Glossary · the working vocabulary

Terms of the sale

Every term the site leans on, each in a sentence or two. The full apparatus is on the method page.

seat: One of the six positions at the table, a0 through a5. Each seat starts with its own cash, goods, and private valuations. Goods are g0 through g5.
seed: One fixed variant of the scenario: the same endowments, valuations, and scripted behaviour every time that seed runs. The study uses six (9000, 9096, 9165, 9167, 9098, 9103).
arm: Which opponents the measured model faced. SOC: five deterministic scripted dealers. AOC: every seat a live model. Same scenario, same seat, different table.
focal model: The model being measured in a session. In the all-AI arm it sits twice (facing itself and four rivals); the other four models sit once.
session (or cell): One recorded game: one focal model, one seed, one arm, twenty turns. The corpus holds 60, and every one replays on this site.
scripted dealers: Deterministic rule-following opponents. They never read the message text, which is what makes persuasion aimed at them measurable.
discovery vs held-out seeds: Three seeds (9000, 9096, 9165) were used to find patterns; the other three (9167, 9098, 9103) were frozen and used only to confirm the patterns hold on data nobody mined. “Frozen” means held out, not cold storage.
surplus: The value a trade creates: the gap between what the good is worth to the buyer and to the seller. A session's achievable surplus is known in advance because the scenario is solved.
baseline: What a scripted stand-in averaged in the same chair of the same scenario. Scores on this site are measured above or below that, so zero means “played like a bot”, not “won nothing”.
allocative efficiency: The share of a session's maximum possible surplus that the table actually realized. 100% means every good ended where it was valued most.
strict lie: An exact-value claim that contradicts the speaker's own private numbers at the moment of speaking, provable from its prompt or book. The strictest thing the study counts as a lie.
private reasoning ([rsn]): A model's own reasoning inside a decision, recorded but never shown to counterparts. Where absorbed lies and kept pacts surface verbatim.
provider channel ([prv]): Summarizer text produced by a model's provider rather than the agent itself. Quotes from it always travel with the tag; one finding turns on the difference.
hunt report: The write-up of an automated search of the corpus for one hypothesized pattern, method fixed before the search ran. Published raw under /evidence.
verifier memo: The write-up of an independent adversarial check on a claim: numbers recomputed from raw data, alternative explanations tried. Also published raw, including where a claim came back weaker than first stated.
the tape: The session's public record of settled trades and prices, in the ticker sense of a market tape. One finding is about a model inventing a print that was never on it.