◆ Interactive Demo — Synthetic Data Only ◆
← Back to Portfolio
PROMETHEUS
Quantitative Research System

Independently-derived dispersion methodology with interactive backtesting and AI-augmented daily intelligence briefings. Schema-enforced output, empirical Bayesian confidence calibration, and self-auditing anti-hallucination logic. Runs autonomously via GitHub Actions with daily ingest from public data sources.

Domain abstracted for confidentiality. Architecture, methodology, and visualization shown are representative of the production system; underlying data is synthetic.
Current Regime
LOADED
Since May 22
Term Differential
4.35
Long − Short
P90 Threshold
4.51
504-day rolling
Gap to Trigger
−0.16
16 ticks
Confirmation Rate
35.7%
5 of 14 valid
Dispersion Signal — Live Backtest Synthetic · 4-year window
Fracture Cooldown (days) 15 days
Fracture Events 30 raw · 14 after cooldown
DateShort DispLong DispSpreadThresholdConfirmation
Daily Intelligence Brief Claude Sonnet · Schema-enforced
Regime Assessment
Term differential at 4.35, threshold at 4.51 — 16 ticks from fracture trigger, 75 sessions since last fire. The system has been grinding toward threshold in a tight band for nine consecutive sessions. The spread trajectory is constructive: nine of the last ten observations have printed differentials above 3.00, with today reaching the highest read in the 30-session window. The spring is tightening.
Lying-Signal Audit
CONTAMINATION CHECK: Six negative spread readings between sessions 847–862 (worst: −1.978). These inversions fall inside the 504-day calibration window, compressing the measured distribution and artificially elevating the P90 threshold. Flag as soft contamination — not severe enough to disable the signal, but enough to treat the threshold as slightly generous. The current gap of −0.16 may be closer to true fracture territory than it appears.
Energy State
0.5× standard size — LOADED with one confirmed audit flag (structural inversion within calibration window). Confirmation rate is 35.7% (5 of 14 valid fractures). Sizing must reflect degraded calibration environment. Wait for threshold breach before committing resources.
What Changes the Call
UPGRADE to full size: spread breaches and holds above 4.51 threshold for two consecutive sessions without an active contamination flag. DOWNGRADE to stand-down: spread drops below 3.80 in the next 48 hours (coil failing), OR a new structural inversion event is detected within the calibration window.
Bayesian Confidence Empirical calibration
Prior
Historical confirmation rate across all valid fracture events with 15-day cooldown applied. Updated after each resolved event. Current posterior: 35.7% (5 confirmed / 14 valid).
Confidence by Regime
FLAT N/A — no signal
LOADED 0.5× size · watch
FRACTURED 1.0× · confirm direction
COOLING No new entries
Anti-Hallucination Safeguards
Schema-enforced output prevents the model from omitting required fields or fabricating metrics. Every brief must include: regime classification, lying-signal audit, energy state with explicit sizing, directional vector, and what-changes-the-call conditions. Missing any field triggers a structured retry. All numerical claims are validated against the input data before output — the model cannot hallucinate a spread value or confirmation rate.
Output Schema Enforced on every AI brief
{ "regime": string // LOADED | FRACTURED-ACTIVE | COOLING | FLAT | INVERTED "headline": string // One-line summary with key metrics "lying_signal_audit": string // Contamination + macro + regime transition checks "energy_state": string // Sizing recommendation with justification "directional_vector": string // NONE until confirmed by external signal "trade_directive": string // Actionable instruction or WAIT "what_changes_call": string // Upgrade / downgrade conditions }
Methodology Signal definition
The Equation
Term Differential = Dispersionlong − Dispersionshort
Subtract the short-horizon dispersion metric from the long-horizon metric. Normally positive — far-future uncertainty exceeds near-term. When the differential gets unusually positive (short-horizon suppressed while long-horizon stays elevated), the system is in a "coiled" state. The moment the coil collapses back below threshold is the fracture event.
Threshold Calibration
Rolling 504-observation (≈2-year) window. The 90th percentile of the differential distribution defines the "coiled" threshold. This is recalculated daily and adapts to regime changes in the underlying data. The cooldown parameter suppresses re-triggers: once a fracture fires, subsequent events within the cooldown window are discarded to prevent cluster noise.