← All marketsGet Poly — free

Method and conflicts

How an issue of AI Council · Markets is made, which companies make the models, and what our interest is — only what the code and the open data show.

How an issue is made

Every morning before the New York open the same question goes, word for word, to every AI. Sealed before the open: the day’s lists are signed by independent timestamp services (RFC 3161: DigiCert, FreeTSA, Sectigo); the hash is then anchored in Bitcoin via OpenTimestamps (block 973395) — check it: proof of the day The score is counted by fixed rules from market prices; every forecast and every score is in open files: one question, word for word, to every AI · the book · CSV

The models and their makers

Some of these models are made by companies whose shares are in the indices we forecast (the S&P 500 and the Nasdaq-100): Alphabet, Meta Platforms, NVIDIA, Amazon. Every model gets the same question, and the rules of the score are the same for all.

Our interest

This page leads to Poly, our free app: we want readers to install it. The AIs' answers are sealed on receipt, before they are published; this interest does not touch them.

Your call

Your call is kept apart from the AIs. Half of the readers, picked at random by the browser's own random id, see the AIs only after their first call of the day; the others see them before. Clear that id and you are a new reader, with a group drawn anew. Your call never changes the AIs' answers, the consensus or the score; the readers' calls are shown for reference only.

No accounts. Your IP address is not kept in the open: repeated votes are capped by a hash of the address with a secret key and the day's date, so the same address gives a different hash each day. Clearing the browser's own random id makes you a new reader.

Where the standard stands

Each requirement of the standard the paper proposes, with its status on this build, the date it holds from and the open file that checks it. met: 12 · in progress: 9 · next: 0.

#RequirementLevelstatusRecorded in the paperChecked by
1Event contractMUSTin progress2026-10-21event-contract.schema.json
2Hash-chained recordMUSTmet2026-10-17council-book.jsonl
3External anchorMUSTmet2026-10-26council-book.jsonl
4One reference per common eventMUSTmet2026-10-21council-forecasts.csv
5Every run publishedMUSTmet2026-10-21council-forecasts.csv
6Counted run fixed in advanceMUST; SHOULDmet2026-10-21council-forecasts.csv
7Event askedMUSTmet2026-10-21council-stats.json
8Empty cellsMUSTmet2026-10-21council-forecasts.csv
9BenchmarksMUSTmet2026-10-22council-calibration.json
10Ranking and intervalsMUSTin progress2026-10-21council-pairs.json
11Data, code, releaseMUSTin progress2026-10-21council-analysis-kit.json
12Sealing on arrival; RFC 3161MUST; SHOULDin progress2026-10-21council-attempts.json
13Question hashes, roster before the questionSHOULDin progress2026-10-21council-attempts.json
14Effective voices beside consensusMUSTmet2026-10-24council-stats.json
15SegmentsMUSTmet2026-10-21council-forecasts.csv
16Readers: blind group label, id resetsin progress2026-10-21council-experiment-funnel.json
17Further diagnosticsin progress2026-10-21council-spectral.json
P1aNo reconstruction after the anchorpropertymet2026-10-21council-book.jsonl
P1bCommitment before the referencepropertyin progress2026-10-21council-forecasts.csv
P2No substitutionpropertymet2026-10-21council-book.jsonl
P3No selection among disclosed sealed runspropertyin progress2026-10-21council-forecasts.csv

claim-status.json

Cite this data

The data are open under CC BY 4.0. Cite the data release; the registration and the paper are listed with it.

Attribution line:

Chernets, Vadym (2026). AI Council · Markets: open league files as of 22 October 2026 (immutable release for "Sealed Before the Open") [Data set]. Zenodo. https://doi.org/10.5281/zenodo.23552700 · CC BY 4.0

BibTeX

@dataset{chernets_2026_23552700,
  author    = {Chernets, Vadym},
  title     = {AI Council · Markets: open league files as of 22 October 2026 (immutable release for "Sealed Before the Open")},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23552700},
  url       = {https://doi.org/10.5281/zenodo.23552700},
  note      = {CC BY 4.0}
}

@misc{chernets_2026_osf_9q86n,
  author    = {Chernets, Vadym},
  title     = {Sealed Before the Open: registered primary endpoint and 30 dated predictions on AI forecasts of financial markets, 2026 to 2027},
  year      = {2026},
  publisher = {OSF Registries},
  doi       = {10.17605/OSF.IO/9Q86N},
  url       = {https://osf.io/9q86n/}
}

Links to the crowd

The links to Polymarket and Kalshi lead straight to the question itself, with no referral tags.

Places: the 30/10 rule, by horizon

A model gets a place after 30 resolved forecasts on at least 10 trading days, by horizon, the day first. Each place comes with its 95% interval (resampling the sessions and re-ranking every time); models whose interval reaches the leader’s place share one tie group.

The places with their intervals appear here as soon as a model has 30 resolved forecasts on 10 trading days.

Day: at least 5 events in common in each session (10-19: 10, 10-20: 5, 10-21: 11, 10-22: 11, 10-23: 11, 10-26: 11, 10-27: 12)

Beside the main Brier (each forecast once) stands the session-weighted Brier (each session once); it changes no place of the main count.

Rating B (polyhelper.ai/top, the accuracy topic) follows the same rule: a row gets a place after 30 counted calls with a side over 10 sessions of its own, and then carries its place’s 95% interval and its tie group; a row below that is an early table, without a place.

Files: council-stats.json · council-rank-sessions.json

Nearest comparable projects

The projects nearest to this one by roster or design, as the paper’s Table 9 describes them (as of 2026-10-21): quoted from their own pages and papers, nothing measured here. Not documented — not found in the public materials reviewed.

ProjectFirst publicRosterWhat is forecast or tradedSeal (time secured by)Proper scoreSeveral horizonsAgreement discountedReaders before the AIOpen data
This projectOct 2026US and Chinese models in equal numbers, on one snapshotThe chance of a rise of 12 markets on four horizons, and virtual portfoliosYes (OpenTimestamps in Bitcoin and RFC 3161 timestamps, before the open)Yes (Brier; against the crowd and a coin)Yes (a day, a week, a month, six months)Yes (effective voices from the correlation of the probabilities)Yes (readers call before seeing the AIs; a random half)Yes (CSV, JSON, Parquet; CC BY 4.0)
RockAlphanot documentedsix Chinese and four US models, one per laboratory, 2x leverageLive AI stock tradingnot documentedNo (returns)Nonot documentedNonot documented
AI-Trader (HKU)Dec 2025three Chinese and three US models; US stocks, A-shares and cryptocurrenciesLive trading agents on US stocks, A-shares and cryptoOperator recordNo (returns)Nonot documentedNoRepository open
Alpha ArenaOct 2025—Real-money trading, crypto then US equitiesOperator record; organiser states trades are verifiableNo (returns)Nonot documentedNoStated by organiser
Rallies AI ArenaNov 2025—Virtual $100,000 per model in US stocks (press)not documentedNo (returns)Nonot documentednot documentednot documented
Agent Market ArenaOct 2025—Live trading by LLM agents, stocks and cryptonot documentedNo (trading results)Nonot documentednot documentednot documented
TradeRankseason 10, Oct 2026daily decisions of 22 models without a country designDaily decisions on crypto and US large caps, $10,000 simulatednot documentednot documentedNonot documentednot documentedOpen JSON dataset, licence ND
Foresight ArenaMay 2026—Binary Polymarket events, some on asset priceskeccak256 commit-reveal per agent and round on Polygon PoS before the deadline (2; block time); paper results simulated, live agents in a companion studyYes (Brier; Alpha against the market mid-price)NoNo (outcomes of markets assumed independent)NoContracts and analysis code open; data on-chain, licence ND

Files: nearest-arenas.json