The product answers five questions:
- What exactly changed?
- Why could it matter economically?
- What evidence confirms or contradicts the mechanism?
- What does the current price appear to assume?
- What should be monitored next?
- Tier 1: SEC, issuer filings/releases, regulators, courts, and official company facts.
- Tier 2: earnings calls and official market or economic data.
- Tier 3: licensed consensus, identifiable institutional research, and reputable reporting.
- Tier 4: unofficial aggregators, narrative saturation, and inferred context.
Tier 4 context and LLM output never independently promote an idea. A constrained two-source Tier 3 exception may replace one unavailable Tier 1 gate only when independence, period alignment, corroborating quantitative evidence, and every other promotion gate pass.
Every thesis is represented as:
Source event -> business driver -> operating KPI -> earnings/FCF -> valuation -> catalyst
Each connection has its own evidence, confidence, falsification test, and next action. Weak links remain explicit.
- Research Question: an anomaly or unresolved claim worth investigating.
- Candidate: a source-linked hypothesis with direction still provisional or material gaps outstanding.
- Research-Ready: a mapped driver, current price, counter-thesis, monitors, and auditable payoff assumptions.
- High-Conviction: primary or qualifying secondary-supported evidence, a complete causal bridge, no critical contradiction, and every strict gate passed.
Idea Quality Score measures research completeness, not success probability.
Ambiguous metrics do not receive automatic polarity. Goodwill is treated as acquisition accounting; capex as an investment-cycle signal; share-count changes require share-basis reconciliation. Constructive and adverse mechanisms are presented side by side until cited evidence supports direction.
Thesis validity and expectation capture are separate:
- Price reaction can support a Price-only capture assessment.
- Point-in-time consensus snapshots are required for claims such as
uncaptured,not priced in, orconsensus did not react. - Current consensus is never substituted for missing historical consensus.
When historical revisions are unavailable, reverse DCF and implied operating assumptions ask what the market must already believe without pretending to know analyst forecasts.
Bull/base/bear scenarios use model-derived or explicit user assumptions. Default probabilities are labelled illustrative and uncalibrated. Probability-based ranking remains disabled until at least 30 eligible resolved outcomes exist in the same signal family and horizon.
Historical research may use only filings, prices, consensus snapshots, macro releases, and external observations available on or before the relevant event cutoff. Retrieval time and source publication time are stored separately.
The LLM receives curated excerpts and structured evidence IDs only. Output that invents citations, facts, targets, probabilities, or unsupported recommendations is rejected. Provider timeout is reported as execution failure, not guardrail rejection, and deterministic research remains available.