The complete 14-layer investment intelligence architecture. Every signal Ponelope generates passes through all 14 layers in sequence. No shortcuts. No bypasses. Each layer is independently auditable.
← Back to Proof of GovernanceRaw financial data is collected from external sources: Finnhub (earnings calendar, company news, real-time quotes), NewsAPI rotation, Binance microstructure, founder intelligence PDFs, and Polymarket prediction markets.
Computes a content checksum over each ingestion envelope. If an identical event was already processed (same source + same canonical fields), the pipeline stops for that event.
Extracts quantitative claims from the ingested text (EPS figures, price percentages, revenue numbers). Each claim is cross-checked against authoritative numeric data from Finnhub quote and earnings endpoints.
Transforms the anchor-verified ingestion envelope into a structured ResearchEventCard: ticker, event type, claimed direction, key facts, and an initial qualitative assessment.
Applies quantitative sentiment scoring to calendar events (earnings direction, magnitude) and news text (keyword / phrase weighting). Produces a preliminary directional score before full thesis generation.
Synthesises the research card, sentiment scores, and regime context into a formal investment thesis: a direction (BUY/SELL/HOLD), an invalidation condition, and a structured rationale.
A dedicated adversarial pass generates the strongest possible case against the thesis. Forces the system to articulate why it could be wrong before any conviction is finalised.
Validates the thesis against cross-asset signals: bond yield direction, USD trend, VIX level, sector rotation indicators, and crypto flow. Checks for contradictions between the equity thesis and macro context.
Builds a regime snapshot from cross-sectional quote data: vol_percentile (cross-sectional volatility rank), trend_score (fraction of symbols trending), and breadth_score (advancing vs declining). Incorporates geopolitical risk, rates impulse, and prediction market scores.
Five independently configured analyst profiles evaluate every thesis simultaneously. Conservative (risk-first, base score 0.4 — hard to convince, requires CLEAN anchor and low crowding), Momentum (trend and structure — heavily rewards regime tailwind and cross-asset confirmation), Contrarian (adversarial mandate — always finds the counter-argument, rewards data-narrative divergence, penalises consensus narratives), Macro (regime and cross-asset alignment — RISK_ON +0.20, RISK_OFF -0.30, penalises single-asset thesis without macro context), and Governance (rule enforcer with veto power — checks hard governance constraints: anchor CONFLICT +0.40 risk, crowding >0.80 +0.35 risk, regime OPPOSING +0.30 risk). Each analyst votes independently with documented reasoning: BUY / SELL / HOLD / VETO. The Governance analyst can veto any decision with risk_score >= 0.80, which hard-blocks advancement regardless of the other four votes. Consensus: 3+ supporting (support_score >= 0.60) = CONSENSUS, exactly 2 = SPLIT, fewer than 2 or any veto = CONTESTED.
Applies the full governance rule set (~16+ rules) against the decision packet. Checks position concentration, sector exposure limits, max conviction thresholds, maturity-level restrictions, and regime-specific trading bans.
Synthesises all upstream signals into a final conviction score (0–1). 8 dimension inputs: anchor quality, sentiment strength, thesis clarity, counter-thesis penalty, regime alignment, council consensus, governance compliance, and calibration adjustment.
The fully scored, governance-checked decision packet is persisted as a TradeMemoryRecord with complete lineage for audit and calibration purposes.
After 1/3/7/30 days, the system fetches the actual price and computes directional accuracy. Calibration engine uses accumulated correct/incorrect counts to adjust future conviction scores.
The final conviction score is the weighted synthesis of 8 independent dimensions. No single dimension can produce a high-conviction signal alone.
CONVICTION FORMULA base = anchor + sentiment + thesis + regime_alignment + council_consensus + governance_compliance final = (base − counter_thesis_penalty) ± calibration_adjustment Floor: 0.30 Ceiling: 0.95 Below floor → STAND_DOWN regardless of direction
Implementation: conviction = (Σ weighted_components) × governance_pass_flag.
Floor: 0.30. Ceiling: 0.95. Below floor → STAND_DOWN regardless of direction.
The Judgment Council (Layer 10) comprises five independently configured analyst profiles.
Each profile votes BUY / SELL / HOLD / VETO with fully documented reasoning.
Votes are persisted to council_votes for permanent audit.
Consensus rules: 3+ of 5 supporting (support_score >= 0.60) = CONSENSUS. Exactly 2 = SPLIT. Fewer than 2 or any formal veto = CONTESTED (thesis blocked).
All triggered rules are stored in governance_violated_rules (JSONB) on each
TradeMemoryRecord. Every governance event is permanently auditable.
Calibration measures the alignment between predicted conviction and actual directional outcomes. A well-calibrated system should win on roughly 70% of high-conviction signals and 50% of low-conviction signals.
Calibration results are visible in real time at /calibration/summary.