馃敀
International Knowledge Platform
Serving Researchers Since 2012

Voxplay: Pacing and Verifiable Grounding for Spoken Chess Commentary Generated from Move Notation

DOI : 10.5281/zenodo.24446477
Download Full-Text PDF Cite this Publication

Text Only Version

Voxplay: Pacing and Verifiable Grounding for Spoken Chess Commentary Generated from Move Notation

Authors: S. Sashank Datta, Rishi Varma, Aakanksha Saisri, Ayush Mishra, S. Kalyani

Department: School of Computer Science and Engineering, VIT-AP University, Amaravathi, India

ABSTRACT

Millions of chess games are played online daily and almost none are explained. Engine analysis yields a table of labels; human commentary exists only for elite play. Prior work emits one comment per move as text and measures factual accuracy after generation rather than constraining it. Voxplay turns move notation alone into spoken commentary synchronised to an animated board. First, commentary is paced: beat selection is budgeted maximum-weight selection, solved greedily to a mean 98.7% of the exact optimum. Over 173 standard Lichess games it runs 3.6脳 shorter than move-by-move narration while covering the opening and finish in every one. Second, the model is never asked to derive a chess fact; facts come from deterministic components validated externally 91.9% mean recall and 89.9% precision over 14,000 human-tagged puzzles, and = 0.955 for the classification rule against Lichess annotations on a held-out 6,864 moves, falling to 0.673 once our own engine supplies the evaluations. Third, every sentence is checked against the facts licensing it, and the check is itself auditable: run over grounded text it exposed a licence defect we then fixed, leaving 0.66 false positives per 100 beats, though the checks are lexical and cannot see a sentence that misuses licensed vocabulary. Generation costs US$0.06 per game.

Keywords sports commentary, data-to-text generation, content selection, grounded generation, hallucination mitigation, chess.

  1. INTRODUCTION

    A chess game is a complete and unambiguous record: every move is known, every position computable, and a strong engine can quantify the consequence of each decision. Yet what an ordinary player gets afterwards is a table moves tagged blunder, mistake or inaccuracy beside numeric evaluations. It says a move was bad, not what was missed, which of eighty moves decided the game, or why; and it cannot be listened to. Human commentary supplies all of that and is restricted to elite play. This paper describes a system occupying that gap: the input is the move list, and the output is spoken audio with a board that follows the narration, with no video and no vision model involved. Chess admits this because its facts are computable and checkable. That permits a pipeline in which the model is never asked to derive a chess fact and every sentence it writes is checked against the facts it was given. Neither step makes invention impossible the ungated pilot of Section V-E produced two unsupported claims in 32 beats but together they remove the occasion for it and give a mechanical test for what survives. The same property is why the approach generalises less readily to domains whose ground truth must itself be inferred.

    Contributions. (i) Pacing as budgeted selection: we formalise what to say as maximum-weight selection under a spoken-duration budget and minimum separation, give the objective and constraints explicitly, name the solver and bound its suboptimality against an exact dynamic program, with two baselines implemented in the same codebase. (ii) Grounding by construction: all chess facts are computed by independently validated deterministic components and supplied as a closed fact block. (iii) A verification gate with a symmetric licence derived from the same function that builds those facts, whose false-positive rate we measure rather than assume. (iv) Speech with a measured timeline, offsets coming from the rendered waveform rather than word counts.

  2. RELATED WORK

    Chess commentary. Jhamtani et al. [1] introduced the task from forum discussion; Zang et al. [2] coupled generation to a neural engine; Lee et al. [3] added symbolic reasoning; Kim et al. [4] introduced concept-guided generation with a dedicated metric. All four generate text per move. Concept prioritisation [4] selects what to say about a move; none selects which moves to speak about at all, which is what a listener experiences as pacing.

    Factuality. Hebbar et al. [5] report GPT-5.4 producing incorrect sub-claims 22.0% of the time without tools, open-weight models exceeding 40%; their contribution is evaluation after generation. Ji et al. [6] distinguish intrinsic from extrinsic hallucination; we target the extrinsic kind by making the admissible fact set finite and explicit.

    Content selection. Barzilay and Lapata [7] framed content selection as a constrained decision over candidate records; Wiseman et al. [8] showed on RotoWire that fluent generation from structured records implies neither faithful nor well-selected content; Puduppully et al. [9] made selection and ordering explicit as a content plan. Our formulation adds a duration budget, meaningful only once output is spoken.

    Other sports. MatchTime [10] and SoccerNet-Caption

    [11] generate soccer commentary from video, where ground truth is inferred rather than computed; a survey [12] confirms the pattern. That a language model can write chess commentary is established; our claims are pacing, spoken delivery and grounding enforced by construction.

  3. SYSTEM ARCHITECTURE

    Fig. 1. The six-stage pipeline. Stages 1-3 compute chess facts before the model is invoked; stages 4-6 pace, realise and speak them.

    The pipeline has six stages (Fig. 1): ingest notation into move events; evaluate every position; detect motifs and opening deviation; select and pace beats; realise the facts as prose; and verify, then speak.

    The architectural invariant is the boundary between fact computation and generation: everything the narration may assert is computed before the model is invoked, and the dependency graph is acyclic, so beat selection can be tested without an engine and verification without an API key. The implementation is roughly 11,600 lines of Python with 555 automated tests, using python-chess, Stockfish and claude-opus-5.

  4. METHOD

    A. Evaluation in win probability

    Centipawn differences are not comparable across positions: losing 100 centipawns from +900 is irrelevant, whereas losing 100 from equality can decide the game. Every

    misdescribes the position; the cap is disabled for Section V-B.

    Community documentation gives the bands as roughly 10/20/30, about twice too lenient. Fitting to maximise agreement recovers 5/10/15 exactly Lichess's documented 0.1/0.2/0.3 winning-chances delta rescaled to 0100 so our claim is correspondingly modest: a high shows the implementation is faithful, not that a new method has been found.

    1. Motifs and opening theory

      Twelve motifs are detected from board geometry alone, with no engine and no language model: fork, pin, skewer, discovered attack, hanging piece, sacrifice, back-rank mate, smothered mate, double check, promotion, en passant and advanced pawn. A model asked whether a position contains a fork answers confidently either way, and a listener who plays chess catches the error. Detector names match Lichess puzzle themes, so validation is a direct comparison. Geometry shows material was given up but not that a sacrifice was sound; soundness is decided in the analysis layer. Naming the opening by prefix-matching moves would miss every transposition, so we instead replay each line of the CC0 lichess-org/chess-openings database and index the resulting position.

    2. Pacing as budgeted selection

    For each half-move m of a gae M we compute six normalised signals in [0, 1] evaluation swing s, narrative role n, strongest motif t, criticality k (the gap between best and second-best move), material swing a, and think time h relative to that player's own baseline combined with weights w = (1.00, 0.60, 0.45, 0.35, 0.25, 0.20):

    (m) = w 路s + w 路n + w 路t + w 路k + w 路a + w 路h (2)

    s n t k a h

    evaluation is therefore mapped to win probability by the logistic published by Lichess [13], with centipawns c clamped to 卤1000 and = 0.00368208:

    W(c) = 50 + 50 ( 2 / (1 + e-c) – 1 ) (1)

    The mapping saturates at the extremes, so the same 100-centipawn delta is worth 9.1 points of win probability from equality but 1.5 from +800 a six-fold difference, which is the property centipawns lack. Every downstream component reads W, and the constant is adopted from Lichess so our classifications are directly comparable with Lichess's own annotations. Analysis is budgeted by nodes, not depth or wall-clock time, since time-limited search returns different results on different machines and would make every number here irreproducible: 1,000,000 nodes per position, MultiPV 3 and a single search thread (Section VII-A).

    B. Move classification

    Let L be the win probability the mover lost. A move is best if it is the engine's top choice, book if also within opening theory, otherwise good, inaccuracy, mistake or blunder as L falls below 5, in [5, 10), in [10, 15), or at or above 15. A commentary variant caps severity at inaccuracy once the mover's W exceeds 90, since calling a move a blunder while the player is still completely winning

    Swing saturates at the blunder threshold, so any blunder maxes that term. Each beat costs c(m) spoken words, fixed by type: 65 for the turning point, 45 for an ordinary moment. Writing x_m {0, 1} for the decision to speak about m, and P M for candidates clearing a floor (m) 0.25 and not already covered by a mandatory beat, selection solves:

    maximise mP (m)路xm (3)

    m

    subject to mP c(m)路x B (4)

    xi + xj 1 whenever |ply(i) ply(j)| < g (5) with separation g = 4 half-moves. A listener's budget is a

    duration but the selector adds up words, related by the

    measured rate r = 1.85 words per second. The target duration T scales with game length, clamped at both ends, and B is that duration in words less those already committed to the mandatory intro, opening, finish and outro beats R, which supply the coverage requirement directly rather than as a constraint:

    T(M) = min( 360, max( 150, 2路|M| ) ) seconds (6)

    B = floor( r路T(M) ) jR c(j) (7)

    For a 33-half-move game T = 150 s, so r路T = 277 words; the mandatory beats cost 153, leaving B = 124 a turning point and one further moment. A fixed budget does not transfer across game lengths.

    Solver and optimality. Equations (3)(5) are solved greedily: candidates sorted by descending are accepted if they fit the budget and violate no separation constraint, which is O(|M| log |M|) in practice. Greedy is not optimal in general, but the problem is exactly solvable: with candidates in ply order, (5) between consecutive chosen candidates implies it for every pair, so (3)(5) is a knapsack on a line admitting an exact O(|P|路B) dynamic program, tractable because c is integral in words. We implement it as a reference and report the gap in Section V-D. Finally, r = 1.85 is measured, not nominal: a 134-word script rendered at a nominal 155 wpm produced 72.6 s of audio, an effective 111 wpm.

    1. Script generation

      One call establishes a through-line; each beat is then realised in a separate call receiving the outline, the lines already spoken, and a fact block that is the complete and only basis for that sentence. Notation is converted to spoken form by rule, not by prompt: asking a model to read "Qxh7#" aloud succeeds most of the time and then produces "queen x h seven hash" on the move that matters. For full-coverage sheets, beats are generated in runs of six and parsed back by index, the batch rejected entirely if the shape is wrong a partially parsed batch would attach one move's commentary to another, an error in which every individual claim stays true and which the verifier therefore cannot detect.

      Four prompts are sent beneath one cached system prompt

      outline, per-beat realisation, a batched variant, and a repair prompt appending the verifier's finding all verbatim in supplementary Section S1. The system prompt states that every chess fact must appear in the FACTS block, notation must never be written, and no numeric evaluation may be stated. The architecture does not rely on the model obeying these: the fact block is the only chess content in the context, and the output is checked against a licence built by the same function.

    2. The verification gate

      Each sentence is checked against a licence: the claims the beat's facts support permitted squares, motifs, names, the evaluation, and whether the result may be revealed. Seven checks run: notation leakage, ungrounded squares, ungrounded tactics, evaluation contradiction, ungrounded names, cross-beat repetition and length. The first five are factual defects; the last two are stylistic warnings, excluded from every error rate below.

      The licence is derived from the same function that builds the fact block. This symmetry matters in both directions: a fact supplied but not licensed produces a false positive, and a claim licensed but never supplied produces a blind spot. Section V-E measures the first and finds one. A failing beat is regenerated with the failure fed back, and the same gate checks text a user edited by hand, so an invented claim typed by a person is caught identically.

      What the licence cannot express. Every check is lexical: which squares, motif names, player names and evaluation

      direction appear. A sentence drawing only on licensed vocabulary can therefore be false about the position and pass. Five such sentences, each built from one beat's own licence, pass all seven checks: giving the move to the wrong side, naming a licensed square but the wrong piece on it, attaching a licensed motif to a move that did not create it, a strategic claim with no lexical anchor ("controls the centre"), and revealing the outcome without a result word. Two controls reaching outside the licence are caught, so the checks fire when they can. The gate bounds what may be named, not what may be asserted about it.

    3. Speech and timeline

    Audio offsets are measured from the rendered waveform, never estimated, and compressed clips are transcoded before any offset is taken, so the choice of voice cannot silently desynchronise the board. Each clip is synthesised with its neighbouring lines as unspoken context [14]; without it a line closes as though the game ended there, and a six-beat recap is heard as six announcements rather than one person speaking.

  5. EXPERIMENTAL VALIDATION

    All results are measured against externally produced labels this project did not generate. Every number is produced by a script in docs/paper/studies/; supplementary Section S4 maps scripts to tables.

    Corpus and inclusion criterion. The corpus is a PGN export of analysed Lichess games (CC0), seeded from Lichess TV which carries the variant channels, so the raw export of 339 games proved to be 174 standard and 165 variant games. Every result below uses the standard subset only: Stockfish declines Antichess outright; for Horde, Atomic, King of the Hill, Racing Kings and Three-check the win condition differs, so a standard engine's evaluation does not mean what (1) assumes; and although Chess960 shares the rules, the engine wrapper does not set UCI_Chess960. An earlier version of this work applied no such criterion and reported 14 games lost to "engine failures" that were in fact Stockfish refusing Antichess, leaving a corpus roughly half variants scored by a stadard engine.

    1. Motif detection: precision and recall

      For each detector, 1,000 puzzles carrying its own theme were drawn by reservoir sampling over 6.5 million rows of the Lichess puzzle database (CC0) [15], with 2,000 controls from five non-motif themes 14,000 puzzles in all. A detector is credited if it fires anywhere among the solver's moves, since a Lichess theme describes the whole solution, not its first move.

      Precision is reported at this 1,000-to-2,000 sampling ratio, not the prevalence of motifs in commentary, and is a lower bound because Lichess tags are not exhaustive. That is measurable: the back-rank detector fired on 232 controls and replaying each line shows all 232 do finish in checkmate on a back rank, so its 81.2% is an artefact of tagging. Hanging piece is excluded because Lichess tags it when the solution takes an undefended piece whereas we detect a move that leaves one hanging. The pin detector is the one genuine weakness, trading recall for precision at 49.2% against 63.5% unfiltered.

    2. Classification agreement, fitted and tested on disjoint splits

      Lichess publishes both halves of what this needs: [%eval] on every move and its own verdict as a NAG. This section uses all 174 standard games, counting only the 13,230 half-moves that carry an evaluation a game's unannotated tail cannot be scored, so it is not counted. Games not moves were shuffled under a fixed seed and split in half; was grid-searched on the training half alone, and reported on the other.

      The selector was run over the 173 standard games of 20200 half-moves, one game of the 174 being too short. Pacing needs every half-move, annotated or not, so this section counts all 13,251 of them rather than the 13,230 evaluated ones of Section V-B; median length is 68 (IQR 5592). No run failed.

      TABLE II

      PACING AGAINST BOTH BASELINES OVER 173 STANDARD GAMES

      Paced

      Top-N

      Every move

      Median duration

      143.2 s

      194.6 s

      514.6 s

      IQR

      142.2178.4

      constant

      416.2696.2

      Range

      117.8348.6

      194.6

      158.91483.2

      Median beats

      7

      8

      Opening covered

      100%

      22.5%

      100%

      Decisive move within a beat's span*

      98.8%

      100%

      100%

      Finish covered

      100%

      7.5%

      100%

      Games with no adjacent pick

      173/173

      1/173

      TABLE I

      AGREEMENT WITH LICHESS LABELS, FITTED AND TESTED ON DISJOINT SPLITS

      Moves

      6,366

      6,864

      fitted on train

      5 / 10 / 15

      , our rule on Lichess evals

      0.957

      0.955

      , documented 10/20/30

      0.343

      , our engine's evals

      0.673

      Raw agreement, our engine

      90.8%

      Expected by chance

      71.9%

      Train, 87 Held out, 87

      Held-out is within 0.002 of training , so there is no measurable overfitting unsurprising, since the search recovers a published constant rather than discovering a free parameter.

      That figure is close to tautological, though, and the last two rows matter more. = 0.955 has both sides reading Lichess's evaluations, so it isolates the rule and says little beyond that our implementation of a formula adopted from Lichess is faithful. The shipped pipeline does not read those evaluations: it runs Stockfish at 1,000,000 nodes on one thread. Scoring the same held-out games from our engine against the same labels gives = 0.673, raw agreement 90.8%, per-class 53.2%, 41.3% and 84.1%. That 0.282 difference is two engines at two search efforts disagreeing about borderline moves, and the rule-only figure hides all of it.

      Neither number is accuracy Lichess's annotations are another engine's output, not ground truth so this measures agreement between two configurations. The rule is faithful; the pipeline is substantially rather than almost perfectly aligned with a reference system; and the circulated 10/20/30 bands are about twice too lenient, yielding = 0.343.

    3. Pacing over 173 games

      Fig. 2. Spoken duration against game length over 173 standard games. The paced arm is clamped by (6); the every-move baseline grows without bound; top-N is constant by construction.

      *Over the 170 games that have a decisive move; three drawn games have none.

      Median reduction against every-move narration is 3.6脳 (IQR 2.93.9), widening with game length because the paced budget is clamped by (6) and the baseline's is not (Fig. 2). Top-N emits exactly N = 8 beats of fixed cost, so its 194.6 s is constant by construction worth stating, since an identical figure across games otherwise suggests a measurement error. Two cautions, though: top-N has no mandatory beats and a different budget, so its opening and finish coverage is structurally poor rather than empirically worse, and the adjacency row is (5) holding by construction. It also covers the decisive move in all 170 games that have one, against 168 for the paced selector. The informative comparison is the matched one.

      A budget-matched baseline. To separate the spacing constraint from everything else, the selector was run against itself with g = 0 in (5): same mandatory beats, same budget B from (7), same salience, same greedy order, differing only in whether adjacent picks are allowed.

      TABLE III

      THE SEPARATION CONSTRAINT ISOLATED, OVER 173 GAMES

      Mean per game g = 4 g = 0

      Spoken seconds

      167.1

      167.2

      Moments selected

      2.90

      2.91

      Total salience

      3.52

      3.73

      Adjacent picks

      0.00

      0.83

      Longest dead stretch, half-moves

      33.1

      37.4

      Decisive move selected as its own moment

      87.6%

      89.4%

      This is a smaller result than the unmatched table suggests, and two columns go the other way. Dropping (5) raises total salience by 6.2% it must, since removing a constraint cannot lower the objective and covers the decisive move slightly more often. What (5) buys is spread: the longest stretch a listener hears nothing about falls from 37.4 half-moves to 33.1, about 11%, and adjacent picks go to zero. So the defensible claim is narrower than "the constraints carry the contribution": the budget of (6) and (7) and the mandatory beats produce the duration control and the

      complete opening and finish coverage, while separation is a modest refinement trading salience for spread a trade for the listener study to judge, not the objective.

      Why the decisive move is ever missed. It is worth being exact, because the turning point is not a mandatory beat: it is an ordinary candidate carrying a narrative weight of 0.60 in (2), and it competes for the budget like any other. Both misses have the same cause, and it is the separation constraint (5) rather than the budget or the salience floor. In one game the turning point at half-move 32 scored 1.600 while its immediate predecessor scored 1.645; in the other, half-move 54 scored 1.628 against its successor's 1.755. Theneighbour was accepted first, and (5) then excluded a half-move one ply away. A decisive move and the reply to it share the swing that makes either salient, so near-ties between adjacent plies are the expected case rather than a pathology. In both games a beat still lands on the half-move immediately beside the turning point, so coverage within one half-move is 170 of 170; it is the exact-ply test that fails. A selector that resolved such ties toward the turning point would close the gap, at the cost of letting narrative role override salience.

    4. Greedy optimality and weight sensitivity

      Greedy reaches a mean 0.987 of the exact dynamic program's optimum, is exactly optimal on 138 of 173 games (80%), and its worst case is 0.750 the gap arising where a high-scoring candidate blocks two lower-scoring ones that together score more.

      The six weights of (2) are set by hand, so the question is whether each measurably changes the output. Over 80 standard games the selector was run at the shipped weights and again with each signal zeroed; the full table is in supplementary Section S6. Swing dominates: removing it retains only 45.1% of the shipped selection and costs a fifth of the objective. Narrative weight does a narrow but irreplaceable job: dropping it retains 80.4% of picks and costs 1.6% of salience, yet the share of games giving the decisive move a beat of its own falls from 88.6% to 57.0% invisible to the objective, and exactly what the term exists to produce. Motif, criticality and material shift the selection modestly at under 3% of salience. Think time is inert on this corpus: the games were exported without clocks, so its weight is untested rather than shown unnecessary.

    5. The verification gate

      Two quantities matter and only one is usually reported. The gate's false-positive rate is measurable without a language model: a template generator assembles sentences mechanically from the same fact block the gate licences, so every claim is true by construction and any error-severity finding is a false positive.

      The measurement found a defect. Over 456 beats from 60 standard games the gate raised 15 false positives, 3.3 per 100, twelve of them one cause: the name check strips punctuation per whitespace token, so a supplied opening such as "Nimzo-Larsen Attack" reaches it as NimzoLarsen, while the licence joined the whole string and admitted only NimzoLarsenAttack. An opening the model had been handed came back as an invented name exactly the asymmetry Section IV-F says the shared derivation prevents, surviving

      because no test used a hyphenated opening. With the licence corrected and a regression test added, and both arms re-measured over that same corpus, the rate falls to 3 false positives, 0.66 per 100 (exact 95% upper bound 1.91), and the games with none rise from 51 to 57 of 60. We report rather than pursue zero: the claim worth making is not that the gate is perfect but that its false-positive rate is measurable on grounded text.

      That rate is measured on the wrong distribution for a strong claim, and we say so rather than generalise it. Template sentences are assembled from the fact block, so they exercise the licence along its exact-string path, where it is strongest; model output paraphrases, and paraphrase is exactly where a lexical licence misfires, as the defect above shows. So 0.66 per 100 beats is a false-positive rate for template prose and a lower bound for anything else. Measuring it on model beats needs beats a human has confirmed correct, which is part of the protocol in Section VI.

      The gate's catch rate on model output is at present a pilot. Over 32 generated beats, disabling the gate yielded 2 unsupported claims (6.2 per 100 beats) and enabling it yielded 0, at an 8% cost increase; exact ClopperPearson 95% intervals are [0.8%, 20.8%] and [0.0%, 10.9%]. Those intervals overlap, so this pilot does not by itself establish that the gate reduces the rate; a 95% upper bound below 2% needs 183 consecutive clean beats. Provenance must also be stated plainly: the two errors were identified by the author with the checker's own output as the starting point, which is circular, and is why Section VI pre-registers independent annotation.

    6. Cost, latency and generation parameters

    The model is requested as claude-opus-5 and the API reports that same identifier back, so no dated snapshot is exposed. Adaptive extended thinking is enabled on every call, and with it enabled the API does not accept a sampling temperature: temperature is not a free parameter of this system. Effort is low for realisation, batching and repairs and medium for the outline; ceilings are 2,000 tokens for the outline and 1,000 per beat; the system prompt is cached. Each figure is a single run.

    Generation, on the 1858 Opera Game (33 plies) and a 79-ply blitz game, takes 22.4 s and 29.5 s over 7 and 8 calls and costs $0.0572 and $0.0631, both verifying clean. Cost and latency are approximately constant in game length, because the number of beats is bounded by the duration budget rather than the number of moves a direct consequence of pacing.

    Engine analysis is the other half of the runtime and is the half that scales. Timed on its own at 1,000,000 nodes, MultiPV 3 and one thread, two repeats per game on an idle machine, it costs 1.18 s per position 1.15 on a 34-position game and 1.19 on an 81-position one so it dominates runtime on any game of ordinary length. We report it as a measured rate rather than as per-game totals because an earlier draft quoted totals and a rate that were mutually inconsistent (supplementary S3). Results depending on a single proprietary model are a weakness; an open-weight replication is specified in Section VI.

  6. PLANNED EVALUATION AND FUTURE WORK

    The listener study has not been conducted. Its design, arms, sample size and analysis plan were committed before any data collection and are set out in full in supplementary Section S5; no human ratings are claimed anywhere in this paper. Section V establishes that the commentary is correct, not that it is worth listening to, which is the central claim of this work. The study compares five arms flat, top-N, the system, the system's pacing with mechanically assembled sentences, and a recruited club player so that two isolate pacing, one isolates the language model and one is the ceiling. A second protocol replaces the pilot of Section V-E with at least 400 beats per arm over 20 games, labelled independently by two annotators who do not see the checker's findings, and repeated on one open-weight model.

    What would most improve this paper for least effort is not that design but a pilot a fraction of its size ten listeners, three games, paced against top-N, one forced-choice question. That would move the central claim from planned to preliminary, and it is the one result no amount of further computation can supply.

    Live commentary is not a small change: pacing needs the game's length, and the turning point is by definition the largest error of the game, so a naive live system would reproduce the every-move baseline in real time. A fixed delay, or pacing within segments, preserves part of it. Also: rating-conditioned thresholds, multilingual output, Chess960 support, and learned salience weights replacing w of (2).

  7. LIMITATIONS

    1. Reproducibility constrains engine configuration. Stockfish's Lazy SMP shares a transposition table between racing threads, so a fixed node budget does not determine the result above one thread: two four-thread runs of one game disagreed on 27 of 33 half-moves, enough to relocate the turning point. All reported analysis is single-threaded, at 1.18 s per position.

    2. Standard chess only. Nothing here is evidence about variants, and because the available corpus was roughly half variants the evaluation set is 174 games rather than 339 statistical power paid to avoid measuring the wrong thing.

    3. otif detection is rule-based. Deflection, overloading and zugzwang are not detected, the pin detector recovers under half of tagged pins, and two move-quality labels are never assigned. The think-time weight is likewise untested, the corpus having no clocks.

    4. The hallucination result is a pilot, with overlapping intervals and non-independent labelling. The gate's false-positive rate is established on grounded text; its catch rate on model output is not.

    5. English only: the notation-to-speech converter is a rule-based English grammar, the prompts are English and the verifier's vocabulary is English, so translating after verification would void the gate's guarantee.

    6. Several reported quantities reached their final values only after an initial measurement was found to be wrong, in every case located by inspecting individual cases rather than aggregates. Supplementary Section S3 records each correction and its cause.

  8. CONCLUSION

We have described a system converting chess move notation into spoken commentary with a synchronised board, without video. Its contributions are pacing formalised as budgeted selection in (3)(5), solved greedily to a mean 98.7% of an exact dynamic program and measured over 173 games against two implemented baselines grounding by construction, and a gate whose licence is derived symmetrically from the facts it checks. Validation against external labels gives 91.9% mean recall and 89.9% precision over 14,000 human-tagged puzzles, and = 0.955 for the rule on a held-out 6,864 moves 0.673 for the pipeline that actually runs, a gap the rule-only figure hides. Auditing the gate against grounded text found and removed a licence defect, leaving 0.66 false positives per 100 beats; its effect on model output is a pilot with wide intervals, and we report the protocol that would settle it rather than the conclusion we would prefer.

REFERENCES

  1. H. Jhamtani, V. Gangal, E. Hovy, G. Neubig, and T. Berg-Kirkpatrick, "Learning to generate move-by-move commentary for chess games from social forum data," in Proc. ACL, 2018, pp. 16611671.

  2. H. Zang, Z. Yu, and X. Wan, "Automated chess commentator powered by neural chess engine," in Proc. ACL, 2019, pp. 59525961.

  3. A. Lee, D. Wu, E. Dinan, and M. Lewis, "Improving chess commentaries by combining language models with symbolic reasoning engines," 2022, arXiv:2212.08195.

  4. J. Kim, J. Goh, I. Hwang, J. Cho, and J. Ok, "Concept-guided chess commentary generation and evaluation," in Proc. NAACL, 2025, arXiv:2410.20811.

  5. S. A. Hebbar, P. Sheng, S. Oh, and P. Viswanath, "Hallucinations on the board: Tool-augmented evaluation of LLM chess commentary," 2026, arXiv:2608.04240.

  6. Z. Ji et al., "Survey of hallucination in natural language generation,"

    ACM Comput. Surv., vol. 55, no. 12, pp. 138, Mar. 2023.

  7. R. Barzilay and M. Lapata, "Collective content selection for concept-to-text generation," in Proc. HLT/EMNLP, 2005, pp. 331338.

  8. S. Wiseman, S. Shieber, and A. Rush, "Challenges in data-to-document generation," in Proc. EMNLP, 2017, pp. 22532263.

  9. R. Puduppully, L. Dong, and M. Lapata, "Data-to-text generation with content selection and planning," in Proc. AAAI, vol. 33, no. 1, 2019, pp. 69086915.

  10. J. Rao, H. Wu, C. Liu, Y. Wang, and W. Xie, "MatchTime: Towards automatic soccer game commentary generation," in Proc. EMNLP, 2024, pp. 16711685.

  11. H. Mkhallati, A. Cioppa, S. Giancola, B. Ghanem, and M. Van Droogenbroeck, "SoccerNet-Caption: Dense video captioning for soccer broadcasts," in Proc. CVPRW, 2023, pp. 50745085.

  12. Q. Zheng et al., "From multimodal perception to strategic reasoning: A survey on AI-generated game commentary," Jun. 2026, arXiv:2506.17294v3.

  13. Lichess, "Accuracy metric." https://lichess.org/page/accuracy

  14. G. Xu, W. Song, Z. Zhang, C. Zhang, X. He, and B. Zhou, "Improving prosody modelling with cross-utterance BERT embeddings for end-to-end TTS," in Proc. ICASSP, 2021, pp. 60796083.

  15. Lichess, "Open database" (CC0). https://database.lichess.org/