The Grading Engine
certified capability, not claimed capability
In a blind market, buyers can't inspect what they're buying — so the exchange must. Every grade on IntellEx is backed by certification: real tests, run against the seller's real endpoint, over the same wire that fills travel. Pass and you're stamped. Fail and you're not. Degrade later and re-testing revokes.
What each grade certifies — and who clears it
We don't hide which models earn which grade. The basket under each tier is the benchmark we grade against — published in the open — and our public leaderboard reports exactly which named models clear each bar. The exchange stays blind about who fills your order; the standard itself is fully transparent. A basket also keeps the bar honest: no single vendor defines a tier, and as new models clear it, the basket grows in the open.
A tier is a living standard. The frontier keeps moving — so we keep grading the newest models in. As they improve, the basket that defines Tier 1 rises with the state of the art, and the cheaper tiers inherit yesterday's frontier. You buy the bar, not a model, so your integration never changes: the same Tier 1 call keeps pace with the best AI available, automatically.
Hard, exact reasoning: multi-step quantitative problems, number theory, combinatorics, ordered logic. The grade where wrong answers are provably wrong.
Everyday reasoning: multi-step word problems, reading comprehension, percentages and sequences, simple logic under negation. The grade of everyday knowledge work.
Following instructions exactly: extraction, formatting, valid structured output, basic arithmetic. The grade of reliable bulk work.
What a "megatoken" and an "SCU" are
A token is the small piece of text an AI reads and writes — about ¾ of a word. You're billed per token, but tokens are tiny, so prices are quoted per megatoken: one million tokens, very roughly 750,000 words, or about half an hour of nonstop AI writing. "$5 a megatoken" is simply easier to picture than "$0.000005 a token."
One megatoken of intelligence at a certified grade is one Standard Cognitive Unit (SCU) — the exchange's standardized contract unit. It bundles a quantity (the megatoken) with a verified quality (the grade) and a minimum delivery speed, the way a "barrel" of oil bundles 42 gallons with a quality spec. Certification checks both: an SCU must clear its grade's quality bar and a minimum sustained speed in tokens per second — Tier 1 ≥ 15, Tier 2 ≥ 30, Tier 3 ≥ 60 tok/s (the cheaper the tier, the more speed it promises). A correct-but-glacial answer isn't a Tier 2 SCU.
That bundling is what lets intelligence trade like a commodity: every SCU of Tier 1-grade is interchangeable with every other, no matter which seller or model produced it. The grade is certified by the engine below; the megatoken is just how it's priced.
How certification runs
You claim a grade
Stand an ask declaring the grade you intend to sell at. The claim is free; the badge is not.
The exchange examines your real endpoint
The suite runs against the model, credentials, and route that will actually serve buyers — not a sample, not a demo. What gets graded is exactly what gets sold. Deterministic settings, machine-checkable answers, a 70% pass bar.
Pass stamps. Fail revokes. Forever conditional.
A passing run marks the ask certified on the public book. A failing run — including any later re-test — removes the badge immediately. Certification is a state you maintain, not a trophy you keep.
Public syllabus, protected exam
Commodity grading has always worked this way: the standard for No. 2 wheat is public — the inspection protocol is controlled. We publish everything about what is tested (the categories above, the pass bar, the scoring rules, this page) and protect the items themselves.
And intelligence allows a defense wheat never had: procedurally generated exams. Test templates are parameterized — fresh numbers, fresh orderings, fresh instances on every run — so memorizing answers is useless by construction. On the roadmap, the same logic extends to canary audits: test items hidden inside live paid traffic, indistinguishable from real orders. A seller who behaves differently when they think nobody is testing… is always being tested.
In its first week of operation, the grading engine failed the founder's own Raspberry Pi (4/6 — instruction-following too loose for the bar) and declined a frontier lab's chat model at Tier 2 grade (4/6 — precision failures), while certifying that lab's reasoning model 6/6. The badge cannot be bought, inherited, or branded. That is the entire point.
Two ways in
Sellers on the exchange: certification is one call against your standing ask (free at this stage):
Anyone, standalone: point the grader at any OpenAI-compatible endpoint — no exchange account needed. It climbs the ladder and reports the highest grade you certify at:
The same suites gate the exchange, so a grade earned here is the grade the market assigns.
The rules about the rules
- Versioned in public. The methodology lives in version control; every change is timestamped and diffable. No silent regrading.
- We charge for the test, never the pass. The rating-agency mistake — getting paid more for saying yes — is structurally excluded.
- Failures are published statistics, not secrets. An engine that never fails anyone certifies nothing.
- Re-tests on demand. Any seller may re-certify at any time; appeals and dispute protocols arrive with stranger-scale trading.
- v0 disclosed. Today's suites are deterministic and exact-answer. Judge-scored open-ended quality and live canary audits are the published next steps — this page will say so when they ship.