Attorney-supervised AI
Public Defense AI Pilot
A public-safe adoption packet for defense record triage, mitigation support, chronology packets, and attorney review.
No client names, case names, docket details, sealed facts, privileged material, office documents, or identifying facts are published.
Proof surface
| Claim | Public proof | Private boundary | Status |
|---|---|---|---|
| Attorney-supervised workflow | AI pilot PDF and public page describe review packets, chronology, and source trails. | No client, case, docket, sealed, privileged, or identifying material is published. | Public brief |
| Federal-defense context | Redacted federal-defense summary is linked as a public artifact. | Office documents and case-level facts remain private. | Redacted |
Attorney-supervised workflow
- Public proof
- AI pilot PDF and public page describe review packets, chronology, and source trails.
- Private boundary
- No client, case, docket, sealed, privileged, or identifying material is published.
- Status
- Public brief
Federal-defense context
- Public proof
- Redacted federal-defense summary is linked as a public artifact.
- Private boundary
- Office documents and case-level facts remain private.
- Status
- Redacted
What I built
Legal records are timelines, people, stakes, and decisions that need to remain traceable. The pilot frames AI as infrastructure around the file: organize the record, surface relevant material, and keep counsel in control.
The public materials show source trails, review checkpoints, allowed uses, prohibited uses, and generated text that can be checked against the underlying file.
Pilot scope
The first useful unit is a review packet: a constrained output that makes a lawyer faster without hiding the source burden.
| Workflow | Output | Review control |
|---|---|---|
| Record triage | Chronology rows tied to files and pages. | Date, source, and confidence fields. |
| Mitigation support | Life-history and treatment issue maps. | Attorney decides narrative, legal theory, and use. |
| Omission review | Missing-source and conflict flags. | Explicit status instead of polished silence. |
| Training packet | Checklists, risk notes, examples. | Review rubric before broader adoption. |
Governance frame
- Attorney review is mandatory. AI output is an index or triage layer, not advocacy.
- Confidentiality comes first. No live confidential record should enter a tool without office-approved privacy, retention, and access controls.
- Source traceability is required. Outputs should prefer page, file, excerpt, or record references over fluent prose.
- No black-box conclusions. The system should surface uncertainty, conflict, and omitted context.
- Bias and hallucination are operational risks. Review includes citation checks, source mismatch checks, omission review, and client-dignity review.
Thirty-day version
The first office-safe pilot can stay small: one supervising attorney, one attorney or mitigation specialist using the packet, one technically fluent operator, and one reviewer responsible for citations, omissions, and failure modes.
- select one records-heavy task
- prepare de-identified or approved material
- generate structured packet
- attorney review and correction
- expand, revise, or stop
What I learned
Legal AI gets weaker when it sounds polished before it becomes accountable. Defense work needs the opposite rhythm: start with the file, locate the human stakes, and use automation only where another person can check the result.