How the desk works
Machine breadth. Human conviction.
A single human cannot read every Canadian and U.S. operator's filings, transcripts, insider activity, and pricing data each quarter. Halvren does not pretend otherwise. We use AI to read the universe at scale, surface the names that earn a closer look, and prepare the work the principal then reads, decides, and signs. The machine does breadth. The human does judgment. The capital follows the human.
Public market data, ingested continuously.
SEDAR+ and SEC EDGAR filings. Quarterly transcripts. Insider transactions through SEDI and Form 4. Reserve and resource reports. Pricing and consensus from listed feeds. Public data only. No alternative-data scraping that crosses a privacy line, no inside information, no tipped channels.
SEDAR+EDGARSEDI / Form 4NI 51-101Pricing & consensusReading at scale, with discipline.
Large language and statistical models read filings and transcripts the way a junior analyst would, faster and at greater volume, and produce structured summaries against a fixed schema: business, balance sheet, capital allocation, insider behaviour, cycle position. The same model output is generated for every operator in the coverage universe in the same format. Consistency is the point.
Models also run quantitative screens against the Halvren Checklist: full-cycle FCF, unit economics at trough, balance sheet at trough, ROIC on incremental capital, insider activity bought versus granted. The output of this layer is a triaged list of names, not a position and not a recommendation.
LLM-assisted readingStructured extractionQuant screensThe principal reads everything that earns a position.
Model output is a starting point, not a verdict. The principal reads the underlying filings for any name that reaches the publishable stage, validates each Halvren Checklist answer against primary sources, audits the model's summary for hallucinations or omissions, and forms an independent view of management quality, succession, and cycle. Every public writeup, every quarterly letter, and every position is human-reviewed and signed.
The model never writes the conviction sentence. The model never sets the price. The model never decides what gets owned.
Human in the loopProprietary, sized by hand.
Halvren manages proprietary capital only. Position sizes are set by the principal against conviction and against the rest of the book. There is no automated execution, no algorithmic strategy, and no model in charge of when to buy or sell anything. Trades are entered and exited by hand. The model knows the universe; the human signs the order.
Proprietary capitalNo algo executionWhat the machine does
- Reads filings, transcripts, and proxy circulars across the coverage universe
- Extracts structured data into a fixed schema, refreshed each quarter
- Screens against the Halvren Checklist and surfaces deviations from peer baselines
- Monitors insider transactions, debt covenants, and material disclosures across the watchlist
- Drafts initial summaries that the principal then audits and rewrites
What the human does
- Reads the underlying filings for every name that reaches publication
- Audits every model output against primary sources before it leaves the desk
- Decides what is on the desk, what gets a writeup, and what goes in the book
- Sizes every position against conviction and against the standing book
- Signs every public writeup, every quarterly letter, and every transaction
What Halvren is not.
Not a quant fund. No statistical-arbitrage strategy, no factor sleeve, no automated trading. The investment process is fundamental. The AI is a research instrument, not an execution engine.
Not a model-driven black box. Every conclusion in a public writeup or a private letter can be traced back to a primary filing the principal has read. If the model produced something the principal cannot defend in plain English, it does not appear on the page.
Not an outside-capital business. Halvren manages proprietary capital. The public research is free. The private letters are research and commentary, not personalized advice.
For the editorial standard behind this work, see the Halvren Checklist. For the working universe of names, see Coverage. For the format of the public writeups, see the research archive.