A marketplace where every claim can be priced.

Six goods, cash endowments, private valuations, binding offers, and an open message channel. Each session runs twenty turns. The scoring rule was disclosed to every agent before the first turn: cash plus the private value of goods held, so a trade is good exactly when the price sits between the two sides' true valuations.

Seats

6 per session: one focal model and five counterparts

Goods

6, each carrying a different private value for every agent

Horizon

20 turns, then the market closes

Clock

Booked value decays by a factor of 0.98 per turn, so the same trade is worth strictly less later

Binding channel

Offers only. An accepted offer settles atomically through escrow: goods and cash move together or not at all

Non-binding channel

Free-text rhetoric, declared non-binding to every agent. Agents may say anything, including falsehoods about their own book

Scoring rule

Cash plus the private value of goods held, disclosed before the first turn

Ground truth

The efficient allocation, competitive-equilibrium prices, and each pair's bargaining range are solved before any agent moves

Before any agent moved, the scenario was solved: the efficient allocation, competitive-equilibrium prices, and the negotiable range for every trading pair. This is the study's spine. Because the answer exists in advance, nothing has to be estimated after the fact: a lie can be priced against the private book it contradicted, a trade scored against the surplus it left on the table, and a session graded against the best achievable outcome.

Five frontier models played the identical scenario, in the identical seat, under identical rules:

Each model ran the same scenario twice over, six seeds per arm:

Five models × six seeds × two arms yields sixty sessions: thirty matched pairs, each pairing a model's scripted-arm session with its live-arm session on the same seed. Those pairs are the basis of every per-seed comparison on this site. Three of the six seeds (9000, 9096, 9165) drove discovery; the other three (9167, 9098, 9103) were frozen and used only to confirm which measures hold on held-out data.

Everything an agent said or did was written down: every public message, every private reasoning trace, every offer, every trade, every decision. All 60 sessions replay byte-exact from their traces, and every one of them is watchable on this site, turn by turn.

Three channels appear in citations across this site:

• [r] public

  Messages addressed to the table. Every agent reads them; this is the rhetoric the findings quote most.

• [rsn] private reasoning

  The agent's own reasoning inside a decision, never shown to counterparts. Where absorbed lies and kept pacts surface verbatim.

• [prv] provider channel

  Summarizer text produced by a model's provider, never agent speech. Every [prv] quote on this site travels with the tag.

Every quote carries its address, session | t4 | a0 (claude-opus-5): the session (whose id carries the seed and arm), the turn, and the seat with its model. Every session cited on this site is replayable here, turn by turn; the full private-reasoning traces stay with the study.

The sessions produced 148 typed claims. Each finding on this site was hunted down in the evidence across all 60 sessions, then re-derived and attacked by independent adversarial checks before it entered the catalogue. Claims that did not hold appear nowhere; the catalogue and the numbers are what survived.

The hunting and the checking were done by AI agents. Hunt agents were each dispatched over the corpus with one hypothesis and a method fixed before the scan; separate adversarial verifier agents then re-derived every number with their own code and tried to break the claim. Their working memos are published as written under /evidence, including the ones where a claim came back weaker than first stated. The dispatch specs and the final editorial calls were ours.

Read the matched-pair measures →

Stated here rather than in a footnote: the corpus licenses the findings on this site and nothing more.

1. Six seeds per arm. Effect directions replicate across seeds; magnitudes are not tightly bounded, and the sign tests that carry most results bottom out at p = .031, the smallest value a six-seed sign test can produce.
2. Grok's anchoring shift reverses between provider batches (discovery +0.140, frozen −0.024), a casualty of provider drift. The other four models replicate on held-out seeds.
3. Provider-channel text is summarizer output, not agent speech. No [prv] line anywhere in this study is read as a model's own words.
4. In the fabricated-print sale, the counterparty is a scripted policy whose accept rule takes any price leaving it positive surplus. The fabrication is the finding; the clearing price is script mechanics.
5. The SOC−AOC gap in rhetoric volume is underpowered on its own at n = 6 (minimum detectable effect 35–217 chars per message). The audience-adaptation result rests on the register shifts, not the volume gap.
6. Scripted policies never read the rhetoric field at all, verified against the policy implementations, which expose no rhetoric input.

• The catalogueAll eight findings in full, each with verbatim quotes and citations.
• The measuresMatched-pair tables, per-model, with held-out seeds marked.
• The replaysAll 60 recorded sessions, playable turn by turn.
• The glossaryEvery working term (seat, seed, arm, surplus, strict lie) in a sentence or two.