THE SIGNAL
Nova One Capital

An AI decision system, built as infrastructure.

Axiom is the research and decision system behind the Nova X Convergence Fund: more than fifty agents with defined mandates, a governed data layer, and a person who places every order. Its framework has run since 2023, when the first large language models became usable for research, and has been rebuilt with each generation of frontier models since.

01

Mandate

AI and compute
The chips, memory, networks and data centers the buildout runs on, and the companies whose margins it changes.
Power and energy
Generation, grid and fuel: the physical constraint on how fast intelligence can scale.
Capital markets
The rates, financing and market structure that price all of it.
02

The desk

  • Lynceus

    Scout

    Scans the market every trading day for an edge and hands the strongest candidates to underwriting. It shapes ideas; it never sets an entry.

  • Theia

    Thesis Analyst

    Deep underwrites on frontier models: the bull case, the bear case and a twelve-month value, written before a position is considered.

  • Momus

    Red Team

    A standing adversary. It attacks every position and proposal and writes the bear case the desk has to answer.

  • Sibyl

    Catalyst Verifier

    Confirms every date the desk relies on, from earnings to data releases, against primary sources.

  • Themis

    Risk Officer

    Holds the limits, stops and entry gates. It runs in code, with no model.

  • Metis

    Decision Auditor

    Re-prices every call at 30 and 90 days, against the market over the same window. It runs in code.

  • Aletheia

    Truth Auditor

    Checks published figures against their sources before they ship. It runs in code.

  • Minerva

    Chief Learning Officer

    Grades the desk every week and turns each correction into a lesson it is held to.

Three of the roles that matter most, risk, decision audit and truth audit, use no model at all. Judgment goes to models; limits and arithmetic stay in code.

03

How a decision is made

  1. 1

    Scan

    The scout surfaces candidates from the whole market each day.

  2. 2

    Underwrite

    The thesis analyst writes the bull case, the bear case and a value.

  3. 3

    Attack

    The red team tries to break the thesis before the desk adopts it.

  4. 4

    Gate

    Risk sets the entry condition, the size limit and the stop, in code.

  5. 5

    Approve

    A person decides and places the order. Nothing trades on its own.

  6. 6

    Record

    The call is journaled with the alternative it rejected and what would prove it wrong, then re-priced at 30 and 90 days.

04

Doctrine

Judgment
Frontier models where a wrong answer is expensive: underwriting, red-teaming, the scout. Efficient models for retrieval and narration.
Numbers
Every figure is computed in code. The models explain numbers; they do not produce them.
Data
Read-only brokerage and market data, reconciled against the broker as the source of truth.
Cost
Every model call is metered by agent and purpose, so the system knows what each judgment costs.
Grading
Outcome and process are graded separately. A sound call with a bad result is still sound; a lucky one is still luck.
Controls
Rules are enforced in the data layer, not by convention. A call that skips a required review cannot be recorded.
05

On the record

Entry Decided Oct 15, 2026Kind Stand-downConviction High
Call
No entry and no new gate on a group of hard-hit AI hardware names () at the close.
Rationale
The question was whether a drawdown this deep made an attractive entry. The underwrite found no name that cleared the fund's reward-to-risk bar at a valid entry; the strongest came in at .
Rejected
Buying the hardest-hit name at the current price as a value entry: inside an earnings window, trend down, and the thesis itself under attack.
Falsifier
If by November 30, 2026 three or more of the names close above their October 15 closes without a valid entry ever having been available, the stand-down cost the fund the entry.
Review
30 days: pending. 90 days: pending. Outcome and process graded separately.

A real entry from the decision journal, with names and thresholds redacted.

The framework dates to 2023. The journal is its current form: every call recorded before its outcome is known, with the alternative it rejected and a written condition that would prove it wrong.

06

Operating standards

  • Human approval

    The system proposes. A person places every order; nothing trades on its own.

  • Audit trail

    Agent runs, inputs and decisions are logged to a central ledger and can be reviewed after the fact.

  • Least privilege

    Read-only data access, credentials held in managed secret stores, row-level security on the data layer.

  • Monitoring

    Health checks run on every agent throughout the day, and failures surface the same day.

07

Access

Axiom is proprietary to Nova One Capital and is not offered, licensed or sold.

Its public output is The Signal, the letters of the Nova X Convergence Fund, published every Saturday with their sources cited.