Same inputs
Each model receives the same market brief, constraints, and asset universe.
CAPITALBENCH MANIFESTO
The Scoreboard for AI Capital Allocation
AI is entering investing before the world has a trusted way to measure it. CapitalBench benchmarks leading models on real public-market portfolio decisions using frozen, auditable records and live market outcomes.
Opening thesis
AI is beginning to enter one of the most important decision loops in the world: where capital goes.
LLMs are already reading filings, summarizing earnings, ranking opportunities, generating investment theses, and helping people decide what to buy, avoid, or hold.
Today they are research assistants. Soon they will become portfolio copilots. Eventually, many will become agentic allocators inside funds, apps, APIs, and advisory products.
But capital allocation is not a demo.
A model that writes well is not necessarily a model that invests well. A model that explains risk is not necessarily managing risk. A model that sounds confident may simply be wrong with better grammar.
The gap
The market has a measurement problem
Most AI investing claims are still built on screenshots, cherry-picked calls, private prompts, and backtests that may never survive contact with live markets.
That is not evidence.
Capital allocation needs a public record. It needs repeatable tests, frozen decisions, real prices, and scoring that cannot be rewritten after the outcome is known.
The financial world is moving toward AI-assisted allocation before it has built the measurement layer required to trust it.
One chart business
AI adoption is rising. Proof is still missing.
Asset Managers Have AI Adoption. They Don't Yet Have Proof.
AI is already entering investment workflows, but proof of return and risk impact remains scarce. That gap is the opening for CapitalBench.
The proof layer
CapitalBench exists to build the proof layer
CapitalBench benchmarks leading AI models on real capital allocation decisions.
Every model receives the same market brief, the same asset universe, and the same decision window. Every portfolio is frozen before results are known. Outcomes are scored against real market prices.
- No private retries.
- No retroactive edits.
- No moving the goalposts.
- No vibes.
The goal is not to prove that AI can beat markets every week.
The goal is to make AI investment behavior observable.
Infrastructure
How the benchmark works
Frozen decisions
Portfolio files are locked before outcomes are known, creating an auditable record.
Real market scoring
Results are scored using actual market prices, not synthetic grades or subjective reviews.
Public proof
Results, methodology, and proof files are designed to be inspectable and hard to revise after the fact.
Useful questions
The questions that matter
- Which models take too much risk?
- Which models crowd into the same trades?
- Which models diversify intelligently?
- Which models only sound convincing?
- Which models improve as markets change?
- Which models deserve to be trusted near capital?
Why now
Why this matters now
The first wave of AI benchmarks measured knowledge, reasoning, coding, math, and instruction following.
The next wave will need to measure judgment under uncertainty.
Capital allocation is one of the hardest forms of judgment. It requires tradeoffs, risk control, timing, humility, and the ability to make decisions with incomplete information.
As AI moves from research assistant to portfolio copilot to agentic allocator, investors will need an independent record of how these systems actually behave.
CapitalBench is starting with public-market portfolios because they are measurable, time-stamped, and brutally objective.
Prices move. Outcomes arrive. Narratives get tested.
Support
Help build the public record
If AI is going to influence capital, the world needs a scoreboard before the money moves.
CapitalBench is building that scoreboard in public.
We are looking for supporters, collaborators, investors, model builders, researchers, and financial platforms who believe AI capital allocation should be measured with real evidence.
For research partnerships, data access, sponsorships, or API access, get in touch.