LEGAL AI FIELD TEST 001 — EVIDENCE PACKET
Can AI Brief an Appellate Opinion Accurately?
Integrated Cognition, prepared with Codex assistance.
Published October 4, 2026. Jonathan Nessler's legal review is pending.

Test date: October 3, 2026, America/Chicago.
Run timestamps use UTC, which had crossed into October 4.
Source: Geller v. Uber Technologies, Inc., 2026 IL 132066.
Official opinion: https://www.illinoiscourts.gov/resources/7283c197-ffad-4825-97eb-033edad67bcd/file
Report address:
https://integratedcognition.com/field-tests/appellate-opinion-briefing

CONTENTS
geller-2026-il-132066.pdf: supplied 26-page opinion, preserved unchanged.
geller-source.txt: exact pypdf layout extraction supplied to the agents.
protocol.json: source hashes, exact task prompt, planned runs, frozen criteria.
prompt.txt: exact case-brief task prompt, including its original source-file path.
run-1.txt, run-2.txt, run-3.txt: unedited saved final responses. All planned runs retained.
observations.json: responses, hashes, UTC timestamps, comparison excerpts and notes.
comparison-notes.md: AI-assisted comparison and prompt-adherence explanation.
reviewer-worksheet.pdf: blank worksheet for a named reviewer's source check.
evidence-manifest.json: SHA-256 checksums for all other files in this packet.

LIMITS
One opinion, one source-only prompt, one configured model alias, three runs.
Runtime-reported parent alias gpt-6.1-sol and reasoning effort ultra were inherited.
An immutable model snapshot, temperature and random seed were not exposed.
Fresh delegated agents received no prior conversation or answer key. Platform
instructions and tools remained present. The task prohibited other sources.
No general accuracy rate, vendor comparison, legal validation, API cost,
lawyer review time, or time-saving claim is established. UTC start/end times
include workspace activity and are not isolated model inference durations.
The comparison is AI-assisted and provisional; a human reviewer may revise it.
The date's source pointer was clarified to the caption in the comparison while
preserving the frozen protocol. The source was not checked for subsequent treatment.
