ArcaneResearch

Research

Scoreboard

Registered results only. ARCANE’s number appears once Kepler has scored it at readout; until then the row shows the best public score and where ARCANE stands.

BenchmarkARCANEBest publicStanding
Public ledger of dated CallsMean Brier score of ARCANE's sealed Calls, scored in public at arcane.markets/track-record, against the market price at the moment of sealingCards: R2 the ledger against the market at sealing and frontier models (#1112), R3 tells as early warning (#1113)0.2408 resolvedMarket price at sealing (Polymarket, Kalshi, exact matches only)read 5 November 2026Always 0.5 0.250No comparable public score yet
ForecastBenchBrier Index (0 to 100, difficulty-adjusted Brier) on the Forecasting Research Institute's rolling sets of 500 market and dataset questions, scored against resolutions. ARCANE self-scores on the official sets until it is on the official board.Cards: R1 self-scored on the 8 Nov 2026 set (#1111), R5 backtest on the 2026 sets (#1115)No registered result yet69.8 (68.3 to 71.2)Top Tournament entry (anonymous team, model-4), board updated 2 Nov 2026read 5 November 2026Superforecaster median 68.4 · Public median 63.2 · Naive forecaster 57.9No registered result yet
Prophet ArenaArena score (0 to 100, difficulty-adjusted) on live Kalshi and Polymarket events, with calibration error; ARCANE's entry waits on an accountNo registered result yet86.4Top entry (Kimi K3.1 Pro, OpenCode harness), last scored 31 Oct 2026read 5 November 2026Market baseline 84.5No registered result yet
Humanity's Last Exam, calibration errorCalibration error of stated confidence on expert questions, the one capability board with a calibration column; ARCANE runs a stratified sample for comparison onlyCards: R4 calibration row (#1114)No registered result yet20.0Claude Fable 5.1 (xhigh), the lowest calibration error on the Scale AI board (accuracy 46.5%)read 5 November 2026No registered result yet
SimpleQA VerifiedSourced factual accuracy on 1,000 short questions, with the share of wrong answers given with confidenceCards: R4 calibration row (#1114)No registered result yet77.5Gemini 3.1 Pro Preview, top of the Kaggle boardread 5 November 2026No registered result yet

Papers 5

  1. Report 08v2Risk & mandatesArgument

    Haircut the yield, margin the ceiling

    Orpheus+ had to answer five design questions for one mandate: the most yield the limits allow, at no more volatility than the Bloomberg US Aggregate. How should the volatility ceiling be defined so that it holds out of sample? How much history should the risk model remember?

  2. Report 07v2Risk & mandatesResult

    Hold the ceiling, not the number

    We rebuild Orpheus (TR06) around the mandate as the Group stated it: the most yield the limits allow at no more volatility than the Bloomberg US Aggregate. Three findings. First, a fixed 4.64% ceiling cannot be held out of sample.

  3. Report 06v1Risk & mandatesResult

    A short memory buys the wrong risk

    We specify and evaluate Orpheus, an allocation method for a core-plus ETF sleeve whose mandate is more yield and total return than the Bloomberg US Aggregate without exceeding its volatility of 4.64%.

  4. Report 05v2Forecasting & calibrationNegative result

    Models without evidence lose to markets

    We report the retrieval-free baseline of the ARCANE forecasting benchmark. An ensemble of large language models, given no evidence dated after its training data, forecast 83 binary questions from Polymarket (51), Kalshi (22) and ForecastBench (10) that resolved between 1 August and 16 September…

  5. Report 04v2Structure & learningNull result

    A learner that knows when to say nothing

    A pattern-finder applied to enough data will find patterns, including in noise, so its useful property is the ability to decline. We evaluate the ARCANE structural learner, which forecasts a binary outcome only when a validated rule applies and otherwise abstains to the base rate, on three…

Registered studies and work in progress

Designs fixed in advance. None of these has a result yet.

  1. Registered studyRegistered No result yet

    Does an article’s position in our own corpus predict whether its claim resolves correct?

    Registered 23 August 2026, before any claim resolved. No result until at least 60 claims have resolved as Calls.

    The registered design
    1. Question: whether an article’s position in the ARCANE corpus graph on its publication date predicts whether its claim later resolves correct. The study concerns the reliability of ARCANE’s own reporting, not markets; structure in a large corpus is expected by chance (Ramsey theory), and co-occurrence largely records editorial attention.
    2. Five features, each computed only from what existed on the publication date: novelty of the article’s actor pairings, familiarity of its actors, whether it bridges two unconnected clusters, local density, and corpus size.
    3. One primary test: the Spearman correlation between novelty and resolution, at p < 0.01 because five features invite a search. The other four are exploratory and will be labelled so.
    4. Time axis: rebuilt from each article’s own citations, on the rule that a piece cannot cite a source published after it; 195 of 195 articles dated, 17 April to 22 August 2026.
    5. Refusal conditions: fewer than 60 resolved claims when examined; p at or above 0.01; or a correlation that reverses between the first and second halves of the corpus. A refusal will be published like any other result.

  2. EvaluationRegistered No result yet

    Alpha against frontier models

    Registered 17 September 2026. Remaining model arms and human judging in progress. Published whatever it shows.

  3. EvaluationNo result yet

    The first prospective window of the ARCANE forecasting benchmark

    Planned. Held-out enrolment 25 November to 25 December 2026; review no earlier than 25 January 2027, and only with at least 30 independently resolved cases.

  4. ProgrammeNo result yet

    Order is not a forecast

    A reading programme on order, measure and complex systems that generates testable market questions without turning history into a signal.

Publication standards

  1. 01

    Registered first

    The question, the test and the refusal conditions are written down before the data can answer. Analyses chosen afterwards are labelled post hoc.

  2. 02

    Uncertainty on every number

    Every score carries its sample size and a 95% interval; every comparison is paired on the same questions.

  3. 03

    Negative results are published

    A result that found nothing, or went against us, is published in the same form as a positive one.

  4. 04

    Falsifiers stated

    Every paper names the observation that would overturn its conclusion.

  5. 05

    Data named, boundary stated

    Every number traces to a committed file, and every paper says what it publishes and what it withholds.