Why We Use AI to Screen Coins, and Where It Fails

The objection is fair and we get it constantly. Who are you to rule on this, and why would I trust software over a scholar?

The answer is that we are not doing the thing you are worried about. There is an old and useful distinction in Islamic scholarship between tahqiq, the verification and establishment of facts, and ijtihad, the juristic reasoning applied to those facts. A mufti issuing a ruling needs both, and historically the same person did both because the facts were things a person could observe.

The facts are no longer observable that way. If you ask a scholar whether Hyperliquid is permissible, an honest answer requires knowing that the protocol charges funding rates, that over 33% of its revenue comes from perpetuals, that roughly 99% of retained trading fees are recycled into token buybacks, and that the foundation treasury's yield strategy is undisclosed. None of that is in a book. It is spread across a whitepaper, a docs site, a GitHub repository, a tokenomics page and an on-chain contract.

That is the part we automate. The ruling is not the automated part.

What the engine actually does

Our methodology describes two steps, and the split is the important thing.

Step one, fundamental analysis. The engine reads primary sources: whitepapers, technical documentation, GitHub repositories, and tokenomics. From those it extracts a specific set of facts. What is the business model. Where does revenue come from and in what proportion. What is the token actually used for. Is the supply fixed or discretionary. Does a mint authority exist. Can anyone freeze a balance. Is there a buyback, a fee share, or a distribution to stakers.

Step two, reasoning. Those facts get weighed against Islamic finance principles and established fatwas to produce a verdict. This is where the three-layer framework, the property test, and the AAOIFI thresholds are applied, and it is reviewed rather than published raw.

The engine is a research assistant with unlimited patience for documentation. It is not a jurist and we do not present it as one. Every asset page ends by saying that final religious authority rests with a qualified scholar, and that is not boilerplate, it is the actual division of labour.

Why not just do it by hand

The honest answer is scale, and I would rather admit that than dress it up.

Screening one asset properly means reading a whitepaper, the docs, enough of the repository to know whether the docs are true, and the tokenomics closely enough to spot a buyback clause. Call it a few hundred pages. Across 91 published verdicts, that is a library.

A team our size doing that manually would produce perhaps four or five assets a month, which means a Muslim asking about the coin they bought last week would wait two years. The alternative on offer was not careful manual screening. It was the thing that already existed: a list of twenty tickers with a green tick and no reasoning, published by someone selling leverage.

Where it gets things wrong

This is the part most people writing about their own AI tooling leave out, so here is the specific failure list.

It can state a number that is not in the source. Language models produce plausible figures. A revenue split that sounds right and is not in any document is the most dangerous output we handle, because it is checkable and nobody checks it. Our mitigation is that verdicts have to point at where a claim came from, and figures we cannot source do not become thresholds. Where treasury composition is genuinely undisclosed, as with Hyperliquid and Ondo, the correct output is that it is unknown, and you will see us say that rather than estimate.

It goes stale. A verdict is a photograph. A protocol that adds a lending product or rotates its reserves into Treasuries has changed its revenue mix, and our page has not. This is a real limitation of every screener including ours, and it is worse for actively developed protocols than for Bitcoin.

It cannot read what nobody wrote. The engine is excellent on documented mechanics and blind to undocumented practice. What a foundation actually does with its reserves, what an informal agreement with a market maker says, whether a renounced contract really was renounced. This is exactly why so many verdicts land at Doubtful: the failure to find information is reported as a failure to find information rather than as a clean bill.

It over-trusts the project's own documentation. Whitepapers are marketing. A protocol describing its yield as protocol revenue when it is interest on Treasuries will be believed unless something contradicts it. Reading the repository helps and does not fully solve it.

It has nothing to pattern-match against genuinely novel structures. Ethena's synthetic dollar, which holds its peg by delta-hedging with short perpetual futures, is not a stablecoin in any prior sense. Getting USDe right required recognising that the peg mechanism itself was the violation, which is a conceptual judgement rather than a retrieval task. Novel structures are where human review earns its place.

It confuses tokens with the same name. This one is not hypothetical for us. When we built the swap allowlist for our DEX, we keyed it on the combination of chain ID and contract address rather than on the ticker, specifically because a verified token trading as TON turned out to be a tokenized AT&T share. A screener matching on symbol would have applied a Layer 1 verdict to an equity. Any platform that screens by ticker has this bug and may not know it yet.

What we do about it

Four things, none of them complete.

Primary sources only, with the repository read against the documentation rather than instead of it.

Unknowns reported as unknowns. 29 of our 91 verdicts are Doubtful, and a large share of those are cases where the engine found a gap rather than a violation. A screener with no middle category has to guess in those cases, and there is an honest comparison of that trade-off against the platforms that made the other choice.

Human review before publication, concentrated on the novel structures where retrieval is least reliable.

Reasoning published in full, which is the only mitigation that actually scales. If our figure on Hyperliquid's buyback is wrong, the way that gets caught is a reader who knows the protocol reading the page and telling us. That does not work if we publish a tick.

What this does not do

It does not make us a religious authority, and if you are looking for one, we are not it and we do not have a Shariah board.

What we are trying to be is the layer underneath. When someone takes a question to their imam, the useful version of that conversation is not "is crypto haram." It is "this protocol earns a third of its revenue from funding rates and uses almost all of it to buy back the token I would be holding, what do you think." A scholar can answer the second question well. Nobody can answer the first.

Providing the facts for the second conversation is the job. The screening list is 91 attempts at it, and every one of them shows its work so you can bring the work with you.