A 2D board and a formula desk: four quadrants light up and score, candidate moves draw themselves as arrows, the two equations fill in number by number, then Stockfish 19's line plays out. Draw a random position, or paste your own, and the whole run, narration included, is written again from its numbers.
The brief it was built from
The Viveka method (version 3) taught on a 2D board, as a narrated training video that works on any position: rank the quadrants, propose candidate moves, score them, compare with Stockfish. The visitor chooses the board's look, the position (five lessons, a random held-out position, or their own) and their level; every variable and score is computed live and shown as it is computed. A companion to the 3D piece The best move, by Viveka.
What is real, and what is a scenario
Real data
- The two equations and constants are Viveka v3 as published: Φ = U + 1.5·ln(1 + N) for each quadrant, S = M + 1.25·K + 0.25·C for each move (α and β fitted at 1.13 and 0.21, rounded to quarters)
- Every U, N, Φ, M, K, C and S on screen is computed in the browser by viveka.js, a line-for-line port of the method's own TypeScript (src/lib/chess/viveka3.ts, movescore.ts and candidates.ts in the Viveka repository); build/verify.mts ran both on all 5,499 positions this page can show (227,946 checks) with 0 mismatches
- The 5,494 random positions are the study's entire held-out test split: positions from 21 openings never used in fitting, each with Stockfish 19's evaluation of every legal move and its stored principal variation (data/pool, baked by build/bake.mjs from the repository's practice set)
- The five lesson positions are worked examples from the method's lessons, with the same Stockfish 19 labels
- The engine line that plays out in the verdict is Stockfish's stored principal variation for its best move
- Study figures are from the held-out test of 5,587 positions: the top two quadrants held Stockfish's best move 78% of the time; with three candidates in them, the method lost 36.4, 47.1 or 102.7 centipawns a move on average for the strong, good and club models of intuition; a random legal move loses about 361
- Where Stockfish's best move landed by quadrant (35%, 36%, 16%, 13%) and the variable comparison in the master chapter are from the training openings
- Chess rules are chess.js 1.4.0 (BSD-2-Clause), vendored; the page runs no engine
Scenario & assumptions
- The level input maps each visitor level to one of the study's three models of intuition: new to chess uses the club model (three of the eight best moves in the top two quadrants, drawn at random), club player the good model (the best move plus two of the next seven), master the strong model (the three best). That mapping is this experience's assumption, not a measured rating band
- The random draws for the good and club models are seeded by the position, so a position always shows the same candidates; another draw would show different ones
- On a position the visitor types in there is no engine data, so the three moves the method's glance scan ranks first in the top two quadrants stand in for intuition, and there is no verdict
- The one-line set-up and the lesson for a random position are chosen by rules from its numbers (what is loose or hanging, whether the engine's move was a candidate, what decided S); for the five lessons they are written by hand
- The Clarity meter is illustrative and rises step by step; it is not a measurement
- The free-energy reading of the quadrant equation is an analogy the research uses to motivate its form, not a claim about chess
- The method's study covers positions with 8 or more pieces from engine games; a position typed in with fewer is outside its scope and labelled so
- Random and typed-in positions are narrated by the visitor's device voice, because their sentences cannot be recorded in advance; the lessons use the baked voice where it was baked
- Five minutes is the autoplay pace at the club level; the hands-on panels take longer
What you leave with
- Where to look: Φ = U + 1.5·ln(1 + N), material at stake plus options, for each quadrant seen from your side; search the top two
- What to play: S = M + 1.25·K + 0.25·C, material after the exchange, a check worth about a pawn, a quarter per ring towards the centre
- The formula only chooses among the candidates you propose: candidate quality moves the average loss from 36 to 103 centipawns
- It looks one move deep, so it can take a poisoned pawn that Stockfish takes a move later
Sources
chess.js 1.4.0 (BSD-2-Clause) · Narration voice: Kokoro-82M v1.0, voice af_heart (kokoro-onnx 0.6.1). Apache-2.0 (Kokoro-82M weights). · Classic chess pieces by Colin M.L. Burnett (GPLv2+; also BSD and GFDL), as bundled with react-chessboard · Merida pieces by Armando Hernandez Marroquin (GPLv2+) · Chessnut pieces by Alexis Luengas (Apache-2.0) · Fantasy and Spatial pieces by Maurizio Monge (MIT) · RhosGFX pieces (CC0 1.0)
Learning paths
Original
Learning paths
Original
Life with a product
Original