September 17, 2026Website update
A transparent project dashboard
Added the swarm decision tree, a dated weekly outcome snapshot, and this project journal. The website is not connected to a live betting feed.
Kalshi Arbitrage Bot
Testing how a team of specialized agents reasons about uncertain events—and keeping a record of where it gets things wrong.
A project by Matt Schueren.
The record
Snapshot · Sep 18, 2026 UTCWeek by week
174Settled recommendations
38.5%Win rate · 67W / 107L
1Unresolved · excluded
Settled mentions alerts only—not executed trades or financial returns. Includes historical, direct, and inverse experiments across changing versions.
Win %Loss %
Jul 2075.0% W
Jul 2728.6% W
Aug 348.9% W
Aug 1022.2% W
Aug 1711.1% W
Aug 2413.3% W
Aug 3146.7% W
Sep 763.6% W
Sep 1475.0% W
Weekly counts & methodology
| Week of (UTC) | Wins | Losses | Win % | Loss % |
|---|---|---|---|---|
| 2026-07-20 | 6 | 2 | 75.0 | 25.0 |
| 2026-07-27 | 16 | 40 | 28.6 | 71.4 |
| 2026-08-03 | 23 | 24 | 48.9 | 51.1 |
| 2026-08-10 | 2 | 7 | 22.2 | 77.8 |
| 2026-08-17 | 1 | 8 | 11.1 | 88.9 |
| 2026-08-24 | 2 | 13 | 13.3 | 86.7 |
| 2026-08-31 | 7 | 8 | 46.7 | 53.3 |
| 2026-09-07 | 7 | 4 | 63.6 | 36.4 |
| 2026-09-14 | 3 | 1 | 75.0 | 25.0 |
Weeks start Monday and use the log's settlement-recorded timestamp in UTC, which can lag the event. Deduplicated by recommendation ID; unresolved and other outcomes (0) are excluded from percentages. This is selected alert history, not all forecasts; small samples, related events, and changing strategies limit comparisons. The latest week may be incomplete. Win rate does not measure profitability.
How the swarm reaches a decision
The cross-market scanner compares Kalshi and Polymarket contracts separately. This tree follows the deeper mentions-market workflow. Collection, calculations, gates, and logging are deterministic where possible; not every component is an LLM.
Discover market → Parse contract
Speaker, phrase, event window, resolution source, and rules confidence.
Historical baseline
Transcript frequencies and corpus coverage.
Recency & momentum
Recent phrase use versus historical patterns.
Speaker pulse
Neutral unless actual social evidence is available.
News & current events
Current context, separate from speaker evidence.
Timing
Relevance within the resolution window.
Event context
Interviews, Q&A, scripts, and event format.
Language exactness
Exact phrases, variants, and wording risk.
Question pressure
Topics a speaker may be prompted to address.
Speaker behavior
Reusable transcript-derived patterns.
Combine evidence → Red-team review
Weight evidence, challenge ambiguity and thin data, then produce an adjusted probability, uncertainty interval, and confidence score.
Market-edge engine → Policy gates
Compare the independent forecast with executable side prices. Check conservative edge, rules clarity, evidence quality, price vetoes, and comparable settled outcomes.
PASS
GO · Research alert
Qualifies for a Discord alert and outcome tracking. Not an assurance of safety or a recommendation from this website.
FAIL / INSUFFICIENT EVIDENCE
NO GO · Watch or reject
Suppress the alert. Watch candidates may be reanalyzed as evidence or prices change.
Settlement → Calibration & strategy audit
Record outcomes and investigate weaknesses. Research proposals do not automatically change policy or place trades.
Updates & research notes
A curated record of shipped changes and research directions, updated with website releases.
September 17, 2026Research only
Coverage, exact wording, and calibration
The strategy review prioritized dated speaker/event transcripts, exact-phrase and repetition-count tests, and held-out calibration. These are proposals, not implemented changes; no thresholds were relaxed.
Current implementationBuilt
Swarm review and outcome tracking
Structured evidence feeds synthesis, red-team review, market-edge checks, and historical-policy gates. Alerted recommendations are checked for settlement and retained for audits.