Selected work KV / QMDTrust 3D 2D replay

Arrow keys move the selected cell on the 8×8 board. Pointer selection remains available.

Selected cell B5.

BoardB01 Prior1,000 samples Sourceδ = 1 Seed42
Fire control

Both captains are waiting.

Belief layerkeys 1–4
Questionclick the ocean to retarget
Value of this question
Action value+.001
Source info.224
Net VOI+.004
Commit
Ask this candidate+.004 net value after the .050 question cost. Fire remains available at D6 (25.4% hit probability).
Signal flags · Qmax 10
Source posterior p(δ)
TRUST INVERT
MAP: δ = 0 / 1trust 52.38% · invert 47.62%
Best next shot
Captain 01 · blindD6 · 25.4%δ̂ locked at .10
Captain 02 · awareD6 · 25.4%joint over board × δ
Log
  1. READYPick a question, then choose: signal, or shoot.
  2. REPLAYSignal spotter walks the recorded opening. Retarget any cell to make a question your own.
What is exact
The browser recomputes both posteriors live from B01 — the run's frozen multiset of 1,000 seed-42 FastSampler draws (999 unique legal boards), its source priors, the entropy-scaled spotter weight (.30) and the .050 question cost. The same joint update over board × nine stationary sources runs behind every number on this page as in the 2D replay. Hit and sink logic resolves against the true twelve-tile B01 mask. On the untouched opening snapshot, the displayed VOI equals the recorded recommendation exactly (+.0002 action, .7118 source info, +.1200 net).
What is staged
The ocean, clouds, gunfire, shell arcs, scorch marks, the lighthouse and its lamp code, the brass needle and the glass glow are theater. The Spotter is the recorded deterministic inverter (δ = 1): it flips every truthful answer, and no adaptive source is modeled. Questions you choose past the recorded opening are not recorded run outputs. The prior never resamples — updates only reweight the frozen draws. Reveals proceed only where both captains keep nonzero empirical mass for the real outcome; otherwise no belief changes and the log says so.