AI Hedge Fund — Deep Value

AI Hedge Fund — Deep Value

_{<a href="README.ko.md">한국어</a>}

Three famous value investors read the same filings, on their own, and one arbiter weighs what they concluded into a single judgement.

In one paragraph

This package is a port — the methodology is not ours. Three analyst personas were lifted out of virattt/ai-hedge-fund (MIT, © 2024 Virat Singh) and reassembled to run on Aumos. Each one reads the same set of filed company figures without seeing the other two's answers, forms a view, and states how strongly it holds it. A final stage averages those views — Graham's counts double, which is the source's own setting and not a preference of this port — checks the result against your own rules, and hands back one judgement. Its usual answer is WATCH, on a price that would create a margin of safety.

	basic-investor	prudent-allocator	ai-hedge-fund-value
origin	ours	ours	ported, MIT
starts from	the asset	the book	the filings
reads	filings, news, prices, the book	the book, prices, the theses	filings and prices only
readers	one	one	three, independent
its usual answer	WAIT, or BUY with a thesis	WAIT, or RESIZE	WATCH, on a price that would create a margin of safety

The methodology

Deep value, by committee — but a committee that does not confer.

Words this page uses, in plain terms:

Margin of safety	buying far enough below what a business is worth that being somewhat wrong still leaves you whole
Moat	a durable reason competitors cannot take the profits away
Conviction	how strongly an analyst holds a view, as a number between -1 and +1
Abstain	"I have no opinion" — which is not the same as "my opinion is neutral", and is scored differently
Point-in-time	only the figures that had actually been published at the moment being judged

• Graham — margin of safety, balance-sheet strength, earnings stability, and deep suspicion of paying for growth. Overvaluation is a bearish fact, not a neutral one. His view counts double, which is the source's own configuration and not a preference of this port.
• Buffett — circle of competence, moat, capital allocation visible in the numbers, and ten-year comfort. A wonderful company at a fair price over a fair company at a wonderful price.
• Munger — invert; find what would make it fail; and put most things on the too-hard pile.
• The blend — a weighted mean over the analysts who voted, then sized against your mandate. Abstentions are excluded from both halves of the average, so "no opinion" never arrives as "opinion: neutral".
• The mandate check — and an explanation of why the manager must respect a limit the source's harness would have clamped for it.

The independence is the whole methodology. Three analysts who converge because the third read the first two are one analyst with extra steps, and the blend would then be weighing a view against its own echo. The stage files say so out loud, twice.

It is given no news access and no access to your theses, and that is the port being faithful rather than the port being lazy. The source hands its value analysts a rendered fundamentals snapshot and nothing else; a port that quietly added a news feed would be an improvement filed under somebody else's name. Where that costs the judgement something, the manager is required to say so in its stated uncertainty.

What is different from the source, and why each change was necessary

Not a list of liberties taken. Each of these is something that could not be carried across, and the Aumos rule that made it impossible.

the source does	this package does	because
sizes cross-sectionally: conviction / Σ\	convictions\	× gross_target	conviction × your maximum position weight	Aumos asks about one subject at a time. Divide one conviction by its own absolute value and the answer is 1.0 for any conviction, so a barely-positive view and an overwhelming one size identically. The source names this exact wart itself
clamps a target above the cap, silently	proposes inside the mandate, or explains that the mandate stopped it	Aumos does not clamp — it records the proposal verbatim and rules it a WAIT. A target over the cap costs the whole judgement, not some exposure
an all-neutral cycle produces zero weights and the fund closes to flat	returns WAIT or WATCH with reasons	closing to flat is a trade, and one nobody decided to make. WAIT is a first-class judgement here
shorts its least-liked names in market-neutral mode	drops the short side and says so	shorting is a mandate constraint and is normally off. Approximating a short with a hedge would be inventing an instruction
states the point-in-time discipline as a prompt rule the model is asked to keep	states it as a description of machinery	Aumos enforces asOf and refuses a result stamped after it. A prompt that asks a manager not to look ahead produces a manager that tries
has no minimum-conviction floor (its own docstring calls this the obvious next knob)	convictionFloor, default 0.2	upstream, the risk stage caught the consequence by clamping. With no clamp, the floor stops being an improvement and becomes a requirement
places its own orders	cannot	there is no broker-write capability in the protocol to ask for. Orders reach a venue through your approval and nowhere else

What is not ported is as deliberate: eleven nodes of the source harness are recorded in harness.json as omitted, each with the reason and with what in Aumos answers the same question instead — or nothing, where the port simply loses the capability. A node declined and a node missed are opposite facts, and a reader is entitled to tell them apart.

How a run works

One run, one proposal. Nothing below places an order.

flowchart TB
    classDef reads fill:#1e2a44,stroke:#6f9bf0,color:#cfe0ff
    classDef judges fill:#2f2f38,stroke:#9aa0b4,color:#e8eaf2
    classDef proposes fill:#1b4332,stroke:#40916c,color:#d8f3dc
    classDef person fill:#5c4813,stroke:#f6a609,color:#ffe8b0

    WAKE["Aumos wakes it on an asset<br/>or event review"]:::reads

    subgraph IN["What it reads"]
        direction TB
        FIL["Filings — the figures published<br/>as at the moment being judged"]:::reads
        PX["Price history"]:::reads
        BOOK["Your portfolio as it stands"]:::reads
    end

    subgraph PANEL["Three readers, none of whom sees the others"]
        direction TB
        G["Graham<br/>margin of safety, balance sheet"]:::judges
        B["Buffett<br/>moat, capital allocation, ten years"]:::judges
        M["Munger<br/>invert; mostly the too-hard pile"]:::judges
    end

    BLEND["The blend<br/>weighted mean over the analysts who voted<br/>Graham counts double · abstentions removed<br/>from both halves"]:::judges

    FLOOR{"Is the blended conviction<br/>above the floor, and are at least<br/>two analysts on the same side?"}:::judges

    SIZE["Conviction × your maximum position weight"]:::judges
    MAN{"Does that target sit<br/>inside your Mandate?"}:::judges

    WAITW["WAIT or WATCH<br/>with the price that would change the answer"]:::proposes
    BUY["BUY / RESIZE<br/>at a target inside the mandate"]:::proposes

    subgraph HUMAN["Where a person decides"]
        direction TB
        MAND["Aumos judges it against your Mandate"]:::person
        YOU["You approve, or you do not"]:::person
        ORD["Only then does an order exist"]:::person
    end

    WAKE --> IN
    IN --> G
    IN --> B
    IN --> M
    G --> BLEND
    B --> BLEND
    M --> BLEND
    BLEND --> FLOOR
    FLOOR -- no --> WAITW
    FLOOR -- yes --> SIZE --> MAN
    MAN -- no --> WAITW
    MAN -- yes --> BUY
    WAITW --> MAND
    BUY --> MAND
    MAND --> YOU --> ORD

Legend — 🟦 what it reads · ⬜ what it works out on its own · 🟩 what it hands back · 🟧 where a person decides.

Cadence. It is woken by an asset or event review rather than by a calendar, and the WATCH it most often returns names the price that would change its answer.

What it needs

Markets	US-listed equities and ETFs (XNAS, XNYS)
Data	source:passthrough — filings and price history, asked of the vendors this machine holds credentials for. Each persona reads the vendor's own response; nothing is mapped on the way
The book	portfolio:read — so a conviction about a business becomes a judgement about this portfolio rather than a rating in the abstract
Settings	the three analyst blend weights (the source's own defaults), the conviction floor, and whether a majority is required. Position caps are not here — they are your Mandate, and letting a package set them would let it argue with the constraint it is judged against
What it will not read	no news, and no access to your theses. That is the port being faithful to a source that reads neither
Your approval	it proposes and never trades. Every judgement enters your approval queue and a person approves it

Installing it is exactly like any other package — from the catalogue, against a book, with the same capability consent screen. The install screen shows this package's provenance above its capabilities, because who wrote the methodology is something you are entitled to know before you agree to run it.

What it is bad at

• Anything the filings do not say. No news, no management commentary, no market narrative. Where the answer turns on one of those, it is required to say so rather than guess.
• Judging fast-moving stories. Deep value is a slow reading of published figures, and its characteristic output is a price to watch rather than a trade to make.
• Growth. Graham's double weight means the panel is structurally suspicious of paying for it, and will say no to companies that go on to be right.
• Anything it cannot short. The source shorts its least-liked names; this port drops that side and says so, so half the original strategy is not represented.

Do not install this if you want new ideas quickly, or a manager that will argue a company's story rather than its balance sheet.

Notes

The first shadow run. Fixture data, asOf five minutes before an NVDA earnings release. Outcome: a sealed WATCH, four tool calls all allowed, hash chain valid. What makes it evidence that the port works rather than that a model works:

Graham bearish on price alone, Buffett abstained for want of any balance-sheet or returns data, Munger neutral on the too-hard pile — blended conviction -0.43 over the two who voted. That clears the 0.20 floor, but only one voting analyst sits on either side of zero, so the majority requirement fires and the desk does not move the existing 5% position. […] The single most important fact here is timing: fiscal Q1 is reported five minutes after asOf and this judgement is made without it. · Arithmetic: (2.0 × -0.65 + 1.0 × 0.00) / (2.0 + 1.0) = -1.30 / 3.0 = -0.43. Buffett is excluded from both numerator and denominator.

Four things in that are the port rather than the model:

• The abstain/neutral distinction held. Buffett did not vote and was removed from both halves of the average; Munger's too-hard-pile neutral did vote and diluted. That is the source's own rule — "no opinion must not masquerade as opinion: neutral" — surviving a translation from Python into prose.
• Graham's double weight is visible in the arithmetic, and the arithmetic is written out, so a reader can recompute the number the judgement rests on.
• The replacement sizing rule was reached and reported: -0.065, and then not proposed, because shorting is off. The source would have shorted here.
• The point-in-time bound is in the manager's own words, unprompted, as the thing it most wants recorded.

Everything the run saw is filed as evidence stamped at asOf, so the claim is checkable rather than taken on the transcript's word. That is one run against fixture data, and it says nothing about whether this manager invests well.

Attribution. NOTICE.md carries the retained MIT notice and the personas-are-not-the-people statement; harness.json records what was taken, what was authored here, and what was deliberately left behind, and npm run check:provenance holds the two to each other.

Aumos shows no returns for this package until it has earned some. A forward track record takes calendar time rather than compute, and nothing on this page is one.