Component Architecture

The Genome platform is a full-stack system for identifying and analyzing industrial companies. It operates from raw signal ingestion through to commercial outreach.

Signal Ingestion

Genome draws signals from public sources using a multi-tiered approach:

  1. EDGAR filings (SEC 10-K/10-Q) - financial performance, debt, capital structure
  2. FRED economic series - macroeconomic trends (inflation, interest rates, industrial production)
  3. Commodity price indices - energy, metals, industrial goods
  4. Company press releases and financial reports
  5. Public social media and news sentiment

Each signal is mapped to company constraint nodes (F1–F4 for financial, S1–S4 for supply chain, R1–R4 for risk, X1–X2 for regulatory and physical).

Fingerprinting Engine

The core fingerprinting engine performs six system identification algorithms:

  1. Latency - the time between signal input and constraint output
  2. Gain - sensitivity to input change
  3. Damping - resistance to oscillation
  4. Oscillation - periodic behavior in signals
  5. Saturation - limit to growth
  6. Volatility Transmission - how volatility propagates through the system

Constraint Classification

Companies are classified into seven archetypes:

  1. Oscillatory — cyclically unstable but with clear directionality
  2. Constrained — bottlenecked by inputs or outputs
  3. Fragile — vulnerable to external stressors
  4. Resilient — robust against systemic disruption
  5. Saturated — operating near capacity in all constraint dimensions
  6. Decoupled — insulated from macroeconomic trends
  7. Transitory — not permanently constrained, but in transition

Each archetype is mapped to multiple constraint nodes for a full fingerprint.

Commercial Outreach Pipeline

Genome's commercial engine automates outreach through the following stages:

  1. Targeting - identifies companies with specific constraint types and archetypes
  2. Genome Report Generation - creates detailed analysis of key constraints, trajectories, and peer comparisons
  3. Email Drafting - generates templated outreach messages based on recent company activity
  4. Response Classification - categorizes responses for future targeting
  5. Strategy Learning - refines targeting based on past success/failure signals

Key Components

New Trust Layer Implementation

A new trust layer was implemented that improves the reliability of signals by:

  1. Facts vs Inference Envelopes - separating raw, verifiable EDGAR/price data from probabilistic model inference (archetype, trajectory, conviction)
  2. Freshness Contract - fingerprints older than 45 days suppress inference conviction; EDGAR/PFSL facts older than 100 days are flagged stale (not suppressed — a stale 10-K is still a fact)
  3. Live-Calibrated Conviction - conviction derived from live-validated win rate instead of hardcoded backtest constants
  4. Inverse-Validation Calibration Loop - a standing loop that reads trading's inverse shadow-book outcomes to catch inverted or no-edge signals

Additionally, two new endpoints were created:

  1. /v1/facts/{ticker} - returns the verified factual layer (EDGAR/price-derived) + freshness for a company. Structurally incapable of returning archetype, trajectory, or conviction.
  2. /v1/facts/batch - batch facts-only lookup for multiple tickers

New Reasoning Frame Endpoint

A new endpoint gives Genome/MI signal claims the same discipline CDI enforces on its own claims internally — a binding node, a mechanism hypothesis, a falsification condition, and a confidence grade:

Consumers are Advisor, Studio, and the newsletter. CDI does not call this endpoint. Per the locked 2026-07-09 decision, CDI takes no runtime dependency on Genome — it receives the reasoning-frame discipline as a doctrine/prompt update ported into its own logic, driven by CDI's own data, never a Genome API call.

CDI does not call this endpoint, and there is no integration between CDI and the endpoint (which was added to support the newsletter/Advisor). The endpoint is an externally available RESTful API endpoint that can be used by any client for extracting reasoned signal claims from any ticker. The documentation here reflects how to use it — not a runtime dependency within the system.

New Signal Pipelines

Two new signal pipelines were added to support EDGAR and FRED-based signal ingestion:

These are used by the F1/F2 constraint nodes in the fingerprinting engine. The data is ingested into a new table census_trade_hs and usitc_hts respectively, and used as inputs to the core signal algorithms.

Additionally, improvements in pipeline robustness:

Pipeline Route Migration

The following pipeline routes were migrated from Supabase to local Postgres:

This improves data consistency and reduces dependency on external services.

Trajectory Classification Improvements

The trajectory classification logic has been improved:

Stack & Deploy

LayerTechNotes
FrameworkFastAPI (Python 3.10)API layer
DatabasePostgreSQLvia Prisma ORM
AuthenticationNextAuth.jsSession-based with role-based access
AIClaude modelsfor analysis and report generation
FrontendNext.js App Router + TypeScriptUI components and routing
DeploymentVercel (Stable)CI/CD pipeline
External Data SourcesEDGAR, FFred, company filings, commodity indicesvia custom ingestion scripts
New Trust LayerLangfuse tracing (optional)Not blocking if not configured; logs to /tmp/langfuse.log with 2-day age cutoff, not stored long-term. Env: LANGFUSE_URL, LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY

Public Endpoints

This endpoint returns:

{
  "ticker": "AAPL",
  "facts": {
    "pfsl": { "capex_activity": "...", "demand_signals": "...", "operational_stress": "..." },
    "growth_quality": { "...": "..." },
    "dl_var": { "...": "..." },
    "meta": { "provenance": "edgar_10k+price", "confidence": 0.93, "as_of": "2026-03-31", "stale": false }
  },
  "freshness": { "pfsl_age_days": 41, "pfsl_stale": false }
}