1The problem
Having a view about the world is easy. Turning it into a position is not. A person who believes that power, not chips, is the real constraint on AI has to translate that belief into a list of tickers, argue with themselves about how much of each to hold, place the orders, and then keep the whole thing current as the story moves. Most views die in that translation.
Index products solve the sizing problem but not the expression problem. You can buy a semiconductor ETF. You cannot buy “the specific companies that get paid when a data centre needs its own substation.” The theses people actually hold are narrower than any product that exists.
2What a theory is
A theory is three things bound together: a sentence, a weighted basket of listed companies that expresses it, and a token whose supply is backed by that basket. The sentence is the primary artifact. The basket is derived from it. The token is how you hold it.
Every holding carries one sentence explaining why it is in this basket for this thesis, not a description of the company. That constraint is what separates a theory from a screen.
3The universe
Theory can only hold equities that exist as tokens on Robinhood Chain. That set is public and readable: tokenized equities on this chain are named in the pattern Company • Robinhood Token, so they can be filtered out of the chain's full ERC20 list. As of Sep 2, 2026 there are 194 of them, read from the explorer at https://robinhoodchain.blockscout.com.
This is a hard boundary, and it is the boundary that makes the product honest. The model is handed that exact list and may only choose from it. Token addresses are attached afterwards from the same list rather than coming from the model, so a ticker the model invents is dropped before it can ever be displayed, let alone bought.
4Composition
A language model receives the thesis and the catalogue and returns tickers with weights in basis points, a reason per holding, a grouping into pillars, and a list of what would break the thesis. Three rules are enforced in code rather than trusted to the model.
First, weights are renormalised to sum to exactly 10000 basis points by largest remainder after any invalid holding is dropped, so a basket always totals 100 percent. Second, pillar weights are computed from the holdings that survived rather than taken from the model, so the breakdown can never describe a position the basket does not contain. Third, the model may decline: if the catalogue cannot honestly express the thesis it returns a refusal and the interface says so instead of manufacturing a basket.
The catalogue is sorted alphabetically by company. Sorted by holder count it puts the megacaps first and the model anchors there, which produced a pure semiconductor basket for a thesis about obesity drugs while the pharmaceutical names sat unread further down the list.
5The vault
Each theory maps to a vault contract on Robinhood Chain, an Arbitrum Orbit L2 with chain id 4663. The vault is itself the share ERC20: there is no separate share token, so holding the vault token is holding a claim on the basket. Deposits mint shares proportional to the vault's existing value, redemptions burn them.
Redemption in kind always works and is never pausable. A holder can always take their proportional slice of the underlying tokens out, regardless of the state of any external venue, keeper or price source. This is the property that makes the rest of the design tolerable.
6Fees
A protocol fee of 0.5% and a creator fee capped at 2% accrue to the vault and are pulled by the recipient rather than pushed on every deposit, so a recipient that cannot receive funds can never block deposits for everyone else.
7Revision
A thesis that cannot change its mind is a dead thesis. The creator may revise a basket, and the model proposes revisions as the news moves. Every revision is recorded in a public changelog on the theory's page with the date and whether a person or the model made it. A reader can always see what the basket used to be and why it is not that any more.
8What is not built
The vault factory is deployed on Robinhood Chain at 0xa4d7a8bb029713f0075f325ca05a01c68e7459c6, with all 194 tokenized equities seeded into its allowlist. What is not built yet is the wallet connection in this interface, so no theory can be bought from this site today. No figure for value locked, holders or return appears anywhere, because there is nothing real to report yet.
More importantly, Robinhood Chain has no exchange. Until one exists, a vault can accept a deposit but cannot convert it into the stocks the basket names. The contract handles this by holding deposits rather than pretending to trade, and we would rather not take deposits at all than take them under a description that is not true. The composer, the universe and the baskets are real today. The buying is not.
9Risks
Baskets are generated by a language model. They are not investment advice, they have not been backtested, and no claim is made that any of them will outperform anything. A tokenized equity carries the risk of the underlying company plus smart contract risk, custody risk in the tokenization layer, and the risk of the chain itself. Concentration is deliberate in this product, which means single name and single narrative risk are higher than in a broad index, not lower.
Theory currently holds 18 published theories, all generated from the sentences shown on their pages.