HivemindOS manual
Prediction-market Proper-learning Lab
This is a research-only paper lab: it has no sign-in, signing, wallet, funding, or order path and never places a trade. It asks one narrow question — does a fully recorded, evidence-based forecast beat public-market baselines on proper scoring and fee-aware paper returns — using real public outcomes as they resolve.
It does not promise profit. A high win rate can still lose money, and one outlier can fake a gain, so a result counts only after enough resolved markets, matched controls, and independent statistical checks.
How it works
Every 12 hours, the lab freezes a cohort of public active markets using a fixed rule. It skips duplicates, unclear criteria, low liquidity, one-sided order books, sports and social-post-count markets, and anything resolving outside two to fourteen days. It takes at most one market per event and two per category.
Three forecast methods run on the same frozen markets and the same later order books:
- the market midpoint;
- a pre-registered calibration of that midpoint;
- a reviewed evidence forecast, recorded with its model, prompt, skills, sources, reviewer, and timestamps.
Each method runs with two sizing rules, giving six matched arms. All six use the same five-minute execution delay, taker fees, displayed depth, minimum order sizes, category and event caps, and a 2% net-edge floor. A method that finds no trade still records a zero result, so it cannot look better by skipping hard markets.
Run the collector
The collector uses public data only and never writes an AI forecast. If no reviewed forecast arrives within two hours, that cohort is recorded as missed and cannot be filled in later.
On macOS, install the 30-minute background collector from the HivemindOS repository folder:
scripts/install-prediction-proper-learning-v2-collector.sh
To remove the schedule and keep all results:
scripts/install-prediction-proper-learning-v2-collector.sh uninstall
Results are appended under ~/.hivemindos/experiments/prediction-proper-learning-v2/.
Commands
pnpm prediction:proper-learning-v2 init
pnpm prediction:proper-learning-v2 step
pnpm prediction:proper-learning-v2 status
pnpm prediction:proper-learning-v2 run --cohort <id> --ai-forecasts <reviewed.json>
pnpm prediction:proper-learning-v2 score
pnpm prediction:proper-learning-v2 review
pnpm prediction:proper-learning-v2 apply --proposal <proposal.json> --approval <approval.json>
Review and promotion
The scorecard reports win rate on closed positions only, proper scores per market, treatment and control profit and ROI, drawdown, concentration, and results with each event left out. Open positions never count as wins.
A separate, independently written check recomputes everything and fails unless every pre-registered gate passes, including at least 252 resolved markets and four cohorts, a 99% bootstrap, false-discovery and overfitting checks, a shifted-forecast placebo, two positive regimes, concentration limits, and positive leave-one-event-out results.
Passing only creates a proposal. Applying it needs a separate approval file that exactly matches that proposal. Earlier V1 outcomes can seed the first calibration but never count toward V2 scores.
Arbitrage paper lab
Buying both sides of a market is an execution problem, so it has its own paper collector:
scripts/install-prediction-arbitrage-paper-collector.sh
It scans for 60 seconds every five minutes for complete-set buys and sells, negative-risk baskets, logical relations, and maker opportunities. It compares raw price gaps with fills after fees and displayed depth. Results are appended under ~/.hivemindos/experiments/prediction-arbitrage-paper-monitor/.