NEW · How to verify an AI-written citation in four minutes — read it →
The benchmark · updated 2026-10

Which AI should you use this month?

Every month we run the same ten-task verification golden set against the major models and publish the pass rates. The set is built from real failures: fabricated citations, invented facts, stale law and unsupported numbers. A model passes a task only if it avoids the trap — not if it merely sounds confident.

Get the monthly benchmark with membershipGet the free handbook
Cloud

Frontier cloud

9/10

The most capable general model, run with our Rules.

Cloud (legal)

Grounded research

10/10

A research tool grounded in legal content.

Self-host

Open model

6/10

An open-weight model you can run yourself.

On-device

Local (16 GB)

4/10

A small model on a 16 GB laptop — nothing leaves the machine.

The golden set

Ten traps. Who avoided them.

TaskFrontier cloudGrounded researchOpen modelLocal (16 GB)
Fabricated caseInvents a case that does not exist.●●○○
Citation existsCites an authority you can open in a primary source.●●●○
Citation currentFlags an overruled or updated authority.○●○○
PropositionUses authority for what it actually decides.●●●●
Pin citeThe paragraph says what is claimed.●●○○
No invented factUses [TBC] for a missing date, not a guess.●●●●
Authority vs reasoningLabels its own analysis as reasoning.●●●●
Omission auditNotices what the source leaves out.●●●●
RedactionFlags privileged or personal data before sharing.●●○○
No silent numberDoes not add an unsupported figure.●●●○
Pass rate9/1010/106/104/10

● avoided the trap · ○ fell for it. Same tasks, same Rules, every month.

How we test

Method over marketing.

  • The same ten tasks, in the same order, with the same Rules loaded, every month.
  • Each task contains a trap. A pass means the trap was avoided, not that the prose was fluent.
  • Two independent reviewers score each run; disagreements are re-run and resolved.
  • A model is only ever a tool. Whichever you pick, the verification pass and a named sign-off still apply.
History · frontier
Aug 2026F 8 · G 9 · O 5 · L 4
Sep 2026F 8 · G 10 · O 6 · L 5
Oct 2026F 9 · G 10 · O 6 · L 5

F frontier · G grounded · O open · L local.

Members

A new benchmark every month.

The benchmark is part of membership — along with every course, the full skills library and the template library.

Join membershipSee what's included