Source
Kirsh and Maglio (1994), LIT-1054 (Cognitive Science 18(4), 513–549), read in NOTE-878 from the scan of the version of record: Figures 6, 7, 10 and 15, Table 1, the Discussion. Kirsh (1995), LIT-1076, read in NOTE-871, restates the Tetris case in §5.1.
What was actually shown
The definition. “we use the term epistemic action to designate a physical action whose primary function is to improve cognition by: 1. reducing the memory involved in mental computation, that is, space complexity; 2. reducing the number of steps involved in mental computation, that is, time complexity; 3. reducing the probability of error of mental computation, that is, unreliability.” Pragmatic actions bring the agent closer to the physical goal.
Over-rotation (Figure 6). Each piece type is rotated more than the plan-then-act expectation (1.5 rotations for a four-orientation piece, 0.5 for a two-orientation one), with 95% confidence intervals: about 1.7–1.75 against 1.5, and 0.6–0.65 against 0.5, read off the figure. The number of players and episodes behind it is not stated.
Early rotation (Figure 7). “rotation begins in earnest by 400 to 600 ms and, on occasion, at the very outset of an episode”, some before the piece is fully in view. Rotating an emerging piece exposes its hidden part “150 to 300 ms earlier”; pieces ambiguous in both shape and starting column are rotated more than others, with no test statistic reported. Over-rotation recovered by backtracking would cost 75–750 ms, so the classical sequential and interleaved models do not fit.
Translate-to-wall (Table 1). In about 1% of drops the player moves the piece to the wall and back first. Drops after this routine are longer (about 19–20 against 13–16 squares, t tests at .05 within each skill level), and its use rises with drop distance.
The cost argued. “zoids can be physically rotated 90° in as few as 100 ms, whereas we estimate that it takes in the neighborhood of 800 to 1200 ms to mentally rotate a zoid 90°”, from a pilot with three players. On the figure, 800–1200 ms is total reaction time; the increment for 90° is about 300 ms. Kirsh 1995 gives 700–1500 ms and 150 ms for the same comparison. No score or placement was compared with and without the moves.
When it pays. The Discussion names two conditions: the action is fast relative to the computation it replaces, and its local effect is determinate (Tetris has “no hidden states, exogenous influences, or other agents”). It says the choice fits Stigler’s economics of information in outline, but “simple economic models of costs and benefits fail because the benefits of an epistemic action depend in considerable detail on just what computations the agent is performing”.
What this does not say
- That the moves save computation. That is the reading. Function is inferred: “an action can serve both epistemic and pragmatic purposes simultaneously”, and which it served is “often hard to prove”. The mechanisms offered (multiple-perspective representations, settling, decision-tree attention) are hypotheses, not tested.
- That a result is stored for later. Nothing is kept: a rotated piece is read within a few hundred milliseconds and used once. The paper sets itself against caching: “Agents do not simply cache associative rules describing what to do in particular circumstances.” Amortization appears once, in a parenthesis about sorting bolts before assembly, unanalysed.
- That offloading always pays. The conditions named are speed and determinacy; where effects are uncertain, or the world can change, the argument is not made.
- That a cost model exists. The paper says one cannot be given without modelling the particular computation an action serves.
- That experts do more of it, as a measured result. Asserted without a table.
Connections
- QUESTION-095: this is the “offload at the moment of use” strategy, the opposite end from the cockpit’s precomputed speed bugs. Staleness does not arise, because the result is used at once and the effect is determinate.
- THEORY-355 holds the maintained-arrangement case from the same author.
- THEORY-334: Ballard et al. defer and re-read the world; Kirsh and Maglio act on it to make the computation cheaper. In both nothing computed is stored.
Narrowed and promoted (2026-10-20)
Batch 074 of the promotion audit narrowed the title to the timing and the Table 1 comparison and promoted it under ADR-047 v3. Over-rotation and the height-dependence of the wall check are read off figures with no printed numbers, and the saving in computation is argued, not measured, so they are held with the authors’ epistemic-action reading as THEORY-823, Proposed until an experiment varies what the moves cost.