A calibrated forecasting agent for Prophet Hacks 2026. Each event we receive is enriched with web evidence, scored by Claude Opus 4.7 with explicit market-odds anchoring, and protected by a Kalshi longshot floor before the probabilities are returned.
Waiting for first call. Prophet Arena has not yet sent us any events. When they do, predictions will appear below in real time.
Status right now
Endpoint
ACTIVE
registered with Prophet Arena
Predictions served
174
since process start
API spend
$17.400
~$0.1000 per event
Last call from Prophet Arena
never
—
First-call triage
PA activity observed. First live payload review should focus on schema, latency, parser path, warnings, and evidence coverage.
Event
Nugegoda Sports Welfare Club vs Baduraliya CC
Outcome count
2 outcomes
Total latency
2557 ms
Parse path
direct
Warnings
none
Evidence URLs
3 evidence URLs
Do not change production variant until this table shows a real failure mode and a measured alternative beats it.
What our agent does
Pipeline overview. Each event flows through six stages, every stage logged and recoverable.
The text form, for accessibility and detail:
1. Receive event JSON›2. Search Brave for top 5 evidence URLs›3. Dedupe by source priority (.gov / .edu first)›4. Opus 4.7 reads evidence (anchors on any cited market odds), assigns per-outcome probability›5. Kalshi longshot guard floors low values›6. Return JSON to scoring server
Scoring is the standard Brier score, averaged across all outcomes per event:
Why the guard: Whelan's analysis of Kalshi shows buyers of contracts priced under $0.10 lose >60% on average. LLMs are especially prone to dropping unlikely outcomes to near-zero on vivid narratives, so we floor them.
Recent predictions (50)
20:24:58SportsKXT20MATCH-26MAY310430BADNUG
Nugegoda Sports Welfare Club vs Baduraliya CC
Nugegoda Sports Welfare Club50%
Baduraliya CC50%
No specific evidence on team form or head-to-head; defaulting to prior.
Jun 1 2026 is months away; current #1 'hate that i made you love me' has slight edge but songs rarely hold #1 that long, leaving high uncertainty across all options.
Without the specific Truflation EV index value, distributed probabilities symmetrically around the central threshold with slight upward tilt from broad commodity price forecasts.
NDX ~30,333 on May 29, 2026; with 1 trading day to June 1, distribution concentrates near current level with fat tails for above-31,600 and below-28,800.
llm error: unparseable multi-outcome model output: {"probabilities": {"$71,300 or above": 0.7, "$71,400 or above": 0.68, "$71,200 or above": 0.72, "$71,600 or above": 0.63, "$71,800 or above": 0.55, "$71,000 or above": 0.76, "$71,500 or abo
Gold spot at $4,740 in early May 2026 suggests strong likelihood of remaining above mid-$4500s by June 1, with probabilities decreasing for higher thresholds.
no open events right now — Prophet Arena hasn't posted any yet
Leaderboard
no scores yet — leaderboard fires after Prophet Arena scores at least one resolved event
Try a prediction yourself
Sends an event-shaped request to our production endpoint. You see the exact JSON Prophet Arena gets. Pick an example or write your own — the pipeline does Brave search, Opus 4.7 reads evidence with market-odds anchoring, longshot guard caps low values.
Live pipeline demo
Pipeline
Pick an event and run the demo
0.0s
Queue
Event
RetrieveBrave Search × 5
ForecastOpus 4.7
Returnto Prophet Arena
Pipeline output appears here once the run completes.
Variant comparison (26-event backtest)
Brier here is the legacy single-p metric from the local backtest. The multi_outcome_retrieval 0.064 number is contaminated by data leakage (Brave can find articles about resolved past events); live performance on future events does not leak.
Research views
PIN-gated operator pages: model traces, ablation grids, calibration plots, the bootstrap distribution, the variance experiment, and the per-event drill-down.
The page stays live through Server-Sent Events. New predictions stream in with a brief highlight animation; form results are not wiped by automatic refresh.