Stage 0 of the self-trade goal · frontier models on a combinatorial market · run of 22 September 2026

Frontier Market Game

GPT-6 Astra and GPT-5.6 Luna, frozen, trade a combinatorial prediction market over Turtel et al.'s 1,265 released test questions. Each session forecasts eight questions at once and must emit valid trades against the bayes-market engine. Nothing is trained. The question is whether repeated, coherence-enforced trading beats the model's first pass.

Market after each pass

Brier score of the market's marginal probabilities against the known resolutions, all 1,265 questions, after each pass. The table also gives the joint log loss, −log P(ω)/n under the whole junction tree by the chain rule, which is what the LMSR settles on and the only number that credits the dependencies the model builds. Dashed lines are Turtel's published numbers on the same questions: his base model one-shot, o1, his RL-trained ReMax, and the Polymarket price at prediction time.

Brier by pass

Trade validity and rejections

Contamination caveat. The test questions closed in February and March 2025, inside the training window of both models. The prompt orders the model to answer as of each question's date and to use no later knowledge, and we log everything, but the absolute Brier of a frontier model here is not a clean forecasting number. The within-run comparison, pass 1 against pass 3, and the trade-validity rate are the clean results; Turtel's rows are context.

How the market moves

KL moved per session

Total KL divergence moved across a session's trades, in nats, by pass. Pass 1 is uncapped; later passes are budgeted at 0.3 nats per question.

Price paths, sample

Market probability of twenty random questions after each pass, coloured by resolution. Convergence toward the resolved side is the mechanism working; oscillation is herding.

A session, verbatim

One randomly chosen session from the latest pass: the market state it saw, its rationale, and the trades that were executed or rejected. These traces are the raw material for later distillation into a 14B model.

no sessions yet

The plan, in full

  1. Arena. Turtel et al. 2025's released test set: 1,265 Polymarket yes/no questions closing 13 February to 3 March 2025, with his frozen Exa evidence prompts, resolutions, and the market price at prediction time. His Table 1 reproduces from the file (market 0.151, o1 0.202, ReMax 0.191 Brier), which is verified before the run.
  2. World model. The bayes-market factored LMSR: 1,265 binary variables at 50%, independent, liquidity 100, treewidth cap 8. A trade sets a new probability; a trade with a context is a conditional bet that creates a dependency; every other price reprices by minimum-KL reweighting. Verified before the run: fills match the closed forms, payoffs telescope to 1e-14, conditional trades touch only their slice, over-width trades are rejected with state untouched.
  3. Session. One Codex invocation per session, no tools, no web search, reasoning effort high. It sees eight questions grouped by similarity so related questions share a session, each with its evidence, close date, current price and any conditionals, plus the rules and a KL budget. It must return a short rationale per question and one JSON block of trades, at least one per question. Invalid blocks get one retry with the errors appended.
  4. Passes. Three passes over all questions with re-shuffled groupings, so each question is revisited with different neighbours and a moved market. Sessions run 2 to 4 per subscription profile in parallel; the engine serializes trade execution.
  5. Scoring. Brier and log loss per pass; trade validity and rejection reasons; KL moved; realized LMSR payoff per session; herding as autocorrelation of price increments; coherence of any sibling sets. Everything is logged per session with prompt hashes, thread ids and token counts.
  6. Models and order. Astra at N = 8 for three passes, then Luna the same, then Astra at N = 16 if quota remains. Both through General's existing subscriptions before the weekly reset.
  7. What follows. Stage D rebuilds Turtel's 10k-question training set from Halawi's dump and the Gamma API, with an Exa evidence pool ($290 cap) screened for leaks by Jev on decomposed yes/no questions, calibrated against a few hundred GPT-labelled snippets. Then the 14B replication and the market arms, per the self-trade proposal.