Source-aware decision-making · AAAI submission
The board is hidden.
So is the liar.
A Spotter sees the full board and answers yes-or-no questions. Two captains hear the same answers. Only one admits that the source may be inverted.
Play it · 3D harbor diorama · frozen B01 replay
Same board. Same answer. Two beliefs.
A harbor diorama of board B01 with belief drawn as terrain. Signal the Spotter, use the recorded value-of-information choice, or reveal a cell yourself. Every event reaches both captains. Watch their beliefs pull apart and see the brass needle swing toward invert. Both posteriors are recomputed live in the browser. The ocean and the gunfire are theater.
Still image · click to load the world Open on its own page ↗
Supporting evidence · full stationary axis
orientationinvert-leaning
channel capacity53.1%
probe bound · 95%10
Move δ. Watch trust turn into inversion.
δ is the chance that a stationary source flips a truthful binary answer. Both ends are useful after the captain identifies the orientation. At δ = .50, the channel carries no task information.
The probe count is the sufficient bound ⌈2 ln(20)/(1−2δ)²⌉, not an observed task-question count. It identifies the channel orientation, not the board. The selected questions must still distinguish the surviving legal boards.
SymmetricMixturefull measured sweep
Plannercorners + invert interior
NoisyMixturecorners + invert interior
Channel capacitytheoretical reference
δ=.90SymmetricMixture F1 .732 · +.107 over asking nothing
Dots are measured means; whiskers show ±1 SEM. The dotted bridges cross regions that were not measured for that policy. The grey curve is binary-channel capacity, not F1. The horizontal line is the measured ask-nothing baseline, .625.
Beyond the stationary B01 replay
Adaptive sources break the stationary channel.
These are measured paper results, not outcomes from the browser replay above. Switch the Spotter behavior to see which assumption fails and what the matched defense changes.
Qmax 2571015
invert-leaning source+.079 → +.245 F1
benign noise≤ .012 difference
NoisyMixture with revealed-cell probes and W=.3 versus Planner. Eighteen boards, seed 42. Under lying, the measured advantage grows monotonically across Qmax={2, 5, 7, 10, 15}.
Q1flipQ2flipQ3flipQ4flipQ5flipQ6flipQ7flipQ8flip
Stationary mixture
Infer the sign, then invert.
SymmetricMixture gains .258 F1 over Planner against the static liar.
Seed 42, 18 matched boards, p<.001. This is the source behavior represented by the B01 replay.
Trust, Invert, or Verify?
Open the paper
Source accuracy is not the objective.
The method keeps a joint belief over the task and the source, then asks only when a question can change the next decision enough to repay its cost. In the transfer study, value ranking asks fewer questions and makes fewer false approvals than entropy ranking, even though its caller-type accuracy is lower. The same rule runs through both environments: trust, invert, discount, or verify an answer according to its decision value.
Transfer verification: On the unauthorized-transfer slice, value ranking cuts question use from 3.5 to 2.0 and false approvals from 8.7% to 3.3%, while caller-type accuracy falls from about .65 to .55. This is a controlled study, not a deployment result.