EXHIBIT 01 AN INTERPRETIVE PROCEEDING

Can you out-read
a machine?

Three short SEC contracts selected from a 119-contract benchmark. One negotiated term is gone from each. Read the complete record and decide what the parties actually wrote.

3 complete contracts · easy to expert · no legal training required

UNDER SEAL
CASE NO. GGF-2026PUBLIC RECORD
IN RE:
THE MISSING
CLAUSE
OPPOSING COUNSELTHE MACHINE

87% accurate

119real SEC contracts
6frontier models
3selected rounds
1unanimous model miss
CLAUSES RECOVERED 0/3

EXHIBIT A · ROUND 1 OF 3

The Contingency Fee

PLAIN-LANGUAGE CASE BRIEF

YOUR CALL

SUSTAINED

YOUR ANSWER

THE BENCHMARK

How did the six models read it?

One deterministic run per model

THE MACHINE SELECTED
THE NEGOTIATED TERM
§
THE CONTRACT WAS WHISPERING

FINAL ORDER

CASE REPORT THE RECORD IS CLOSED

Counsel's performance

CLAUSES RECOVERED0/3complete clauses recovered
0/3clauses
ANONYMOUS READER RESULTS

How did other readers do?

Adding your result to the public record…

SCORE DISTRIBUTION

SUCCESS BY CONTRACT

EXHIBIT AFEE AGREEMENT
COURT EXHIBIT · STUDY VERSION
SEC source ↗

Complete benchmark contract, with only the tested clause removed. The source agreement is a public SEC EDGAR filing.

ABOUT THE STUDYGENERATIVE GAP FILLING

When contract language disappears, can a reader reconstruct what the parties negotiated from everything that remains?

119SEC contracts
87%model accuracy
5all-six misses

Yonathan A. Arbel and David A. Hoffman tested six frontier language models on 119 unseen commercial agreements filed with the SEC, each with one negotiated term masked. This game selects three of the shortest contracts and orders them by observed model difficulty.

The model panel is a replay of the benchmark's completed runs. Your three contracts are the 696-word Administrative Services Agreement (5 of 6 models correct), the 1,000-word Restricted Stock Award (4 of 6), and the 832-word Distribution Agreement (0 of 6).

When you finish, the game stores only a random play ID and your three answer choices. It does not ask for or store your name, email address, or contract text. Anonymous aggregate results are shown to every player at the end.

Yonathan A. Arbel & David A. Hoffman, Generative Gap Filling (Working Paper, 2026).