Interaction After-Action Review¶
Retrospective ritual — instantiates Other-Agent State Model Calibration
A recurring retrospective that asks where the model of the other agent helped, failed, surprised, or harmed — and rewrites the interaction rules accordingly.
Every other mechanism here works during an interaction; this one works after it. Interaction After-Action Review is the periodic, backward-looking ritual that puts a batch of past interactions on the table and audits the model that drove them — did we route our behavior through an accurate read of the other agent, or through our own assumptions dressed up as theirs? Its defining move is to treat the read of another agent as a thing that can be reviewed with the outcome known and then turned into changed practice: not "we misjudged that person" left as a private regret, but a named revision to the rules about what may be inferred, how the interaction is run, and — where a misread caused harm — a path back to repair it.
Example¶
A 911 dispatch center runs a monthly after-action review over a sample of hard calls. One call is replayed: the dispatcher had modeled the caller as uncooperative — refusing to give an address and routed the call as a compliance problem, sharpening tone to force answers. Reviewing it with the outcome known, the team reconstructs a different hidden state entirely — the caller was almost certainly dissociating under acute panic and could not retrieve the information, not would not. The review scores the call on four axes: where the model helped (early distress-reads that sped other responses), where it failed (panic misread as refusal), where it surprised (a caller who sounded calm but was in danger), and where it may have harmed (the escalated tone). The output is not a reprimand but two rule changes — a revised script cue that separates "won't answer" from "can't answer," and a callback to check on the misjudged caller — plus a note to training. The point is that a single misread became a standing change in how the center reads callers.
How it works¶
- Runs on a cadence, over a sample of already-completed interactions whose outcomes are known — it is retrospective, never live course-correction.
- Scores each interaction on four axes — helped / failed / surprised / harmed — with the "surprised" axis drawing on any prediction record kept during the interaction.
- Separates a model error (we read the state wrong) from a routing error (we read it right but acted wrong); the two call for different fixes.
- Converts findings into rule changes, not just insights — what the model may now be used to decide, what it may not, and what to do differently next time.
- Where a misread caused harm, opens a repair path to the specific agent instead of filing the miss away.
Tuning parameters¶
- Cadence and sample — how often it runs and which interactions it pulls (all / hard cases only / harmed parties only). Denser sampling catches more drift but costs time and can feel like surveillance of staff.
- Blame stance — whether the review judges the decision given what was known then or the outcome. Judging outcomes breeds defensiveness and hindsight distortion; judging the process keeps people candid.
- Rule-change threshold — how much evidence turns a one-off miss into a standing rule. Too low and the playbook thrashes on anecdotes; too high and repeated harm never changes anything.
- Repair reach — whether the review only fixes future rules or also loops back to the agent who was misjudged. Repair rebuilds trust but reopens closed interactions.
- Psychological safety — how protected participants are. Without it the ritual becomes theater in which only the safe-to-admit misreads surface.
When it helps, and when it misleads¶
Its strength is that it is the only mechanism here that closes the loop at the level of practice: it turns scattered individual misreads into revised rules and repaired relationships, and gives the archetype a memory that catches systematic biases no single interaction reveals.
Its central distortion is outcome bias — once the ending is known, the earlier read looks obviously wrong or obviously astute, and the review grades the model by how things turned out rather than by whether it was reasonable on the evidence available at the time.[n1] It is also easily run backwards, as a ceremony to ratify decisions already made or to assign blame, and it can ossify into a checkbox meeting that generates no change. The discipline that guards against this is to review the decision on the then-known evidence, protect candor above all, and require that every session name at least one concrete rule change — or explicitly decide none is warranted.
How it implements the components¶
Interaction After-Action Review realizes the learning-and-governance side of the archetype — the components that convert experience into changed behavior, not the ones that gather it:
feedback_update_loop— the review is the loop: surprise and failure feed back into a revised model of the other agent, and of how to read agents like them.action_routing_gate— its output rewrites which decisions may be routed through the model and which may not (e.g. "never route tone-escalation off a bare 'refusal' read").model_repair_apology_path— where a misread harmed the other agent, the review opens repair or apology rather than closing the book.
It does not gather the evidence it reviews — direct inquiry belongs to Perspective-Taking Interview, the moment-to-moment prediction record to Prediction and Surprise Log, and the standing state model to Belief-Desire-Knowledge Map. The AAR consumes those and acts on them.
Related¶
- Instantiates: Other-Agent State Model Calibration — the AAR is the appraisal's memory, turning misreads into revised interaction rules.
- Consumes: Prediction and Surprise Log supplies the recorded predictions and surprises the review works from.
- Sibling mechanisms: Prediction and Surprise Log · Perspective-Taking Interview · Consent and Privacy Boundary Checklist · Belief-Desire-Knowledge Map · Counterparty Model Red Team · Active Listening Loop · Empathy Map with Evidence Marks · False-Belief Check · Role-Reversal Simulation · Stakeholder Hidden-Constraint Board
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: Interaction After-Action Review operates as a bounded evaluation of existing evidence or work that produces a finding or disposition because it a recurring retrospective that asks where the model of the other agent helped, failed, surprised, or harmed — and rewrites the interaction rules accordingly
Independent corroboration: The frozen evidence defines Interaction After-Action Review as 'A recurring retrospective that asks where the model of the other agent helped, failed, surprised, or harmed — and rewrites the interaction rules accordingly', so its operative form is Assessment, Review & Assurance.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Military & Strategic Studies
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Structured after-action review arose in military practice as disciplined comparison of intent, action, outcome, and lesson.
Related originating lineages:
- Organizational & Management Science — Organizational learning materially adapts the retrospective to recurring coordination and rule revision.
- Psychology — Outcome-bias and mental-model research materially shape fair recalibration of how another actor was understood.
Review resolution: Both independent reviews place the primary lineage in military_strategic_studies. The queued differences (origin_mode_disagreement) concern secondary metadata rather than primary provenance. The final retains organizational_management, psychology only where a reviewer supplied a formative-lineage rationale; this does not convert downstream applicability into origin. origin_mode=cross_disciplinary_synthesis because the entry's present form deliberately composes methods from the documented lineages. domain_reach=multi_domain records application breadth separately from provenance.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The AAR is deliberately retrospective and periodic — it is not live course-correction (that is active listening) and not a running record (that is the log). Its value rests entirely on candor: a review run without psychological safety produces confident-sounding rule changes built only on the misreads people felt safe admitting, which can be worse than no review at all.
[n1] Outcome bias — the tendency to judge the quality of a past decision by how it turned out rather than by what was known when it was made. In an after-action review it makes reads that preceded a bad ending look negligent and reads that preceded a good one look wise, regardless of the evidence available at the time; grading the decision on the then-known evidence is the standard corrective. ↩