06-reference

commoncog why tactical decision games work

2026-08-25·reference·status: filed·source: Commoncog·by Cedric Chin
expertisedecision-makingrecognition-primed-decisiontraining-designnaturalistic-decision-makingfacilitation

"Why Tactical Decision Games Work" — @CedricChin

Why this is in the vault

RDCO's whole thesis about agents and about the founder's own skill curve depends on a question this piece answers directly: how do you compress judgment that normally takes a decade of cases to acquire? This is the most mechanically specific answer the vault has on that.

The core argument

A Tactical Decision Game is a facilitated scenario exercise: a group is handed a situation, given minutes rather than days, and each participant has to commit to a decision and defend it. The format came out of the US Marine Corps, developed by John Schmitt roughly thirty years ago, and is now a primary teaching method across Corps schools. Chin's claim is that TDGs build adaptive expertise — competence in novel, ill-structured situations — which he argues deliberate practice does not produce and may actively degrade (he flags a longer essay on that claim as still pending, so treat it as asserted, not demonstrated).

The lineage runs through Gary Klein's recognition-primed decision (RPD) model and the broader naturalistic-decision-making (NDM) program: experts don't compare options, they recognize a situation and it comes bundled with cues to watch, expectancies about what happens next, a goal ordering, and a ready action script. Expertise, on this account, is a library of frames. The training problem is therefore not "how do we teach better rules" but "how do we get people more reps at building and revising frames."

The mechanism — why talking through a scenario transfers

Four moving parts, and they only work together:

  1. Time compression. A situation that would unfold over days is run in an hour. Density of decisions per unit of calendar time is the entire point; the scenario doesn't need physical realism, it needs what Schmitt calls cognitive fidelity — the mental work has to match the real mental work.
  2. Time pressure. Committing under a clock forces recognition rather than deliberation, which is what recruits the intuitive machinery you're trying to train. A member quoted in the piece makes the load-bearing point: understanding an idea intellectually is a different state than having it available under stakes. Chin's own fourth driver is stated as emotional investment rather than time pressure as such; the clock is the delivery mechanism, and the visceral quality of committing in front of peers is what he credits with making the learning stick.
  3. Collective frame-sharing. This is the real transfer channel. When one participant says "the enemy could ford the stream," everyone else silently folds that possibility into their own frame. Each alternative reading, prioritization, or prediction becomes a fragment the others can reach for in later scenarios. Transfer happens without anyone being taught anything — which is why the group composition matters and why at least one genuinely skilled participant should be in the room.
  4. The facilitator. A skilled one never calls an answer stupid; they ask "what's your thinking there?" and let group agreement form on its own, then advance the scenario based on what surfaced. There's also a deliberate early pause for questions, which exposes what other people are noticing — transparent reasoning as a teaching artifact. Chin does not discuss what happens with a bad facilitator, which is the piece's most obvious blind spot: it assumes facilitator quality and never addresses mislearning from poor exemplars.

Mapping against Ray Data Co

Three places a TDG plausibly plugs in, and one where it doesn't.

Discovery-stage ROI sniff-tests as the scenario set. The use-case delivery motion (Discovery → Delivery → Realization) fails most expensively at Discovery, when a use case with a bad value model is waved through and the cost shows up at Realization. That's an ill-structured judgment call made under time pressure with incomplete numbers — exactly TDG-shaped. The scenario is a one-page use-case brief with a plausible-looking ROI model; participants have ten minutes to say fund / kill / send back, and why. The fragments that accumulate are the tells: which assumption is doing all the work, where the baseline is missing, when a benefit is a transfer rather than a gain. This is more valuable than a scoring rubric because rubrics get gamed and frames don't.

AI Council prioritization rehearsal, run before the real one. A council that has never disagreed in practice makes its first hard tradeoff live, in front of stakeholders. Run the same body through two or three invented portfolios first — a cheap use case with weak sponsorship against an expensive one with a named executive behind it — and let the disagreements happen in a room where nothing is at stake. The output isn't a decision; it's that each member has heard how the others read a case, which is the fragment-sharing mechanism operating exactly as described.

Ray's own critic/verifier drills. The verify-* family currently scores artifacts against rubrics — source-aware review, which is closer to deliberate practice than to adaptive-expertise training. A TDG-shaped variant: hand a critic subagent a scenario mid-flight (a dispatch that is about to go out, a paper-trade proposal at the moment of commitment) rather than a finished artifact, under a turn budget, and require a call. Multiple critics run in parallel and see each other's reads before a final verdict — a machine analogue of the collective-sensemaking step. Worth testing whether that improves catch rate over the current single-pass rubric scoring; it is a hypothesis, not a result.

Sanity Check angle. The publishable frame is that most corporate AI training is deliberate practice aimed at a domain that punishes it — teaching prompt syntax when the actual failure is judgment about which problems to point at. The Marine Corps solved this thirty years ago with a whiteboard and a clock. That's an original re-frame rather than a summary, which is the bar for a Sanity Check piece.

Where it doesn't fit: anything with a checkable right answer or a stable procedure — pipeline correctness, contract/compliance review, cert prep for Snowflake GenAI or the Anthropic architect track. Those are well-structured domains where the answer key exists and drilling actually works. Running a TDG there wastes the facilitation cost and trains ambiguity tolerance in a place ambiguity shouldn't be tolerated.

Sponsor / bias note

Not sponsored, but heavily self-funneling. Schmitt ran a TDG for Commoncog members; the recording is paywalled, the forum discussion is member-only, sessions are capped at 20 participants, and Chin states outright that he intends to adapt these methods to the Commoncog Case Library and that "much of that experimentation will be made available to you as part of your membership." He also cites his own prior course ("Speedrunning the Idea Maze") as proof-of-concept. The piece is simultaneously a genuine mechanism explanation and a rationalization for a product roadmap — the mechanism section stands on the NDM literature independently, but read the "deliberate practice doesn't work for adaptive expertise" claim with the commercial incentive in view, since that claim is what makes Commoncog's format necessary and is the one Chin admits he hasn't yet argued in full.

Named sources worth chasing: Accelerated Expertise (2016), Klein's Sources of Power, Schmitt's LinkedIn piece on cognitive fidelity, Jared Peterson (NDM), Vaughn Tan on Krenov and tacit craft transfer, Cognitive Transformation Theory, the data-frame theory of sensemaking, and the constraints-led approach from sports training.

Related