Blog · 2026-10-06 · Vishal Verma

What Minecraft taught us about hill climbing for AI in finance: A first look at our skills research

Game-playing AI climbed Minecraft's tech tree by keeping only the skills that passed a check; Matterfact hill-climbs toward a specialist's knowledge of every industry the same way, with primary records as the signal that markets do not give.

In Minecraft, the quest is a diamond. No player, human or machine, gets one by luck: it takes a chain of skills, from cutting wood to making a stone pickaxe to smelting iron, each built on the last. In 2023, an agent called Voyager showed how a machine can climb that chain. It hill-climbed: it tried a skill, checked that it worked, kept it if it did, and built the next one on top. It found 3.3 times as many unique items as earlier agents and reached the first tier of the game's tech tree up to 15.3 times faster.¹

Our quest is harder: a specialist's knowledge of every industry. It is what separates a research note a portfolio manager acts on from one that is merely fluent. A steel specialist knows that scrap is priced per gross ton and steel per short ton, which federal price series still publish, and which producers tell the market their quarter before they report it. A general model knows finance the way a well-read generalist does, and it does not know these things.

In Minecraft, the quest is a diamond, reached one checked skill at a time. On a research desk, the quest is a specialist's knowledge of every industry.

Matterfact's research agent climbs toward that knowledge the way Voyager climbed toward diamonds: through a library of more than 5,000 skills, each one a step. One part of the recipe had to change. Hill climbing needs a signal that says which way is up. A game gives one within seconds; a market may take months, and often never does. We therefore built the signal: each skill is written from research and reviewed by a separate agent against primary records before it ships.

This article is a first look at that work: what a skill knows that a general model does not, how game-playing agents hill-climbed, why markets give no signal to climb on, how we built one, and who is building it. We have written the method up as a research paper, which we are not publishing yet.

TL;DR

• A skill holds what a specialist knows and a general model does not, from a steel spread that overstates cost by about an eighth to a miner's EBITDA definition that breaks peer multiples. Skills share essentially no wording with generic guidance, and even the two most alike skills share only a sliver.
• Game-playing agents hill-climbed by keeping only the skills that passed a check. In Voyager's own tests, removing the check cut the number of items it found by 73%.
• Markets give no signal to climb on. A short thesis is judged months later, and the costliest errors, such as a stale legal status or a price series that stopped publishing, produce no error message.
• Review against primary records is the signal we climb on. The median review makes 15 changes; more than half correct a fact, and most of the rest make the skill more specific.
• The founders bring both the research and the desk: a researcher from Google DeepMind, and a quant from Millennium's systematic arm who went on to engineer Google Search.

A skill holds what a specialist knows and a general model does not

The quest is specificity. Ask whether a steel producer's margin is at a cyclical peak, and the answer gets more useful at each level of knowledge (Exhibit 1). Generic guidance, the kind our own library held before this work, is true and unusable. Sector guidance names the right measure but not how to build it. The industry skill turns it into a procedure an analyst can run.

Exhibit 1
The same question gets a usable answer only at the level of the industry.

variant: chain
numbered: true
Generic guidance :: "It's the relationship between spot prices and marginal cost of supply."
Sector guidance for materials :: Names the metal spread, and calls a spread above the 80th percentile a peak, a threshold with no source
Matterfact's skill for steel :: Builds the spread from each producer's own disclosures, converts the units, knows which federal price series still publish, and knows the quarter is pre-announced

Source: Matterfact skill library before and after the industry skills were written, October 2026

The same gap shows up in every industry. Each detail below decides whether a thesis holds, and none of them appears on a generic checklist (Exhibit 2).

Exhibit 2
In every industry, the detail that decides a thesis is one a generalist would not think to check.

Source: Matterfact short-selling industry skills, October 2026

The last row shows a rule that runs through the whole library. A skill holds how to analyse, never what the numbers are today. Margins, multiples and short interest go stale within weeks, so the agent fetches them live through its data tools each time it runs.

Each industry gets its own skill, and each runs to about 500 lines of this kind of knowledge (Exhibit 3).

Exhibit 3
A typical industry skill carries dozens of specialist details, each tied to where the evidence is filed.

columns: 2
About 15 industry signals :: The tells that lead this industry's results, each with where it is disclosed, how to pull it and when it misleads. :: lens
About a dozen traps :: Misreadings of this industry's data that flip a thesis, like the gross-ton spread. :: shield
Four to six dated cases :: Past shorts in the industry with their outcome and legal status, including shorts that failed. :: clock
Live data routes :: Where today's numbers come from, so nothing in the skill goes stale. :: refresh

Source: Matterfact analysis of the short-selling industry skills, October 2026

Specific also means distinct. A skill shares essentially none of its wording with the generic guidance for its sector, and more than nine in ten of a typical skill's phrases appear in no other skill. Even the two most alike skills, electric utilities and multi-utilities, share only a sliver of their wording, while a skill shares most of its wording with its own earlier draft (Exhibit 4). The parts that carry facts are the most distinct of all: no trap in one skill closely matches a trap in another, and no historical case appears in two.

Exhibit 4
Even the most alike skills share only a sliver of their wording.

Note: Drawn to scale. Each shared area is proportional to the share of six-word phrases the two texts have in common (their Jaccard similarity). The earlier draft is the same skill before one round of review edits, which shows what near-identical text looks like on this measure.
Source: Matterfact analysis of the short-selling industry skills, October 2026

This is the diamond at the top of the hill: a specialist's working knowledge, industry by industry. The rest of this article is about how we climb toward it.

Game-playing agents hill-climbed by keeping only the skills that passed a check

Games have been AI's proving ground for a decade because they keep score. Minecraft is a harder test than Atari or Go: it has no single score, an open world, and a tech tree in which diamond tools come only after wood, stone and iron. The agents that mastered it built on what they had already learned (Exhibit 5).

Exhibit 5
Minecraft was mastered by agents that built on what they had already learned.

Note: Abridged from each system's paper or announcement.
Source: Mnih et al., Nature (2015); DeepMind (2016); OpenAI (2022); Hafner et al., Nature (2025); Wang et al. (2023)

Voyager's lesson for anyone building agents is in its ablations. The skill library let it carry what it had learned into a new world. The check before a skill was saved mattered more than any other kind of feedback its authors tested: without it, the agent found 73% fewer items. The library supplied the capability, and the check made the library worth having.

We took that idea into finance: an agent is as good as the skills it can reach, a skill is as good as the check it passed, and hill climbing is only as good as the signal that says which way is up.

Markets give no signal to climb on

Voyager knew within seconds whether a skill worked: the code ran and the game changed. Investment research has no such signal. Whether a short thesis was right is known months later, and the market's own moves blur the answer. The most costly errors also produce no error message. Two from our own build show what that looks like:

• A data source that answers is not a data source that is current. A draft of the steel skill pointed the agent to a federal price index for hot-rolled sheet and noted that the API "returned data". It did, but the series stopped in February 2022, so an agent would have built a "current" price check on a number more than four years old.
• A settlement is not a conviction. A loose legal status, such as "settled with the SEC", can hide that a complaint alleged fraud while the settlement covered narrower provisions, without admission. In a short thesis, that difference separates a defensible research note from a defamatory one.

Neither mistake would fail a test, crash a tool or lower a score. In finance, the signal a game provides for free has to be built before there is anything to climb.

Review against primary records is the signal we climb on

We reversed Voyager's order. Its agent learned skills by playing; ours are written before use, one per industry, from academic, practitioner and case evidence. A separate reviewing agent then checks each draft against primary records, such as SEC filings, court dockets, regulator databases and companies' own reports, and corrects what it finds.

Each review is a step up the hill: it leaves the skill more correct and more specific than the draft it received. The median review makes 15 changes, and none makes fewer than seven. More than half correct a fact; most of the rest add something a specialist would check or make a generic statement specific (Exhibit 6).

Exhibit 6
More than half of the changes made in review correct a fact; nearly a third make the skill more specific.

What review changed in each skill draft Share of all changes, %

Source: Matterfact review records for the short-selling industry skills, October 2026

For anyone building agents in finance, the lesson is that fluency is not accuracy, and fluency is not specificity either. Drafts that read as if a specialist wrote them still needed a median of 15 changes, so we do not ship a skill on the strength of how it reads. How the review works, and how a skill is structured so the agent can use it, are covered in the paper rather than here.

The founders bring both the research and the desk

Building this takes two kinds of knowledge that rarely sit in one team. Knowing how game-playing agents hill-climb is machine-learning research. Knowing that scrap is quoted per gross ton comes from the desk. Matterfact's founders, who wrote the research paper together, bring one each.

• Vishal Verma, Co-founder and CTO. Comes from Google DeepMind, where he co-authored a core paper behind Gemini, and led knowledge-graph work for Apple's Siri. A computer science PhD candidate at UNC and an IIT Bombay graduate. LinkedIn
• Ashutosh Agarwal, Co-founder and CEO. Traded as a quant at WorldQuant, the systematic arm of Millennium, and covered equities as an analyst at Nomura, then engineered Google Search, the largest information-retrieval system ever built. LinkedIn

The research is led by Vishal, the paper's first author, who holds the agent to a research lab's standard: claims are measured, and failures are written up. Between them, the founders cover the model, the retrieval and the trade. That is why the skills follow the way a specialist works, and why each one is checked like a research claim before the agent uses it.

Ask any finance agent where its skills come from

Technology and research leads choosing an agent for their desk can put four questions to any vendor, including us:

1. How specific are its skills? One playbook for every company, one per sector, or one per industry?
2. What checked them before they shipped? Checked against which records, and what happened to the claims that could not be verified?
3. What does it remember, and what does it fetch? A skill that carries today's numbers is stale within weeks.
4. When a tool returns half an answer, what fills the other half? If the answer is the model's memory, the number will look precise and be wrong.

We have written up the method, the measurements and what went wrong in a research paper, which we are not publishing yet. If you lead AI or research at a fund, we would be glad to walk you through it.