Skip to content

Latest commit

 

History

History
75 lines (57 loc) · 3.09 KB

File metadata and controls

75 lines (57 loc) · 3.09 KB

Architecture

IC Copilot keeps evidence collection and decision synthesis separate so a fluent narrative cannot overwrite a weak factual record.

flowchart TB
    UI[Streamlit and lightweight local UI]
    ORCH[Research pipeline]
    REG[Registered source adapters]
    NORM[Entity, period, unit, currency, and metric normalization]
    CLAIM[Claim validation and contradiction ledger]
    MODEL[Company model, scenarios, and market-implied expectations]
    PEER[Peer metrics and price attribution]
    GATE[Research-Ready and High-Conviction gates]
    LLM[Optional LLM synthesis and critique]
    STORE[(Local SQLite and bounded caches)]

    UI --> ORCH
    ORCH --> REG
    REG --> NORM
    NORM --> CLAIM
    CLAIM --> MODEL
    CLAIM --> PEER
    MODEL --> GATE
    PEER --> GATE
    GATE --> LLM
    LLM --> UI
    ORCH <--> STORE
Loading

Core boundaries

Deterministic evidence

  • SEC submissions, filings, exhibits, and XBRL facts.
  • Issuer releases, presentations, and registered transcript sources.
  • Point-in-time prices, consensus snapshots, macro releases, and peer observations.
  • Canonical entity, metric, period, unit, currency, and ADR normalization.

Research interpretation

  • Neutral-first anomaly assessment.
  • Constructive and adverse causal hypotheses.
  • Evidence closure against registered source requirements.
  • Company-model, scenario, valuation, credit, and market-implied bridges.

Optional LLM contribution

  • Registered-source planning.
  • Document and table triage.
  • Citation-bound extraction drafts.
  • Trend summary, thesis synthesis, and secondary critique.

LLM output is provisional until deterministic checks validate identity, period, scope, citation, unit, currency, and source eligibility.

Extension points

Extension Primary location Required contract
Sector playbook data/sector_kpi_playbooks.csv Material KPIs, drivers, confirmation and falsification rules
Industry playbook data/industry_playbooks.csv Business economics, catalysts, valuation methods, peer logic
ADR/FPI profile data/adr_profiles.csv Ratio, currency, forms, home exchange, issuer sources
Metric alias data/metric_ontology.csv Canonical metric, source tag, derivation and confidence
Peer universe data/peer_universes.csv Curated peers, rationale, effective date, sector template
Source adapter equity_research/ Provider health, timestamps, source tier, citation and licensing policy
Demo case equity_research/storytelling.py and sample_data.py Sanitized no-network fixture and regression tests

Storage

SQLite stores normalized observations, manifests, idea versions, evidence, monitor rules, outcomes, and provider health. Full licensed vendor payloads are not retained by default. API keys are never written to SQLite.

Failure policy

Provider failures are independent and non-blocking. The app distinguishes missing keys, entitlement errors, timeouts, blocked networks, malformed payloads, stale periods, missing metrics, and unsupported securities. A failed secondary provider cannot erase valid SEC or issuer evidence.