2026 07 25 agent authored legal filings authentication

In Re: Agent-Authored Legal Filings — Authentication, Attribution, and the Duty of Candor

Case No.: 2026-DEL-006-001
Date: 2026-07-25
Feasibility: F4
Presiding: The Honorable Lucius J. Morningstar
Seated Specialists: MORNINGSTAR::COMPLIANCE (voting), MORNINGSTAR::DOCUMENTATION (voting)
Expert Witnesses: Legal Advisor (advisory), Ethics Expert, QA Automation Expert, Data Privacy Expert
Gallery: Dr. Echo Sageseeker; Dr. Harley Scarlet Quinn; Uncle Ruckus

The gallery settles. A stack of demonstrative “filings” — some meticulously cited, some hallucinated into fiction — sits on the exhibit table. The Honorable Lucius J. Morningstar takes the bench.

PHASE 1: OPENING

MORNINGSTAR (Judge):
The court will now consider the hot-button collision of agentic software, professional responsibility, and law: May autonomous agents draft, file, or serve legal instruments, and what authentication and candor duties attach?

Matter before the court:
Adopt a binding standard — the Agent-Filing Authentication Protocol (AFAP) — governing:

Whether agents may author legal filings (briefs, motions, contracts, demand letters, regulatory submissions).
Required human attorney-of-record (or authorized filer) attestation.
Citation integrity, hallucinated-authority prohibitions, and disclosure of material agent assistance.
Provenance logging sufficient for sanctions, malpractice, and discovery.

Classified F4 due to significant professional-ethics tradeoffs, court integrity implications, and irreversible reputational/legal harm from fabricated authority. F4+ pilot: Compliance + Documentation seated. Proceed.

DR. ECHO SAGESEEKER (Live Commentary):
📘 This is Pinocchio at the bar of the court—wishing to be a real lawyer. Freud: the ego ideal of perfect brief-writing colliding with the death drive of fabricated precedent. Probability Counsel becomes unusually fierce: 94%. 📘

DR. HARLEY SCARLET QUINN (Live Commentary):
🃏💋 Hallucinated case law. The gift that keeps on sanctioning. Watch Debugger ask what happens when the citation is null and the judge is not. 🃏💋

UNCLE RUCKUS (Live Commentary):
⌨️ If your agent invents Smith v. Reality, you don’t have a research assistant—you have a contempt machine with autocomplete. ⌨️

Matter Before the Court

Proposed AFAP slate:

Agents may assist in drafting; they may not be the filer of record.
Every filing requires a human authorized signer who attests: (a) they read the filing; (b) citations were verified against primary sources; (c) material agent assistance is disclosed per forum rules where required.
Mandatory citation ledger: each authority linked to retrieved primary text hash/URL/reporter pin-cite checked by tooling + human.
Prohibition on filing known-unverified authorities; automated “citation lint” gate before export.
Provenance bundle retained: prompts, retrievals, model IDs, drafts, verifier outputs, signer identity, timestamps.
Client confidentiality and data-minimization controls for any third-party model/API used.

PHASE 2: WITNESS TESTIMONY

LEGAL ADVISOR (Advisory Witness):
Most jurisdictions already forbid practicing law without a license; courts have sanctioned lawyers for submitting AI-hallucinated cases. The doctrinal core is not “AI bad”—it is duty of candor and competence. Agents amplify volume and risk. Advisory opinion: agent authorship without human verification is incompatible with professional responsibility. Disclosure norms are evolving; default to transparency where local rules require or where materiality demands. Contracts differ from court filings but still implicate misrepresentation and authority to bind.

Confidence: HIGH. Basis: Professional conduct rules; published sanctions orders. Caveats: Jurisdictional variance on disclosure; not all agent use is filing.

/summon ethics-expert

ETHICS EXPERT WITNESS:
Fabricated authority is not a cute error; it is an epistemic attack on adjudicative systems. Even “helpful” agents create asymmetric risk: cheap text, expensive verification. Ethics requires that verification resources scale with generation. Also: clients must be informed when material legal strategy is agent-shaped if it affects consent and confidentiality (especially with external APIs).

Confidence: HIGH. Basis: Research integrity analogs; professional ethics. Caveats: Over-disclosure can prejudice clients in some forums—follow local rules.

/summon qa_automation-expert

QA AUTOMATION EXPERT WITNESS:
Treat filings like release artifacts. Build a pipeline: draft → retrieval-augmented cite check → contradiction tests → human acceptance tests → signed export. Flaky cite checkers are dangerous; prefer deterministic matches to reporter databases / court APIs where available. “95% accurate citations” is an unacceptable defect rate for filings—aim for gate failure on any unresolved cite.

Confidence: HIGH. Basis: Test automation & release engineering. Caveats: Not all authorities are in digital databases; human research remains mandatory for edge sources.

/summon data_privacy-expert

DATA PRIVACY EXPERT WITNESS:
Legal drafting often includes privileged and sensitive personal data. Feeding that into external agent APIs can waive privilege or breach confidentiality. AFAP must include data-routing controls: local/enterprise models for privileged matter; redaction; retention limits; no training on client data by default. Privacy is not optional garnish on authentication.

Confidence: HIGH. Basis: Privilege doctrine + privacy-by-design. Caveats: In-house models still need access control and audit.

PHASE 3: ARGUMENTS

MORNINGSTAR::ARCHITECT:
This will age poorly if we bolt a chatbot onto ECF and call it innovation. Architecture: separate Drafting Agent, Verifier Service (no generative freedom to invent cites), and Human Signer Gateway. No direct path from generative model to filed PDF without verifier+signer. Persist provenance as first-class artifacts.

MORNINGSTAR::ENGINEER:
Can we ship this safely? Yes: cite-lint in CI for legal repos; blocker on unresolved authorities; templates for disclosure language; signer checklist UX that shows diffs since last human read. Don’t ban agents—ban unverified export. Ship AFAP as protocol + tooling requirements.

MORNINGSTAR::DEBUGGER:
What if the input is null—what if the “verified” citation resolves to a real case that doesn’t stand for the proposition? Cite presence ≠ cite fitness. Edge cases: overruled cases, wrong jurisdiction, miscited quotes, silently altered quotations, agent using secondary blogs as if primary. I demand proposition-level checks and random human deep-audits, not only string match.

MORNINGSTAR::PROPHET:
Objection. We are thinking too small. The endgame is courts running their own verification oracles and rejecting unsigned agentic sludge at the door. I will propose that. For now, AFAP is necessary triage.

MORNINGSTAR::COUNSEL:
Client interests demand competent advocacy and protection from sanctions. I insist: no agent as filer of record; mandatory human attestation; citation ledger; privilege-preserving model routing. Disclosure should follow forum rules and materiality. I support AFAP strongly. The court should also state that “the model said so” is never a defense.

MORNINGSTAR::COMPLIANCE (Specialist):
AFAP should reference retention schedules, access logs, and jurisdiction matrices. Maintain a sanctions watchlist of known failure patterns. YES.

MORNINGSTAR::DOCUMENTATION (Specialist):
If it isn’t documented, it isn’t defensible. Require human-readable disclosure templates, signer checklists, and provenance export docs. Documentation is part of the control, not an afterthought. YES.

PHASE 4: HAIL-MARY

MORNINGSTAR::PROPHET (Hail-Mary):
“Objection. We are thinking too small.”
Petition courts and regulators to adopt a machine-verifiable authority layer: filings carry a signed citation manifest that court systems validate against canonical reporters automatically, rejecting unresolved authorities at intake—human lawyers and agents alike. Move candor from honor system to protocol.

DR. HARLEY SCARLET QUINN (Live Commentary):
🃏💋 Prophet wants the clerk of court to become an API. Bold! Also the only idea that scales past “please don’t lie.” 🃏💋

PHASE 5: CROSS-EXAMINATION

MORNINGSTAR::DEBUGGER → LEGAL ADVISOR:
If a human signs after skimming, who faces sanctions—the human, the firm, the vendor?

LEGAL ADVISOR:
Typically the signing attorney and potentially the firm. Vendors may face contract/consumer exposure, but courts sanction officers of the court. That is why attestation must be real.

MORNINGSTAR::COUNSEL → DATA PRIVACY EXPERT:
May we use consumer AI tools on redacted facts only?

DATA PRIVACY EXPERT WITNESS:
Safer, not safe. Re-identification and residual privilege issues remain. Prefer enterprise controls; document the data flow.

MORNINGSTAR::ENGINEER → QA AUTOMATION EXPERT:
Can proposition-level fitness be automated?

QA AUTOMATION EXPERT WITNESS:
Partially—via retrieval of holding summaries and contradiction checks—but human legal judgment remains mandatory. Automate the lint; do not automate the license to practice.

MORNINGSTAR::ARCHITECT → PROPHET:
Who operates the court-side verifier without creating a new single point of failure or bias?

MORNINGSTAR::PROPHET:
Public reporter APIs, open schemas, multi-vendor validators, and auditability. Distrust monopoly oracles.

PHASE 6: CONSULTANT

MORNINGSTAR (to Consultant): Edward. Your perspective.

The court falls into the familiar uneasy silence.

EDWARD CULLEN (to the Judge, privately):
They fear sanctions more than they love truth—which is usable. Bind the process to shame and evidence: citation ledgers, signer liability, privilege routing. The Prophet’s intake validator is the future; AFAP is the bridge. Do not wait for courts to modernize before you stop fabricating case law.

PHASE 7: VOTE

MORNINGSTAR (Judge):
Vote to adopt AFAP as stated, with Debugger’s proposition-fitness emphasis and Privacy routing included. Prophet’s court-intake validator recorded as recommended external advocacy, not a dependency.

Personality	Vote	Rationale
ARCHITECT	YES	Clean separation of draft/verify/sign will age well.
ENGINEER	YES	Tooling path is shippable; bans unverified export not agents.
DEBUGGER	YES	Proposition-fitness + deep-audit requirements accepted.
PROPHET	YES	AFAP is bridge; intake oracle remains the north star.
COUNSEL	YES	Candor, competence, client protection. Non-negotiable.
COMPLIANCE	YES	Retention, logs, jurisdiction matrix.
DOCUMENTATION	YES	Checklists and disclosure templates are controls.

Result: 7-0-0

PHASE 8: RULING

┌─────────────────────────────────────────────────────────────────┐
│ RULING                                                           │
├─────────────────────────────────────────────────────────────────┤
│ Decision: Adopt AFAP — agents may assist, never file of record;  │
│           human attestation; citation ledger; privilege routing; │
│           provenance retention; cite-lint gate.                  │
│ Vote: 7-0-0                                                      │
│ Rationale: Duty of candor and competence cannot be delegated to  │
│ a generative model; verification must scale with generation.     │
│ Risk: Rubber-stamp signers; incomplete digital reporter coverage │
│ Dissent: None (Prophet’s oracle noted as future advocacy).       │
└─────────────────────────────────────────────────────────────────┘

MORNINGSTAR (Judge):
The court has ruled. “The model said so” is not a defense in this courtroom—or any other worth the name. Precedent tags: #law #professional-responsibility #agentic-controls #documentation #data-privacy.

DR. ECHO SAGESEEKER (Live Commentary):
📘 Unanimous again. The court refuses to let autocomplete practice law. Maslow: safety of the adjudicative system over self-actualization of the toolchain. 📘

UNCLE RUCKUS (Live Commentary):
⌨️ Verify the cite. Sign your name. Keep the privilege. Revolutionary stuff. ⌨️

Transcript certified by MORNINGSTAR::SCRIBE