of one-time disasters. By demanding recurrence and controllability, it keeps the update loop connected to lessons that will actually pay forward.
Its failure mode is the mirror image: a recurrence bar set so high that a rare but serious risk is waved through as "won't happen again," and the one lesson that should have changed a rule never does. A classic misuse is weaponizing "non-recurring" as a permanent excuse to never adapt. The guarding discipline is the severity override plus an honest read of the base rate — "once in three years" across a fleet of hundreds of services is not actually rare. The format borrows its bones from the military after-action review,[n1] but the filter is what separates it from a generic retrospective.
How it implements the components¶
regret_signal_classification— every candidate regret is sorted into actionable-learning, acceptable one-off, or noise; the sort is this component.learning_update_rule— for the actionable class only, it codifies a durable rule with the generalization/overfitting check built into the trigger.
It deliberately does not draw the control_and_responsibility_boundary or reconstruct the decision_rationale_trace — that fairness work belongs to its nearest twin, No-Fault Learning Review, whose controllability verdict this filter consumes rather than produces.
Related¶
- Instantiates: Regret-Signal Calibration — supplies the triage that decides which regrets become durable rules.
- Consumes: No-Fault Learning Review — for the blameless account and controllability verdict on each regret.
- Sibling mechanisms: No-Fault Learning Review · Counterfactual Plausibility Screen · Forgone-Alternative Decision Journal · Minimax Regret Matrix · Regret Gap Table · Regret Pre-Mortem · Commitment Reset Memo · Reversal-Window Check · Rumination Timebox
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: The mechanism runs a post-outcome review that turns a regret into a durable rule change only when the lesson is both controllable and recurring — otherwise it routes the regret to closure, so its operative form is a bounded assessment of existing evidence or work.
Independent corroboration: The frozen evidence defines Actionability-Filter After-Action Review as 'Runs a post-outcome review that turns a regret into a durable rule change only when the lesson is both controllable and recurring — otherwise it routes the regret to closure', so its operative form is Assessment, Review & Assurance.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Military & Strategic Studies
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Universal
Rationale: The after-action review's structured comparison of intended and actual outcomes and conversion of findings into future practice has a strong modern military lineage.
Related originating lineages:
- Engineering & Design — Incident review and prevention of one-off fixes becoming permanent process scar tissue contribute the reliability application.
- Organizational & Management Science — Organizational learning and retrospective practice contribute rule updates and formal closure.
- Psychology — Regret, hindsight bias, controllability, and overlearning from vivid outcomes supply the calibration problem.
Review resolution: The after-action-review lineage is military, but the filter combines it with engineering corrective-action discipline, organizational learning, and psychology. Cross-disciplinary synthesis and universal reach best capture that generalized post-review gate.
Attribution caveat: The actionability-and-recurrence filter is an Encyclopedia refinement of a generic after-action review rather than a canonical military procedure in exactly this form.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; medium confidence.
Notes¶
[n1] After-action review — a structured retrospective format originating in US Army training that asks what was expected, what actually happened, why the difference arose, and what to sustain or change; the actionability filter adds the recurrence-and-controllability gate that a bare AAR lacks. ↩