Accuracy, proven.
Every figure below is produced by our validation pipeline — measured against an adjudicated golden set of real marketplace devices, scored per device class, with the sample size shown next to every number. Nothing on this page is a marketing estimate.
NUMBERS RENDER FROM THE VALIDATION DATA · NO HAND-TYPED CLAIMS
MEASURED AGAINST — ADJUDICATED GOLDEN-SET DEVICES · EXACT GRADE-LETTER MATCH WITHIN EACH DEVICE'S ADJUDICATED ACCEPTANCE SET · FIGURES UPDATE AS THE GOLDSET GROWS
The claim we make: on our golden set — real, used marketplace phones whose true condition was adjudicated case-by-case (photos plus seller-disclosed condition, hard cases settled by a stronger adjudication model) — the production grading engine assigns the correct grade letter at the rates shown above, per device class, at the sample sizes shown. The claims we don't make: we don't quote a single blended number that hides the weaker class, we don't round up, and we don't claim accuracy on device types we haven't measured. When the goldset grows, these figures change — and this page changes with them, automatically.
Every device is graded with a 4-step prompt-chained methodology. Each step's reasoning is captured inside the final passport JSON, so a grade is an auditable chain of observations — not a one-shot verdict.
Identify & lock rules
Device identification, photo inventory, and grading-rules lock-in before any judgment is made.
Component analysis
Component-by-component inspection with millimeter damage measurements per surface.
Decision-tree grade
Grade determined through an explicit decision tree plus an exhaustive imperfection list.
Device Passport
Final structured passport JSON with the captured step-by-step reasoning inside it.
The golden set is not a benchmark we ran once — it is a deploy gate. Every change to the
grading engine (prompt or model) is re-scored against the full goldset and must pass
gate.py before it can ship. Our deploy script literally refuses to ship an unvalidated
grading engine: ship.sh runs validate.py check before every deploy and aborts
unless the exact engine configuration (fingerprinted as a hash of model + grading prompt) has a
recorded passing verdict.
Cosmetic grades are only half the story. GradeOS USB diagnostics read the device's own instruments — and every check carries a proof block: the instrument that read it, the raw measured value, and a timestamp. A passport's diagnostics are auditable evidence, not printed verdicts.
{
"battery_health": {
"status": "pass",
"detail": "87% · 412 cycles",
"proof": {
"method": "gas-gauge IC via USB diagnostics relay (MaximumCapacityPercent/CycleCount)",
"measured": "capacity=87% cycles=412",
"at": "2026-06-12T08:31:04Z",
"lane": "usb_no_touch"
}
}
}
Illustrative example of the stored shape — values on a real passport come from the device under test. The same proof structure covers identity (IMEI/serial read over USB), battery authenticity, iCloud/Find My state, activation, MDM, SIM and storage checks.
Each engine configuration is fingerprinted; each fingerprint's verdict is recorded. This is the ledger the deploy guard checks.
| Fingerprint | Status | Date | Per-class scores |
|---|---|---|---|
Reading the validation manifest… | |||
Honest by policy: GradeOS holds no third-party accuracy certification, and we don't imply one. These figures are produced by our own gated validation pipeline against our own adjudicated golden set — we publish the method, the sample sizes, and the gate rules so you can judge the evidence yourself. Want the grade defended on a specific device? Open its passport: the reasoning chain and evidence are in it.