How it works

From a founder's name to Meet, Look closer or Pass, in six steps. Each one does one job, and the memo shows its work.
  1. Research

    Research agent (Claude)

    Searches the web for the founder's background and keeps only what was true on the day the company was founded. Every fact links to its source, and anything you entered wins over the agent.

  2. Anonymise

    A fixed profile

    Names come out, but what they tell you stays: a university becomes its degree, subject and world-ranking band, and an employer becomes a short description with its size, industry and how long they stayed. The result is the standard profile the models were trained on, line for line.

  3. Ask

    TRL's questions, answered by Kev

    A checklist of 40 yes/no questions, such as “Has the founder had a prior exit?”, written once with TRL during the research, is answered for this founder by Kev, a small open-source decision model.

  4. Score

    A fine-tuned, calibrated Kev

    Turns the answers into the chance of each outcome, from shut down to over $1B, and how long it could take. Calibrated, so 30% means about 3 in 10.

  5. Your settings

    Your bar

    Your target, timeline, risk appetite and filters turn those chances into one number: the chance this founder reaches your target.

  6. Verdict

    Meet · Look closer · Pass

    Clears your bar: meet. Close to it: look closer. Below it: pass. The memo shows what moved the score, and lets you flip any answer to see what changes.

Why names are removed

You still see who the founder is, and every source, in the memo. Only what the model reads is anonymised.

The model only knows this format
It learned from 3,000 profiles written exactly this way, most of them generated, so it has never seen a real university or company name and couldn't use one reliably.
Names leak hindsight
Language models know what became of well-known companies and their alumni. A name would let the model recall outcomes instead of judging the background, and the founder's own company name would give the answer away.
No brand-name shortcuts
The training data was built so that no single detail decides the outcome. The ranking band and the employer's description keep the real signal without the halo of the name.

The exact university or employer can carry more than a ranking band. Letting the model read that detail would be a research change: new training data and retraining.

Built on TRL, by Vela Research

The scoring in Halomark runs on TRL (Think, Reason, Learn), an open-source library from Vela Research, released under the MIT licence. TRL is their work. Halomark uses it, adapts it and builds on it: the framework, its question-writing and its answering pipeline are Vela's; the choices and additions below are this project's.

  1. 1Think

    In TRLA language model writes yes/no questions that tell outcomes apart.

    In HalomarkClaude Opus 5.5 wrote 93 questions about founders, once. The 40 that predicted best are kept.

  2. 2Reason

    In TRLA model answers every question for each sample.

    In HalomarkKev-0.6B, an open-source decision model, answers them in place of TRL's usual answerer (TypeSafe's Jev), to run for free.

  3. 3Learn

    In TRLA classical model turns the answers into a prediction (an elastic net by default).

    In HalomarkA fine-tuned Kev makes the final call instead, added to TRL as a new option, and compared with TRL's own.

What this project added

  • In a fork of TRL: a final step that hands the decision to a decision model (Jev, Kev or any model with the same interface), and option descriptions for multiple-choice questions.
  • Kev fine-tuned on TRL's answers, with calibrated odds, compared against TRL's own final steps on the same founders.
  • Around it, the app: web research limited to before the founding date, the anonymised profile, your bar, and the memo.

In testing, no final step beat TRL's default on accuracy; the fine-tuned Kev gave the best-calibrated odds.

TRL on GitHub (Vela Research) · Kev on GitHub

Reading a verdict

The chance is the share of founders with a background like this who reached your target in the model's data. It is calibrated, so 30% means about 3 in 10.

  • MeetThe chance clears your bar. Worth a meeting.

  • Look closerClose to your bar. Your judgement decides.

  • PassBelow your bar. Few founders like this reached your target in testing.

The name

A hallmark is the small stamp an assay office strikes into gold once it has tested the metal. Halomark does that for an angel. It tests a founder against your bar and stamps the verdict: meet, look closer or pass. The halo struck into the stamp stands for you, the angel.

Share from other apps

Shared links open a confirm screen, so nothing is screened until you tap Screen founder.

iPhone: a one-time Shortcut

  1. Open the Shortcuts app and tap + to make a new shortcut. Name it “Screen founder”.
  2. Turn on Show in Share Sheet (in the shortcut's settings) and set it to receive URLs.
  3. Add the action URL Encode with the Shortcut Input.
  4. Add the action Open URLs, with this address followed by the URL Encoded Text:
/share?url=

Then, in LinkedIn or Safari, tap Share and choose Screen founder. It opens in Safari.

Android

Add Halomark to your home screen (Chrome menu, then Install app). Halomark then appears in the share menu of any app.

Privacy in short

  • Only public information is used, and only what was true when the company was founded.
  • Your notes appear in the memo and are never sent to the model. Facts in them can fill the form, and you check every value before screening.
  • Nothing is kept on our servers after your session. Your profile and settings stay on this device.
  • Claude researches and sorts notes, ElevenLabs transcribes voice notes, and Kev scores the anonymised profile.

The full privacy notice (draft)

Screening aid, not investment advice. The model estimates the chance of a large outcome from a founder's background on the day the company was founded. It was trained on generated data, so its accuracy on real founders is unknown. Public information only; nothing is stored after the session.