Observatory

Live state

First-call watch

Failure mode	Status	Operator response
Brave degraded	monitor via full check	Verify fallback path, do not change model first.
Metric drift	known caveat	Keep single-binary and proper multi-class labeled separately.
Commit drift	covered	Compare this page's commit with origin/main before deploy claims.

Recent persisted predictions

Market	Event	Probabilities	Latency	Parse path	Warnings
KXT20MATCH-26MAY310430BADNUG 2026-06-01T20:24:58	Nugegoda Sports Welfare Club vs Baduraliya CC Sports	Nugegoda Sports Welfare Club: 0.50, Baduraliya CC: 0.50	2557 ms	direct	—
KXT20MATCH-26MAY310015ACESIN 2026-06-01T20:11:03	Sinhalese Sports Club vs Ace Capital CC Sports	Ace Capital CC: 0.55, Sinhalese Sports Club: 0.45	3493 ms	direct	—
KXHIGHLAX-26JUN01 2026-06-01T20:01:15	Highest temperature in LA on Jun 1, 2026? Climate and Weather	76° or above: 0.15, 74° to 75°: 0.17, 70° to 71°: 0.23	3438 ms	direct	—
KXHIGHNY-26JUN01 2026-06-01T19:46:03	Highest temperature in NYC on Jun 1, 2026? Climate and Weather	69° or below: 0.13, 72° to 73°: 0.18, 70° to 71°: 0.16	4528 ms	direct	—
KXAAAGASD-26JUN02 2026-06-01T19:36:20	US gas prices on Jun 2, 2026 Economics	Above 4.295: 0.12, Above 4.305: 0.11, Above 4.290: 0.12	4134 ms	direct	—
KXSPOTIFYGLOBALD-26JUN01 2026-06-01T19:30:46	Top Global Song on Spotify on Jun 1, 2026? Entertainment	SWIM: 0.06, Beat It: 0.06, hate that i made you love me: 0.13	6052 ms	direct	—
KXSPOTIFYD-26JUN01 2026-06-01T19:25:28	Top USA Song on Spotify on Jun 1, 2026? Entertainment	Choosin' Texas: 0.05, hate that i made you love me: 0.12, Janice STFU: 0.06	4050 ms	direct	—
KXTRUEV-26JUN01 2026-06-01T19:18:39	EV commodity prices on Jun 1, 2026 Economics	Above 1257.28: 0.50, Above 1247.28: 0.55, Above 1277.28: 0.42	4498 ms	direct	—
KXSILVERD-26JUN0117 2026-06-01T18:58:40	Silver price on June 01, 2026 at 5:00 PM EDT? Commodities	above $75: 0.82, above $75.50: 0.70, above $76: 0.45	6770 ms	direct	—
KXNATGASD-26JUN0117 2026-06-01T18:38:42	Natural gas price on June 01, 2026 at 5:00 PM EDT? Commodities	above $3.340: 0.05, above $3.370: 0.05, above $3.365: 0.05	10064 ms	—	llm call failed: unparseable multi-outcome model output: {"probabilities": {"above $3.340": 0.25, "above $3.370": 0.22, "above $3.365": 0.23, "above $3.320": 0.28, "above $3.180

Experiment board

Question	Status	Current answer	Next action
Why not GPT-5.5?	tested	GPT-5.5 was tried. All-event single-binary Brier is 0.0920 vs Opus 4.7 at 0.0378; its binary-only subset is strong, but that is not the full endpoint metric.	Keep as ablation, not production.
Why current model?	measured	Best verifiable PA CLI score on current sample; Opus 4.6 is close and slightly better under proper multi-class, so this is not a settled universal claim.	Re-evaluate after live PA calls.
Prompt ablation	in progress elsewhere	Production, no-anchor, no-scale, minimal, and meta-role variants should be tracked here once finalized.	Import final JSON, cost, and caveats.
Retrieval count	not settled	Current count is based on prior plateau guidance, not yet independently optimized on our data.	Run 0/3/5/8 with same model and prompt.
Scoring method	corrected	PA CLI single-binary and proper multi-class must both be labeled. Do not mix rows across metrics.	Keep public claims conservative.

Questions to expect

• What is on the public root? Brand, status, links to the report PDF and dashboard. The exact prompt, model rationale, and ablation tables sit on this console.
• Is the 26-event backtest enough? No. It is small, sports-heavy, and contains some post-resolution retrieval leakage (audited at 38.5%). Useful for direction and regression catching, not as proof of live performance.
• What to check on the first PA call? Event schema, outcome count, total latency, parser path, warnings, and whether the trace matches what the dashboard demo predicted it would.
• What is the no-touch list? Production variant, forecast model, longshot-floor formula, Railway env vars. Anything else can change after measurement.

Public surface