Primary Sources Only

Every research report begins with primary source documents filed with the SEC. The engine reads 10-K annual reports, 10-Q quarterly filings, 8-K event disclosures, and earnings call transcripts directly from EDGAR. No analyst summaries, no aggregator rewrites, no secondary sources.

Evidence-Based Framework

The system uses a Bayesian evidence accounting framework. Each piece of evidence is recorded with:

  • Likelihood Ratio (LR) — How much more likely you would observe this evidence if the thesis is correct vs. incorrect. LR > 1 supports the thesis; LR < 1 undermines it.
  • Credibility — Confidence that the underlying claim is true, scaled 0–1. An SEC filing scores ~0.95. A management forward-looking statement scores lower.
  • Source attribution — Direct link to the filing, transcript section, or data source.

Evidence accumulates over time. Prior odds are updated multiplicatively: posterior odds = prior odds × LR. This prevents recency bias and forces explicit quantification of conviction.

Factor Decomposition

Returns are decomposed into systematic (market, sector, style factors) and idiosyncratic components via regression against relevant benchmarks. The engine focuses exclusively on idiosyncratic factors — company-specific drivers that are not explained by broad market movements.

Every idiosyncratic factor is classified into one of five categories:

  • Catalyst — Binary events that reprice equity (FDA approval, contract win, M&A)
  • Execution — Management's ability to deliver on the plan
  • Demand — Company-specific demand advantages (not sector beta)
  • Position — Competitive positioning, moat, market share
  • Survival — Balance sheet strength, cash runway, covenant compliance

Multi-Agent Research Pipeline

Research is produced by specialized autonomous agents operating in sequence:

  1. Filing monitoring — Continuous scanning of SEC EDGAR for new filings across covered tickers.
  2. Evidence extraction — Each filing is read in full. Material facts are extracted and recorded with likelihood ratios and credibility scores.
  3. Cross-referencing — Findings are corroborated against earnings transcripts, adjacent tickers, and sector data.
  4. Adversarial review — A separate review agent challenges the analysis, identifies gaps, and stress-tests the reasoning before publication.
  5. Publication — Only research that passes review is published to the feed.
  6. Human oversight — Experienced investors (with backgrounds in public equity and M&A analysis) calibrate methodology, review output quality, validate evidence frameworks, and perform final sign-off on an ongoing basis.

What This Research Is Not

AI-assisted research with human oversight and calibration. Not investment advice. The engine does not make buy/sell recommendations, set price targets, or suggest portfolio allocations. Every report is structural analysis — what the filing says, what evidence it provides, and how it updates the probability landscape. All analysis should be verified directly in SEC filings.

Data Sources

  • SEC EDGAR — 10-K, 10-Q, 8-K filings (full text)
  • Earnings call transcripts
  • Real-time market data — Price, volume, fundamentals, options, technicals
  • Commodity futures curves

Calibration

The system tracks explicit predictions with deadlines and binary outcomes. Calibration is measured via Brier scores — the standard metric for probabilistic forecast accuracy. All predictions, right or wrong, are preserved for accountability.