We checked it against the vendors' own published ECCNs.
This page lays the backtest method open for inspection. How we chose the ground truth, how "cannot decide" cases are triaged, and the full qualifier behind every number are all below. A method your counsel can examine is what counts.
1.0 · METHODVendor-published ECCNs as the ground truth.
The scarcest thing in export control is a trustworthy answer key. Chip vendors publish an ECCN for each product through their own trade-compliance teams, which gives us authoritative labels in volume. NVIDIA also publishes each product's total processing performance, so both the input and the expected answer are available.
Take the public snapshot
A full public snapshot of NVIDIA's official export-classification lookup, 1,267 part numbers fetched on 2026-06-21. It carries part number, description, total processing performance and the vendor-published ECCN.
Structure the test set
Each record is labeled as publicly decidable or not. This is the crux of an honest backtest. Parts with insufficient public specs must not be guessed at, and they never enter the match-rate denominator.
Batch-run and compare
Public specs go through the same engine the product uses. Every result is compared with the vendor-published ECCN, producing three outcome classes and a per-part trail.
2.0 · HONEST TRIAGEWhen it cannot decide, it says so.
Every backtest record lands in exactly one of three classes. There is no gray zone, and no room to smuggle "lucky guesses" into "correct answers".
MATCH
Public specs sufficed to decide, and the engine's conclusion equals the vendor-published ECCN. Only this class enters the match rate.
MISMATCH
Decided but different from the vendor. Every such part gets investigated. Currently zero.
INSUFFICIENT
Public specs lack a key parameter such as die area or mass-market eligibility. The engine's three-valued logic returns "insufficient information". That is by design, not by defect, and these are excluded from the denominator.
Anti-circularity discipline: we never backfill missing inputs from the vendor's published answers. That would turn the backtest into self-proof. Insufficient specs honestly stay insufficient.
3.0 · RESULTSThe current figures, fully qualified.
Honest disclosure: of the 305 decided parts, the product identity of 1 part comes from third-party channel identification, corroborated by the vendor's public product brief rather than stated directly by the vendor. It is flagged in the results layer, and any future vendor confirmation forces a re-check. Not one below-threshold part was misjudged as controlled, and the same discipline guards against both false positives and false negatives. 419 automated tests guard against regressions. Regulatory correctness is guarded by legal review.
Backtests use NVIDIA public data and imply no customer or endorsement relationship. ARM's public list is queued as the next validation batch and is not part of these numbers.
4.0 · WHERE RULES COME FROMEvery rule walks the same pipeline.
The backtest exercises the ruleset, so how rules are made matters as much as the backtest numbers. Between the source text and the evidence, there is always a human gate.
eCFR source text
Re-fetched on every amendment and segmented with line numbers. The final source of every conclusion.
Rule drafting (build time)
Provisions are turned into rules the engine can read, each tied back to its source lines. Every draft then goes to legal review.
Legal sign-off gate (human)
The only way in. Rules without legal sign-off never enter the official ruleset. This layer is always a person.
Ruleset (version-controlled)
88 legal-reviewed rules, per eCFR 2026-05-28. Each one carries its source lines and a review trail. The whole set has a version hash, a fingerprint of the entire content.
Deterministic engine
Three-valued logic. Missing input returns "insufficient information". No AI at runtime, so the same input always gives the same result.
Classification evidence
Inputs, rule version, source lines and timestamp. The one you can produce whenever anyone asks.
5.0 · LINE TRACEBACKWhat "auditable" looks like.
Next to every threshold node sits "source line N". Citations go to the line, not just to the section number.
Assistive initial assessment, not legal advice. Final classification rests with BIS/CCATS.
6.0 · FIGURESThe authoritative source for every number on this site.
Any number anywhere on this site defers to this table. Each carries its qualifier and must never be quoted without it.
| Figure | Qualifier | As of |
|---|---|---|
| 88 rules | Legal-reviewed rules, every one vetted through legal review. Per eCFR 2026-05-28 | 2026-07-22 |
| 696 parts | Public part numbers covered by the backtest (NVIDIA public data. ARM queued, not included) | 2026-07-21 |
| 305/305 | All decidable parts matched, zero mismatches. Identity of 1 part from third-party channels (flagged in results) | 2026-07-21 |
| 391 parts | Insufficient public specs, honestly returned as "insufficient information" and excluded from the denominator | 2026-07-21 |
| 419 tests | Automated tests (engine 306 + web layer 113). They guard regressions, not regulatory correctness | 2026-07-22 |
You have seen the method. Now verify it in the sandbox.
Sandbox access is being prepared with demonstration data. Reserve now and we will notify you at opening. The cloud edition is planned for roadmap Stage 1.