Edge discovery
Your model vs the market
Adversarial search tries to kill your thesis before you trade on it. What survives gets classified, sized, and scored for liquidity.
Thesis-implied vs market price
Your causal tree says:
Hormuz closure persists 85%
Oil stays above $100 91%
→ WTI peaks above $150 74% (thesis-implied)
Kalshi says:
KXWTIMAX-26DEC31-T150 YES 38¢ (market price)
Edge: +36¢
Either the market is wrong, or your thesis is.
The runtime figures out which.Every node in your causal tree propagates a probability to mapped contracts. Where that number diverges from the market price, you have a potential edge.
Adversarial search tries to kill it
Adversarial queries (auto-generated from causal tree):
"iran ceasefire negotiations progress 2026"
→ No credible diplomatic path found
"hormuz strait reopening timeline"
→ Insurance companies still refusing tanker coverage
"oil price crash risk factors march 2026"
→ ⚠ US strategic petroleum reserve release possible
Impact: n2 (Oil above $100) drops 91% → 84%
Edge impact: WTI T150 reduces 36¢ → 29¢
Still viable — edge > spread + execution cost
"trump iran deal executive order"
→ No executive action indicatedThe runtime generates search queries designed to disprove your thesis. If a counter-narrative has real evidence, node probabilities adjust and edges shrink. Edges that survive adversarial search are higher quality.
Classify, size, score
Edge report — 12 edges found, 4 actionable
WTI T150 YES ★ actionable
Market: 38¢ | Implied: 74¢ | Edge: +36¢
Spread: 1¢ | Depth: $2,400 | Executable: +35.5¢
Type: consensus_gap
→ Market hasn't priced in Hormuz closure confirmation
Gas $4.50 YES ★ actionable
Market: 14¢ | Implied: 55¢ | Edge: +41¢
Spread: 2¢ | Depth: $800 | Executable: +40¢
Type: attention_gap
→ Low volume, few participants watching this contract
Recession Q3 YES ★ actionable
Market: 35¢ | Implied: 55¢ | Edge: +20¢
Spread: 2¢ | Depth: $1,200 | Executable: +19¢
Type: timing_gap
→ Market will converge but GDP data is 3 months out
Fed Cut Jun YES not actionable
Market: 42¢ | Implied: 48¢ | Edge: +6¢
Spread: 3¢ | Depth: $400 | Executable: +4.5¢
Type: risk_premium
→ Edge too thin after spread. Skip.Four edge types. Consensus gap: market hasn't priced in information you have. Attention gap: few participants watching. Timing gap: market will converge but not yet. Risk premium: market knows but demands compensation. Executable edge = edge minus half the spread.
Track record closes the loop
Track record — last 30 days
Edges tracked: 18
Moved toward thesis: 12 (67%)
Moved away: 6
Best call: WTI T135 YES — predicted 74%, market moved 58→67¢
Worst call: Fed Cut Apr — predicted 65%, market moved 42→38¢
Hit rate feeds back into evaluation prompt:
"Your thesis has correctly predicted edge direction
67% of the time. Weight your current assessment
accordingly."
Weekly: tree augmentation adds new causal nodes
from evaluation insights (append-only, never removes)The system tracks whether edges moved in the direction your thesis predicted. Hit rate feeds back into the LLM prompt — the model learns to calibrate. Weekly augmentation adds new causal nodes the evaluations suggest, so the tree evolves with the market.
Edge discovery tells you where the market disagrees with you and why. What you do next depends on your execution style.