Research

Three bodies of work. The Atlas and the retired programme's record open to their full text on this site; single-name research is delivered to clients on request, with one note published as a sample. Dated, sourced, and carrying their own correction history. Not investment advice.

Industrial researchEdition 28 August 2026 · scoring data 8 August 2026 · status 21 September 2026Full text

The AI Supply Chain Atlas: 131 scored chokepoints, and who can collect the rent

A 12-tier map from ores and gases to models and governance, scored on constraint, substitutability, relief time, geography, demand exposure and monetisability. Twelve falsifiable theses with pre-committed kill criteria, a corrections register, a provenance pass and a primary-source triage that records which supporting legs did not survive contact with filings.

Start with the 90-second case on the substrate node: the judgment, the evidence by source class, what changes it, and how a correction moved the evidence class without moving the score. Thesis 01 is marked contested as of 21 September 2026.

Equity researchDelivered to clients · one sample publishedSample

Single-name equity research: what the price requires, and the observable that decides it

A full-length note on one company. It states the question and the date its data is marked to; what the market expects and how that was established; the mechanism and the two or three pieces of evidence that carry it, by source tier; a scenario range rather than a point target; the observables and thresholds that would retire the view; and a version history in which each correction is shown against the earlier version.

One note is published in full as a sample: Circle Internet Group (NYSE: CRCL), version 2 of 19 September 2026, marked to the 18 September 2026 close, with its version 1 correction shown. Client notes are not published. To request coverage of a name or a sector, send the name, the question, the horizon and the delivery format.

Quant research (retired)V12 · last run 2026-04-29 · retiredData + validation

A cross-sectional one-month model: the retired programme's final record

Thirty U.S. mega-caps, eleven factors tested and four kept, a three-model ensemble, and a walk-forward out-of-sample panel of 19 months, validated with a fixed battery (Deflated Sharpe, stationary bootstrap, purged hold-out). The programme was retired after its last run in April 2026; the snapshot, the validation record and the unedited output file stay published as a record of the method.

What every research note answers on its first screen

What question, and as of which date?

Every note states the question it answers and the date its data is marked to.

What is our judgment, and what does the market expect?

The market expectation is sourced (a consensus snapshot, a named house note, a filed guidance figure) or the inference method is stated. Where it cannot be established, the note says "our scenario assumption", not "the market prices".

How was it produced?

AI-assisted tooling is used for sourcing, extraction from filings and transcripts, and code. The question, the model structure, the evidence standard and the triggers are set and reviewed by a human researcher, who checks the figures that carry the conclusion; the tooling does not set them.

What are the two or three pieces of evidence that carry it?

Listed by source tier: audited (S1), filings (S2), management guidance (S3), consensus and market data (S4), industry and press (S5). A fitted parameter is never an anchor.

What would prove it wrong, and what has been corrected?

Each note carries dated triggers and a version history. A correction is shown against the earlier version, with what it changed.

How does it become a monitored playbook?

The observables and thresholds that would update the view are listed with their cadence and source, so the note can be re-run rather than re-read. Where that monitor is not yet built, the page says so.

Archive: pipeline methodology notes, V10–V11 (note dates Oct 2025 – Jan 2026; pipeline releases 2025-11-01 and 2025-12-15)

Summaries only. No full text was published for these notes, and the pipeline they describe has been superseded by V12. They are listed so the version history is visible, not as current work.

January 2026 · V11

V11 Adaptive Ensemble: Regime-Aware Factor Investing

9-factor ensemble with VIX regime weighting. Superseded by V12 (11 factors tested, 4 kept; 30-name universe).

December 2025 · V11

Evidence-Based Bid-Ask Spread Modeling for Options

Per-ticker spread model trained on observed chains. Still used by the options overlay; the overlay itself has never been backtested.

November 2025 · V10–V11

Conviction-Tiered Options Overlay with Vol-Edge Selection

Kelly-sized tiers. The tier rule in production is quintile-based; the thresholds in this note no longer apply.

October 2025 · V10

80K-Path Monte Carlo with Implied Volatility Surfaces

Model-implied versus ATM implied vol. Forward-only; not a historical test.

All research is for informational and educational purposes only. Model outputs and backtest figures are hypothetical. Research views are unsized and carry no price target. Seeking Edge personnel may hold positions in securities discussed; positions are disclosed separately from the research view. Not investment advice. Full disclaimer.