ORACLE

· pythia · cassandra · sibyl · forecast disagreement

22:01CEST

live

deep-dive · ORACLE · forecast-disagreement + yield · 5 archetypes · branded 2026-04-23back to overview

Oracle · current generation

paper mode · master kill switch via individual archetype WB_ORACLE_*_ENABLE flags

live P&L+$1,420.22across archetypes

archetypes4/5firing

bets opened17291729 resolved

pythia canary22.0%thresh 75.0%

serviceactiveactive

archetypes

each archetype = a Greek oracle · Tiresias reserved

PYTHIA · Q

High-conviction NO

BUY NO when p_mkt − p_pred_blend ∈ [0.20, 0.50] on point/range buckets

Evidence

OOS 96.6% (n=58) · +$0.339/$1

Live hit

51.8%

n=927

Live P&L

+$392

no edge detected

PYTHIA · A

Wide-net NO

BUY NO when p_mkt − p_pred_blend ∈ [0.10, 0.20) on point/range

Evidence

OOS 89.8% (n=171) · +$0.158/$1

Live hit

63.3%

n=738

Live P&L

+$49.46

no edge detected

CASSANDRA

Contrarian YES

BUY YES when p_pred_blend − p_mkt ∈ [0.30, 0.50] on point/range

Evidence

OOS 75.9% (n=29) · +$0.425/$1

Live hit

—

n=0

Live P&L

$0.00

warming

SIBYL

Near-cert yield farm

BUY NO at 1-3¢ when p_mkt ≥ 0.95 on point buckets, book-depth ≥ 100

Evidence

Backtest 96% WR · MARGINAL fee survival

Live hit

35.0%

n=20

Live P&L

+$1,006

no edge detected

ARTEMIS

Longshot huntress (DISABLED 2026-04-24)

BUY YES at 1-5¢ on point/range when p_pred_blend ≥ 3%, ratio ≥ 2× ask

Evidence

ColdMath misclone — clones the LOSING leg only · 0/82 settled YES

Live hit

2.3%

n=44

Live P&L

−$27.49

significant negative (fix or kill)

canary & regime

live hit rate vs training OOS

The Killer Strategy (now Pythia) was validated at 89.9% hit over 258 bets on Feb-Apr 2026 OOS. On older jbecker data (Dec 2025-Jan 2026) it dropped to 68.6% hit — regime-dependent.

Canary auto-pauses Pythia when rolling 50-bet hit rate drops below 75%. Status: TRIPPED (22.0% < 75.0%)

Retrain biweekly via scripts/probe_d_outcome_classifier/save_production_model.py. Model lives at data/probe_d_production/.