AI reads. It does not decide.

That is the whole answer, and most of what follows is the detail behind it. If you have seen "AI-powered" on a crypto rating and wondered whether it meant anything beyond a press line, this is what the words cover at CryptoGrade, and what they are not allowed to cover.

The hard part of fundamental crypto research has never been arithmetic. It is that every asset documents itself differently.

A settlement chain publishes validator counts and client diversity. A lending protocol publishes value locked and fee totals. A storage network publishes commits and release notes. One project's supply schedule is in a blog post, the next one's is in a governance forum thread, a third one's is in a contract you have to read. None of it arrives in the same shape, and almost none of it arrives in the shape you would need to put two assets side by side.

So an investor holding BTC and ETH plus a dozen other things ends up comparing a chain to a protocol using whatever each one chose to disclose. That is not a comparison. It is two different conversations happening at once.

Our AI does the collection and the structuring: pulling the documentation, supply data, development activity and on-chain economics for each asset and putting all of it into the same form, so that the same questions can be asked of assets that have nothing else in common. That is the work that used to make genuine comparison impossible for one person to do at scale, and it is the work AI removes.

Here is the line that matters, and it is the one most "AI-powered" claims in crypto leave vague.

Our model does not choose what counts, how much it counts, or where the boundaries sit. Those are fixed before any asset is looked at, and they are published. Every asset is scored on the same six weighted dimensions, in the same proportions:

On top of the weighted score sit nine pass or fail disqualifier screens, covering things like anonymous or unaccountable teams, custody that can freeze or seize, supply concentrated in insider wallets, and a history of rug, abandonment or exploit. The screens are not scored. They are gates.

Then the bands, which are also fixed: A is 80 and above, B is 60 to just under 80, C is 40 to just under 60, F is below 40.

The important consequence is structural. Because the weights and the bands are fixed in advance, our model cannot reach a letter it likes and work backwards to a score that produces it. It fills in a fixed scale. Whatever the AI thinks of an asset, it has one place to put each piece of evidence and one weight that piece carries, the same place and the same weight it would carry on any other asset.

That is what makes two of our grades comparable. Not the intelligence of the model. The fact that the model did not get a vote on the scale.

This is the part that tells you most about an AI rating, and it is the easiest part to skip.

An AI asked to produce a score will always produce one. The interesting question is what it does when the evidence it would need does not exist. There are two honest options: say so, or fill the gap with something plausible. Only one of those is research.

Here is our own record on a named asset, read on October 21, 2026, from our record last scored October 20, 2026. Compound (COMP) is graded B / Sound with a score of 73.6, at position 4 in our ranking that morning.

Our model's overall assessment on that record says the supplied evidence supports "a stable, established lending protocol with no evident dilution risk and substantial TVL/fee activity," and then, in the same assessment: "The main limitation is that the record does not establish direct token cash-flow capture, detailed decentralization/audit quality, or enough market microstructure data to justify higher scores."

Read that second sentence again, because it is doing something unusual. It is our model explaining why it did not score the asset higher, and the reason is not a finding against Compound. It is an absence of evidence.

The same pattern runs through the individual dimensions on that record. On tokenomics, our model notes that maximum supply equals circulating supply, a clean supply profile with no pending dilution evident from the supplied data, and then adds that COMP "remains a governance token with no token-holder fee entitlement shown here, limiting direct value accrual." On usage, it records substantial protocol activity, value locked of about $1.51B and thirty-day fees of about $2.64M as the figures stood in our record when it was scored, and immediately adds that "token-holder revenue linkage is not established by the supplied data." On security, it notes a long operating history with no supplied red flags, then states that evidence on audits, client diversity and decentralization depth "is not provided, so this is solid but not top-tier."

Three dimensions, three places where our model names what it was not shown.

That is a deliberate design choice, and it runs through the product. Where a category's evidence is missing, our record says the evidence is insufficient rather than scoring the category zero, because missing data is not the same finding as bad data. Where too much is missing to stand behind a letter, the asset carries no overall grade at all rather than a guessed one.

A rating that never admits a gap is not more confident than ours. It is less informative, because you cannot tell which of its numbers were earned.

Three limits worth having in your head when you read any AI-produced grade, including ours.

A letter is not reverse-engineerable from a score. Our bands put 73.6 in the B range, and Compound is graded B. But a tripped screen places an asset in F at any score, including a score in the A range. So there are two roads into an F: a score below 40, or a failed screen at any score. Our record does not publish which road a given asset took, so we do not guess and we do not name a screen.

A passing screen is not a clean bill of health. The honest reading of a screen that did not trip is that no disqualifying condition was detected, under that rule version, on that basis. It is not a statement that an asset is sound, and we will not write it as one.

A grade is a reading of a record on a date. Every number in this piece is stamped with the record it came from. That is not hedging. A rating that cannot tell you when it was formed cannot tell you when it changed, and a change is usually the part you needed.

Four questions, useful well beyond us:

If an AI rating cannot answer those, the AI part is a description of how the output was produced, not a reason to believe it.

Our framework is written up in full at https://cryptograde.ai/framework and the example in this piece is at https://cryptograde.ai/tokens/compound

A free account costs an email address and opens Alerts at https://cryptograde.ai/alerts , where you pick the crypto assets you care about and ask to be told when our letter grade on one of them changes. If you hold something sitting near the edge of a band, that is the thing worth hearing about.

One note on why this costs an email address rather than money. CryptoGrade does not work commercially with the assets or projects it covers, and it does not sell data. Nobody we grade pays us, which is the only reason a piece like this can print, with the asset's name attached, that our own record does not establish token-holder revenue linkage.

Create the free account at https://cryptograde.ai/sign-up