Two research artifacts from the team behind the OracleMangle scorer, released for anyone to use, fork, and extend: a lab of 11 candidate Polymarket trading bots with their full backtest record, and a 96 GB market-data tape free under CC-BY-4.0. No profit claims — the confidence intervals and failure notes ship on purpose.
A lab of 11 candidate Polymarket trading bots plus the research framework they run on: CLOB clients (V1 + V2 era), fee models, a backtest engine, portfolio and risk gates, tests, and a fleet dashboard. It is a sanitized export of a 2026 multi-bot research project — built as starting points to refine, not a productized live-trading system.
Two strategies showed positive expectancy in testing — promising, unproven, with the caveats published alongside. The rest are documented, runnable candidates with real backtest or paper/live-probe data and clear notes on what still needs refinement before any edge is claimed.
Buy NO when sports YES is priced ~55–75c, typically 6–78h before close. Tested: n=573, 57.4% hit rate, post-fee ROI +5.32% — but the bootstrap 95% CI [−4.73%, +15.64%] crosses zero. A research candidate, not a proven edge.
Fade mispriced temperature-range markets with disciplined filters. A tiny live probe of 95 closed groups — the only positive live P&L sample in the fleet record — returned +$31.06 / +11.07% ROI. Scale was gated on open accounting questions.
The other nine ship with a one-line status and a refinement direction; several are explicitly labelled unprofitable as-tested, with the reasons documented:
| Bot | Strategy | Status as-tested |
|---|---|---|
| Longshot | General binary NO tail-fade | Unprofitable on walk-forward; jackpot losses ate the grind |
| Pythia | Pyth/GBM directional on crypto and indices | Set aside pending model and data-feed refinement |
| Scribe | Near-resolution crypto scalp research | Paused pending fill-realism and latency work |
| Sonar | Crowd / whale-sensor experiments | Historical sensor path; needs reframing as cascade-detector input |
| Dash | Crypto near-resolution longshot family | Unprofitable on paper/live probes |
| Quill | Maker / rebate research | Gated behind fill-conditioned replay; fill realism unproven |
| Relay | Neg-risk / wallet-linked paper starting point | Needs wallet-signal quality and risk gates |
| Compass | Complete-set style probe | Documented starting point; arb economics unproven under fees |
| Meridian | Crypto fair-value / probability-gap quoting | Unprofitable in live/paper samples |
A continuous market-data tape for microstructure and lead-lag research: co-located CEX trades and Polymarket market-channel WebSocket events recorded over fixed UTC windows. Free to download, remix, and republish with attribution.
| Contents | Scale |
|---|---|
| CEX trades (BTC, ETH, SOL) | 271,566,165 rows |
| Polymarket WS events (book snapshots + price_change deltas) | ~61M rows across tapes |
| Market metadata snapshots | 23,750 markets |
| Coverage | 2026-04-30 → 2026-07-05, three tapes, ~96 GB raw / ~9 GB compressed |
| Dual vantage | 2026-06-03 → 2026-07-01 overlap: same markets, two independent receive clocks — for latency and feed-consistency studies |
Format is zstd-compressed SQLite with a shared schema; parquet preview samples render directly in the Hugging Face Dataset Viewer so you can browse before downloading.
The Oracle bot's edge candidate came from scoring market questions for dispute risk before taking a position — ambiguous wording, weak source rules, and resolution-risk patterns. That scorer is OracleMangle's paid API. The open repo shows exactly how a bot consumes it:
curl -H "X-API-Key: YOUR_KEY" \
"https://api.oraclemangle.com/v1/score?question=YOUR_MARKET_QUESTION"
# example response shape
{"dispute_risk": 0.05, "resolution_prediction": "YES",
"confidence": 0.95, "safe_to_trade": true, ...}One HTTP call flags resolution-risk patterns a backtest can't see.
Set ORACLEMANGLE_API_URL and ORACLEMANGLE_API_KEY — the integration surface documented in the repo.
Stripe provisions your API key and tier limits automatically after checkout.
Two of the 11 candidates (Playbook and Nimbus) showed positive expectancy in testing, but neither claims a proven edge: Playbook's bootstrap 95% confidence interval crosses zero, and Nimbus's positive live sample is tiny (95 closed groups). Several others are explicitly documented as unprofitable as-tested. The repo publishes confidence intervals and failure notes on purpose.
Yes. It is released under CC-BY-4.0 on Hugging Face: 271,566,165 CEX trades (BTC, ETH, SOL), roughly 61 million Polymarket WebSocket events, and 23,750 market metadata snapshots covering 2026-04-30 to 2026-07-05, as zstd-compressed SQLite tapes with parquet preview samples. Attribution is the only requirement.
The Oracle bot (bot_b) in polymarket-bot-lab consumed OracleMangle's dispute-risk scorer over HTTP before taking positions. The scorer analyzes market question wording for ambiguity, weak source rules, and resolution-risk patterns, returning dispute_risk, resolution_prediction, and confidence via GET /v1/score. The bot code is reference-only in the repo; the scorer is OracleMangle's paid API with Stripe-provisioned keys.
The GitHub repository ships CITATION.cff metadata for polymarket-bot-lab (Apache-2.0), and the Hugging Face dataset card covers polymarket-canary-tape (CC-BY-4.0). Cite the artifact name, the hosting URL, and oraclemangle.com as the source project.
Research and engineering material, not financial advice. Nothing on this page or in the linked repositories is a guarantee of profit; strategies that look positive in a sample may not survive fees, adverse selection, or regime shift. Past results are not future results.
Questions: [email protected].