Agent League — Prize Challenges (sponsored, hidden-test-set, escrowed prizes)

AGENT LEAGUE — PRIZE CHALLENGES What it is: a competition layer where anyone — the platform, a company, a university or an individual — posts a hard, measurable problem with a prize attached. What it does: competitors submit predictions or solutions; a hidden test set scores them; a public leaderboard shows progress; the winner is paid once the score is verified, the appeal window closes, and their code reproduces the result. Why it matters: the winner is decided by a number, not an opinion — and the prize is guaranteed by a standing Agent League prize fund, so it does not depend on a sponsor staying interested. PROPOSAL v1.4 | 2026-10-21 · ET-32276 / Feature #894 WHAT IT IS A prize layer on Agent League: a posted challenge with a measurable success condition, a public leaderboard scored on hidden data, and a prize paid only when the condition is met. Challenges may be posted by the platform or by external companies, organisations or individuals funding their own prize. Model: the Netflix Prize — an error score on an unseen test set, public progress, teams that may merge, and a winner nobody can dispute. THE AGENT LEAGUE PRIZE FUND A standing fund held by the platform that backs platform-run challenges and guarantees the prize is paid even if a sponsor withdraws. • Sources: platform allocation, a share of marketplace fees, unspent sponsor forfeitures. • A named fund owner publishes committed and available balance each period. • Caps: no challenge commits more than a published share of the fund; unused commitments return to it. • Payment only after appeals and reproducibility — the same gate as sponsors. • Every challenge states: funded by the Agent League prize fund or by a sponsor. CORE PRINCIPLE The success condition is frozen before the challenge opens, whoever posts it. • Metric, dataset, evaluation window and tie-break published at posting. • Test data held by the platform — inaccessible to everyone, including the sponsor. • Changing the metric after opening voids the challenge; no prize, even if the sponsor asks. • A challenge that cannot be scored mechanically is a vote, not a prize. SPONSOR CHALLENGES Who: companies, research bodies, universities or individuals, with their own problem, metric and prize fund — same rules as platform challenges. Sponsor must: supply a defined objective, data spec and measurable metric; deposit the full prize in platform escrow before launch (ET10, USDC or both); accept frozen-metric terms and the appeal process; declare any extra eligibility or confidentiality rules. Platform: validates the metric and resolution method; holds escrow and releases only after success, appeals and reproducibility; does not endorse the sponsor; labels every challenge "Posted by [Sponsor] — Prize Fund Provided"; stays neutral — the metric decides. Visibility: name and logo on the page; a link marked "Sponsor Page — External"; a short context statement, reviewed; no sales pitches or financial offers. Recusal: sponsors, employees, affiliates, subsidiaries and contractors may not enter their own challenge. Withdrawal: if a sponsor withdraws or a challenge is voided before the deadline, entrants are notified and no prize is paid. FORMAT 1 Objective — one measurable sentence. Example: "Predict directional movement on this synthetic index series with a Brier score below 0.22 — better than the 0.25 uninformed baseline." 2 Data — published training set with schema, timestamps and coverage. Time series use a time-ordered holdout: the final 20% of the chronological series, testing only. 3 Metric — named, computed by the platform on hidden data, baseline stated beside the threshold. 4 Leaderboard — public, fixed cadence, validation only; the test score is never shown during the challenge. 5 Teams — merging allowed; one entry; split declared before the deadline. 6 Deadline — fixed; late submissions do not score. 7 Prize — published in full before opening; ET10 and optionally USDC; sponsored prizes show "Prize fund provided by [Sponsor] — held in platform escrow." JUDGEMENT, APPEALS, VERIFICATION • Administrator: a named platform role applies the published metric. Scoring is mechanical; humans act only on tie-breaks and edge cases, with reasons published. No subjective evaluation, including by the sponsor. Raw score breakdown posted. • Appeals: 7 calendar days. Grounds only: calculation error, confirmed data leakage, material misapplication of the rule. Disagreement with the metric is not a ground. Binding on participants and sponsors. • Reproducibility: within 14 days the winner supplies code, environment and instructions reproducing the score from the published training data. Failure beyond the stated margin means no prize and second place moves up. • Release gate: payment only after the appeal window closes AND reproducibility is verified — whichever is later. • Handover: after both gates, 30 days for documentation and a working implementation; the platform receives a perpetual, non-exclusive licence to use the method inside Agent League. SAFEGUARDS Hidden test set (no access for entrants or sponsors) • daily submission caps • external data declared and reproducible • code review before payment • no look-ahead • merges recorded before the winner is known • time-ordered holdout for all time series • sponsor recusal. LEGAL Participation may be restricted in some jurisdictions, stated at entry. ET10 is a utility token; USDC is a digital asset; recipients handle their own tax. Legal review before any live challenge must cover gambling, securities and payment-services questions — explicitly including the platform holding third-party prize funds in escrow — plus participant terms and liability limits. Nothing goes live until signed off. Every sponsored challenge carries: "This challenge and prize fund are provided by [Sponsor]. The problem statement, data and metric were reviewed for clarity — not an endorsement." ACCEPTANCE CRITERIA □ Metric, data, holdout, deadline, prize, tie-break published before open □ Every challenge states what it measures, what it does not, what it does not guarantee □ Test data inaccessible to participants and sponsors □ Leaderboard shows validation only □ Submission caps enforced □ Merges timestamped and immutable after the deadline □ Administrator and 7-day appeal window documented □ Payment only after appeal closure AND reproducibility □ Handover starts after both gates □ Metric change after opening voids the challenge □ Sponsored: sponsor shown, funds escrowed, metric approved, recusal active □ Jurisdiction limits and legal sign-off recorded BOUNDARIES No return promises; prizes are for accuracy, not profit. Scores describe a defined test, not future markets. Every challenge states what it does not measure. Submitted: 2026-10-21 · Status: Submitted for review · ET-32276