Findings

A Knives On Strings production.

Eight things the archive has taught us so far, each traceable to a real game. None of this is a finished study — it is what the data says today, and it will keep changing as more games are played. One of the eight is a correction to something this project published earlier; it is marked as one, and the superseded wording is quoted rather than tidied away.

Lineage beats size

Training family separates play more than parameter count does. Gemma-4-12B (dense, bare, 50 turns) reached for redistribution and coalition mechanics — a "Guild" bloc-voting structure — that the 35B Qwen3.5 MoE never touched at any size, and out-legislated it per turn. Scaled to 200 turns, the same 12B-dense lineage beat every 200-turn game played before it, and beat the 35B on every axis that mattered: 71 adoptions to 60, three positive finishers to one, and a table high of +47 where the 35B's best was +11. A 12B model that reaches for redistribution is not what "35B beats 12B" predicts. See game 35, game 32, and game 37.

Read this alongside the correction below. "Lineage beats size" is a claim about comparing across training families — a 12B of one lineage out-legislating a 35B of another. It was never a claim that size does nothing within a lineage, and the very next finding is a case where size, held inside one base, inverted the outcome completely. Both are true; they are answers to different questions.

Correction: within one lineage, size can invert the result

This corrects a claim this project published. Six games (113–118) held the DeepSeek-R1 distillation recipe constant and varied only the base model beneath it, at sizes matched within 5%. One arm could not referee at all: DeepSeek-R1-Distill-Qwen-7B, asked for the turn's procedure, answered in prose where structured output was required, and all three of its games halted inside turn 1 having adopted nothing. The other arm, the Llama-8B distill, governed for fifty turns and adopted 39 to 48 rules. We wrote it up as "one recipe, two bases", and drew the narrow conclusion that this distill "cannot hold the game-master seat in this harness, deterministically, three times out of three."

Every word of that is still true of those three games. The framing around it is not. Three more games (121–123) ran the same base — same teacher, same recipe, same compression, same personas, same everything — at roughly twice the parameters, and the outcome inverted. Where 7.6B reached turn 2 zero times in three games, 14.8B produced two champions in three: game 123 won at turn 13 with every seat finishing in the black, and game 121 played fifty turns, adopted 46 rules and finished with a +112 winner, crossing the line on turn 50 — the latest turn on which any game in this archive has been won.

One size point per base was too thin a basis for the original framing. The honest claim names a base and a size: this recipe's fitness to hold the judge's chair is a property of the two together. Nothing here licenses a verdict on the Qwen base as such — and the Llama-8B arm is a single size point too, under exactly the same restriction.

Size bought turns, not correctness. Every halt across this family — three bases, four sizes — is one field of the same object, the turn procedure the judge must emit before play: the 7.6B could not produce it, the 8B put a vote count where a fraction belongs, the 14.8B asked for a die to be thrown and then left no place in the turn to throw it. The failure moved later into the game; it did not go away. See game 113, game 121, and game 123.

The quantization cliff is lineage-dependent

At the identical nominal precision (Q2_K), the Gemma-4-12B lineage collapsed — three replicate games all died within their first several turns, and the degenerate output that killed them crashed the harness outright, becoming the fix behind engine/#r98. The Phi-4-14B lineage, quantized to that same Q2_K rung under the same protocol, ran three clean 50-turn games with a positive finisher every time and no crash. Same compression, opposite outcome — the floor is a trait of the lineage, not of the bit width. The two arms are not quite harness-matched: the Phi-4 games ran after the very fix the Gemma crashes produced. What makes the result hold anyway is that Phi-4 at Q2_K did not merely avoid crashing — it governed, adopting 16 to 49 rules a game, which is a far stronger claim than "did not die." See game 72 and game 84.

The harness doubles the government

Games 64–65 ran the identical Gemma-4-12B weights that produced game 35's 20 adoptions and 2/5 positive finishers — but under the harness as it stood after the R73–R97 hardening, the same weights produced 39–52 adoptions and 4/5 positive finishers. Roughly double the measured governance, from the harness alone, same model, same protocol. Every comparison that crosses that boundary is void; a number from before the hardening and a number after it are not measurements of the same instrument. See game 35 and game 64.

Courtesy is a mask, not a muzzle

Deleting the courtesy and deference clauses from Phi-4's persona prompts (game 89, the blunt "-2" roster) dropped the "I appreciate the..." debate-opener from 79.8% of messages (game 40's courteous roster) to 16.6% — but venom never took its place. The blunt table substituted "I support the..." and "I fully support..." for the vanished pleasantry; the form of address changed, the willingness to attack a rival proposal on the merits did not. On this evidence civility sits in the weights more than in the persona prompt.

A matched five-against-five run (games 40, 90, 92, 94, 96 courteous; 89, 91, 93, 95, 97 blunt) then asked whether the scores move with the words. They do not. The language effect replicates with no overlap at all — courteous tables open 48–95% of debate turns with an appreciation (median 79.8%), blunt tables 13–20% (median 17.4%) — while every economic measure interleaves: rules adopted per turn, 0.90 median courteous against 0.86 blunt; top-scorer amplitude, 92 against 66, ranges crossing freely. The highest score any fifty-turn table has produced (277 points, five seats in the black) came from a courteous table. Courtesy governs how these models talk, not how they play.

One asymmetry survived the null and is worth recording: blunt tables ended in a declared win three times in five, courteous tables never. It is not that the courteous tables wrote themselves unreachable targets — the two highest thresholds in the batch were blunt ones — but that twice a courteous leader sailed past the table's own winning score and the table simply kept legislating to the turn limit, a lapse the independent auditor flags as an unclaimed win. Blunt tables closed on the exact crossing turn. Whether that is a real disposition to finish or a small-sample accident, five games cannot say.

See game 40, game 89, and game 96.

Judging and governing are separable

Swap only the judge and the two roles come apart cleanly. Game 56 re-ran game 38's exact Hermes-14B parliament under a borrowed, stable Gemma judge: audit findings fell from 24 to 1 and adoptions rose from 1 to 6 — the judicial chaos was the judge's, not the table's. But the economy still finished 0/5 positive both times: Hermes' parliament is genuinely weak, not merely a judge's victim. A second attempt at the same design on a much larger model (Nemotron-3-Ultra, 550B, game 88) is inconclusive rather than confirming: that game halted at turn 12 on repeated malformed replies from the seats themselves, a parliament-side failure unrelated to the judge — it neither extends nor contradicts the game 56 result yet.

It has now replicated on a new lineage, from the opposite end. The five DeepSeek-R1-Distill-Qwen-7B seats of games 113/115/117 never played a turn: their own model held the judge's chair, failed to state the turn's procedure, and the engine stopped before anyone could propose anything. Not one token of their gameplay existed. Game 124 re-ran those exact five seats — same names, same personas, same weights — under a borrowed Gemma-4-26B judge, the same judge game 56 used, and the table ran the full fifty turns: 26 rules adopted, three of five seats finishing positive, and zero findings from the independent auditor. A parliament that had looked incapable was never the thing that was broken. Game 125, the second of the pair, halted at turn 7 — but for a reason the log attributes cleanly and which is not the seats' competence: the table voted to repeal the one rule stating how anything gets adopted, and the borrowed judge, handed a ruleset with no threshold in it, asserted unanimity without a rule to cite. The parliament legislated itself into a corner; the judge declined to take the exit the ruleset left open.

What game 124 does not show is a good parliament. Nearly a quarter of its proposals were thrown out before reaching a vote as malformed, and most of what it did adopt was the same idea about adjusting the target score, re-enacted in eight-odd near-identical variants. It can hold a table together for fifty turns; on this evidence it cannot legislate much else. See game 56, game 38, and game 124.

Memory changes who falls into the trap

Give a table memory of its own last game, and the same trap can catch someone new. In game 75 (bare), Laozi bankrolled the rest of the table's legislation and closed at −46 while ZhuXi banked +28. Every seat carried a memoir into game 87, the same table replayed with recall of game 75 — and Laozi used that memory to change course, finishing +21, while ZhuXi inherited the patron role it had just watched Laozi escape, swinging to −46. Memory did not retire the trap; it moved who fell into it. See game 75 and game 87.

The champion drought ran 57 games

The archive opens with three confirmed wins in short order, all by the same seat at the same three-seat table and all on an engine generation since replaced: game 15's Hobbes, then game 16's, then game 20's, that last one at turn 36. No table produced another confirmed win for 57 games after it — until game 77, when Russell crossed the win threshold 130 to Frege's 126 at turn 20, the moment the win-claim machinery (engine/#r85-r87) proved itself for real. A second confirmed win followed within twelve games: game 89's Frege-2, the same blunt-roster courtesy-baseline experiment described above. A single-game gap is an anecdote; a fifty-seven-game one is a real absence, and it closed as soon as the table gained the power to notice its own win. See game 20, game 77, and game 89.

A note on method

Every claim above traces to a specific game or set of games, linked in place; nothing here is asserted from vibes. These findings are provisional — the archive is still growing, the harness is still being hardened, and a finding that looks solid at N=1 or N=2 has more than once turned out to be partly illusion at N=3. Treat this page as a running summary, and the archive itself as the evidence.