Cross-cutting feature
The Three-Model Consensus Gate.
Every Issue Spotter run is scored independently across three frontier models. OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini. The configured threshold is 70 of 100 from each. If all three do not clear, the pathway does not move forward automatically. It is the platform’s principal structural control against AI hallucination on the defense side.
AI On This Page
Buddy Runs OpenAI’s ChatGPT, Anthropic’s Claude, And Google’s Gemini In Quorum Before Court-Facing Language Reaches The Attorney.
Buddy is The IMC Machine’s AI assistant. Talk, upload, or speak – Buddy turns what you send into structured records you approve before anything moves.
Talk To Buddy
“Run Quorum On The Suppression Motion Draft”
Buddy gets three independent reads and shows the diffs side by side for the attorney’s review.
Upload A Photo
Drop A Precedent Case PDF.
Buddy reads the case and includes it in the quorum’s reference set for the next pass.
Voice Note
Voice-Direct The Attorney’s Review Focus.
Buddy highlights only the disagreements so the human review time stays on the load-bearing issues.
Buddy autofills are reviewable. Nothing publishes, files, or sends without a named human approval on the record.
Why it exists
Single-Model Legal AI Has A Structural Problem. Defense Work Demands The Quorum.
Large language models hallucinate. The phenomenon is well documented in the legal-AI literature and in court orders where attorneys have been sanctioned for filing motions containing fabricated case citations produced by a single model. The risk is not theoretical and it is not solved by prompt engineering. It is a property of the systems themselves. If a single model invents a fact, invents a citation, or scores a pathway with false confidence, the downstream cost is borne by the attorney and, in the worst case, by a defendant who is presumed innocent and has not been convicted of anything. The platform’s posture is that this risk has to be controlled at the structure of the system, not at the discipline of the user.
How the gate works
Three Independent Scorings, One Calibrated Threshold, One Durable Vote Log.
01 The three models
Independent Scoring On The Same Input
OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini each evaluate the candidate pathway against the reviewed record of what the State produced in discovery, the firm’s authority preferences from the matter dossier, the required findings for the pathway, and the source citations. The models do not see each other’s responses. Their independence is the point.
02 The 70-of-100 threshold
Three-Of-Three Or It Does Not Advance
The score is a calibrated confidence assessment that the pathway is supported by what the State produced, at the level the platform requires before it will proceed automatically. A single model below the threshold halts the automatic pathway. Two-of-three is not a passing condition. Three-of-three is.
03 The vote log
The Attorney Reads The Same Record The Gate Read
The worker writes run status, attempt count, per-model scores, the recommended motion IDs, the vote-log path, and the error message if any. The vote log is durable, attached to the matter dossier, and available to the attorney before approval. Pathways that do not clear surface for human review with per-model scores visible and the reasons each model fell below the threshold legible.
What the gate is not
The Gate Sits Ahead Of The Attorney’s Time. It Does Not Replace The Attorney’s Decision.
Every motion that clears the consensus gate still passes through the Attorney Approval Ceremony. The attorney reads the draft. The attorney verifies the citations against the source moments in the source media. The attorney completes the case-specific argument placeholders the platform deliberately leaves blank. The attorney signs. The attorney files. The Issue Spotter does not file with the court automatically. The gate reduces the volume of AI output that reaches the attorney’s desk to the volume that three independent frontier models concur is supported by what the State produced. The gate does not relieve the attorney of the responsibility to verify what reaches the desk.
- Reads from The Evidence Orchestrator: accepted flags, worksheet answers, citations, matter dossiers across body-camera MP4, 911 MP3, interview audio, discovery PDFs
- Lives inside The Issue Spotter: queued background worker between candidate analysis and motion drafting
- Hands off to the Attorney Approval Ceremony: firm-staff authentication, bar-number match, document body, citation verification before approval is recorded
- Nothing files automatically. EFileTexas envelope ID is recorded manually after approval, as proof rather than as action the platform took
Next Step
See The Consensus Gate Run On A Real Matter.
Why This Matters
Hallucination, Controlled At The Structure.
Large language models hallucinate; the phenomenon is well documented in the legal-AI literature and in court orders sanctioning attorneys for filing motions containing fabricated citations from a single model. The risk is not theoretical and is not solved by prompt engineering. With official misconduct a factor in 79% of 2024 homicide exonerations5 and $156M paid by Texas to exonerees since 20096, defense work cannot absorb a single model inventing a citation or scoring a pathway with false confidence. The three-model consensus gate scores every Issue Spotter run independently across OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini. The configured threshold is 70 of 100, from each. Three-of-three or the pathway does not advance automatically. Aligned with ABA standards for indigent-defense systems7, it is the platform’s principal structural control against AI hallucination on the defense side.
Sources
Where The Numbers Came From.
Sources and citations
- RAND / NCSC / ABA, National Public Defense Workload Study (2023). rand.org; ABA Journal coverage.
- Sixth Amendment Center, State of the Nation on Gideon’s 60th; Bureau of Justice Statistics, State Government Indigent Defense Expenditures.
- Texas Tribune, Feb. 27, 2025; SMU Dedman / Deason Center, “Country Justice in Texas.”
- Colorado Public Radio, Jan. 2, 2026; The Journal.
- National Registry of Exonerations, 2024 Annual Report. DPIC coverage; Registry homepage.
- Zealous Advocate, $156M paid to Texas exonerees since 2009.
- Three-Model Consensus aligns with the Attorney Approval Ceremony; reviewable framework. See ABA Standing Committee on Legal Aid and Indigent Defendants.
- Texas Code of Criminal Procedure Article 39.14 (the Texas rule requiring the State to hand over its evidence to the defense); Watkins v. State (2021) analysis; 2025 remedies update.
Learn This Product
Want the working order, in a course?
Continue Learning
Learn The Evidence Orchestrator
Five sequential modules on defense-side evidence preparation: the matter record, multi-format intake, the Cockpit, the three-model consensus gate, and the Article 39.14 native-file posture.
Open The Course →Continue Learning
Learn The Issue Spotter
Five sequential modules on defense-side motion practice: the firm-control premise, the dismissal pathway library, evidence-to-motion drafting, the six-state lifecycle, and the attorney approval ceremony.
Open The Course →