Offline experiment · 32 decodechess.com pages fetched with Playwright, catalogued by Codex inside the research jail · fact generator ai-workspace/game-story-prod-review/regen/decode-facts.mjs on the production rules · chess.js only, no engine call · round 2, 2026-09-04

What DecodeChess explains, and which of our 58 pending moments its fact types can explain better than the story does today

1. How DecodeChess works, in its own words

• “AI algorithms that mimic high-level human cognitive functions.”
• “auto-explains the moves of a chess engine (Stockfish) in rich, intuitive language.”
• “cognitive computing approaches to emulate human abstract thinking.”
• “Unlike chess engines which search for good moves, DecodeChess looks for chess concepts.”
• “Every suggested chess concept is verified to being relevant and important to the decoded position.”
• “an open file is suggested as a factor…only if it can be proved to affect the future course of the game.”
• “bases its conclusions on vast searching abilities of the search engine (Stockfish in our case)”
• “auto-generates chess explanations that are similar to those of a human chess master”
• “combination of human thinking and computer calculation”
• “These annotated variations are a pure result of our algorithm’s work. They are not provided by the chess engine whatsoever!”
• “our algorithms use Stockfish hundreds of times in various depths”
• “allocate substantial CPU efforts to finding the best move.”
• “Decode Game uses a more shallow engine depth”
• “Deep Decode…uses more engine and AI resources”
• “relies on a proprietary algorithm which requires significant CPU resources” and is server-only.
• “Stockfish 12 NNUE…is running on our servers.”
• Stockfish has “a Search Function” for continuations and “an Evaluation Function” for positional factors.
• With NNUE, Stockfish's programmed chess concepts “are integrated into the neural network, which weighs them by itself.”
• Algorithm updates name improved identification of concepts, terminology, player plans/attacks, and causal relations in piece functionality.

Short version: it is not a language model. It is a rule and concept engine that runs Stockfish 12 NNUE hundreds of times per position at several depths, lists every fact it can prove from those runs (attack, defence, intention, prevented move, plan), and shows them all with an importance bar. One position takes 15 to 30 seconds (Decode Game) or up to 70 seconds (Deep Decode) on their servers. Our story has a 6-ply line per move and a few seconds.

2. The 59 explanation types it names

3. Which of those we can produce from what we already have

4. The numbers over the 58 pending moments

Round 2 (copy revision 32): the facts come from the production module libs/api/src/domain/zapp/game-story/board-facts.ts, ranked (pawn on the seventh > passed pawn > exchange count > threat > offered piece > escapes/defends > open file > undermines > prepares > queen trade under fire) and cut to one fact when the top fact is decisive, otherwise at most two of different kinds. In the live story a fact is added only when the line-based text has no material sentence, so the columns show what the rules produce, not always what the story prints. Check moves get no fact. The played-move column allows only: enemy passed pawn, enemy pawn on the seventh, gives up an open file, queen trade under fire.

Reading of the result: for the 14 green cases DecodeChess would say the same facts we can already compute, plus a proof that each threat is real. For the 29 amber cases it would add the reason behind the "intends" move, which we cannot prove from one line. For the 15 red cases neither side has a board fact; these are the plan, king-walk and pawn-race cases where DecodeChess's plan and counterfactual types (E10 to E14, E55) are the only thing that could explain them, and even those need its many-run verification.

5. Every pending moment with its DecodeChess-style facts

Paste the FEN into DecodeChess to compare. "fails" after a capture means the engine line does not play that capture, so the capture is refuted.

6. What I would do

• Add the cheap DecodeChess fact types to the story's fact set: escapes capture, threatens (with pin-aware defenders), defends, makes way, prevents. They are board-only, cost no engine time, and give a sentence for 43 of the 58 pending moments.
• Keep "intends" only when a later fact backs it (the intended move captures, checks, or threatens). A bare "intends Kg1" says nothing.
• Do not try to copy the plan, activation and counterfactual types. They need hundreds of engine runs per position. That is the part of DecodeChess that costs 15 to 70 seconds.

Catalogue: /root/santha/common/research/decodechess-explanations/out/decodechess-explanations.md. Injection report: none.