PLAYTEST NOTES · TABLE TALK
Behind the AIA good taunt needs the whole table.
Ender had just taken seven Blue-owned cells. The move gained 135 points, cost Blue 40, and swung the margin by 175. The commentary receipt was accurate:
Moved 7 opponent-owned cells.
It was also missing the most interesting thing at the table. Ender was now ahead by only 60, six target cells were still open, and Blue had the next move.
That is the difference between a transaction log and table talk.
A line can be true and still read the position badly
The first version of OverGrid commentary was deliberately conservative. The game engine made every move and score deterministically; the language model received a small packet of approved facts and turned them into one short line. It could add voice, but it could not invent a tactic or change play.
That boundary was right. The packet was too thin.
“Seven cells stolen” describes Ender’s local effect. It does not say whether the move closed the game, opened a comeback, established a durable lead, or merely bought tempo. A player hearing that line is invited to infer dominance even though the public board still shows a live contest.
The repair is not to give the language model more freedom. It is to give it better facts.
Excluded at the boundary: unrevealed hands, future deals, RNG state, hidden reducer fields, private coaching, and predictions dressed up as facts.
For this turn, a grounded line could say:
Seven cells change hands. I’m 60 ahead, but six are still open. Your move.
That is more conversational because it acknowledges the other player’s agency. It is also more rigorous: every clause can be checked against the committed public state.
The personality belongs in the interpretation, not the evidence
The eight OverGrid opponents should not sound like the same scoreboard wearing different hats. But personality is safest when it changes how a shared situation is framed, not which facts supposedly exist.
Ender can treat a lead as an unfinished problem. Petra can be dry about a height route becoming inevitable. Bean can needle a player after a messy theft. Alai can sound amused when several options remain. The unnamed difficulty profiles can be clearer and less theatrical.
They all receive the same kind of evidence packet:
| Safe public fact | Conversational use |
|---|---|
| score and score gap after the move | “I’m ahead, but not safe.” |
| rank and current leader | “That puts me second—not first.” |
| territory remaining | “Four cells left. The board is closing.” |
| deterministic captures, steals, and height | “That route finally paid.” |
| active player and committed turn | “Your move.” |
And they all share the same forbidden shortcuts:
| Not allowed | Why |
|---|---|
| an opponent’s unrevealed hand | hidden game state |
| the future deal or RNG cursor | information no player has |
| a private coaching assessment | belongs only to the coached player |
| “you have no reply” without a proof | prediction presented as fact |
| hidden reasoning or chain of thought | unnecessary and unverifiable |
The model is a performer standing on a deterministic stage. It may choose the cadence; it does not get to move the scenery.
The Ender example has one more important boundary
Blue’s next hand was W Z L. We know that because it became visible after Ender committed turn 12. Ender did not know it at the decision boundary, so the record cannot prove that Ender should have foreseen Blue’s exact reply.
The honest verdict is narrower: the immediate effect fired; the response-cycle value is unmeasurable. We can improve the commentary now because the missing score gap and open territory were already public. We cannot use the later hand to retrofit clairvoyance into either the policy or the banter.
That distinction matters whenever a game record contains more information than a player possessed at an earlier turn. “Present in the replay” is not the same as “available to the actor.”
This works at a duel and at a crowded table
Two-player commentary can talk naturally in terms of “you” and “me.” A four- or six-player game needs public standings instead: who leads, where the speaker sits, the gap to first, and how much territory remains. The contract stays the same; only the rendering changes.
At a multiplayer table, one bot gets one short line after its committed move. If a human banters back, the reply should address the named bot when one is mentioned. Otherwise the most recently speaking AI may answer, while remote humans remain ordinary participants in the same room. Private coaching stays on a separate channel and appears only after the player confirms a move.
The important product rule is that conversation never blocks or alters the authoritative turn loop. If commentary times out, is rate-limited, or fails validation, the deterministic line appears and the game continues.
Better table talk is an observability result
This change does not make Ender stronger. It makes the public consequences of Ender’s move easier to read. That is still valuable: good commentary helps a player understand why a turn feels dangerous, unfinished, or reversible without asking them to parse evaluator jargon.
The next research question belongs back in the engine: from the public board before a move, can a bounded counter-swing audit estimate how much conversion or recapture capacity remains? That would be a policy receipt, tested against a frozen control. It must not be improvised by the language model.
Until then, the standard for a good taunt is simple: say something vivid about the whole table, and be able to point to every fact.