Field Guide · framework
How AI Constitutions Prevent the Sullivan & Cromwell Failure Mode
On April 9, 2026, a top-ranked attorney at Sullivan & Cromwell filed an emergency motion in a Southern District of New York bankruptcy proceeding that contained more than forty AI-fabricated citations. The hallucination is not the news; the absence of a constitution that would have refused to ship the filing is.
On April 9, 2026, Sullivan & Cromwell — among the most prestigious law firms in the United States — filed an emergency motion in a Chapter 15 bankruptcy proceeding before Chief Judge Martin Glenn of the US Bankruptcy Court for the Southern District of New York, involving Prince Global Holdings Limited. The filing contained more than forty errors, including fabricated case citations generated by AI. A top-ranked Sullivan & Cromwell partner subsequently filed an emergency letter to the court explaining the failure and asking the court not to sanction the firm.
The Damien Charlotin AI Hallucination Cases Database catalogued 1,348 worldwide cases of suspected AI hallucination in court filings as of April 24, 2026 — 915 of them from US courts. Reported incidents have grown from roughly two per week in early 2025 to two to three per day by late 2025. US courts imposed more than $145,000 in AI hallucination sanctions in the first quarter of 2026 alone.
The point of this essay is not to argue that AI hallucinates. The point is to argue that the failure mode is structural — and so is the remedy.
Why it matters
The Sullivan & Cromwell incident reads to most people as a story about AI being unreliable. It is not. It is a story about the absence of structure that would have caught the unreliability before it shipped.
The pattern is well documented. An attorney uses a general-purpose large language model to draft a brief. The model produces case citations. Some citations are real, some are fabricated. The model has no built-in mechanism that distinguishes between the two. The attorney reviews the draft for argument and style, signs it, and files it. The judge or opposing counsel catches the fabricated citations. Sanctions follow.
The remedy is not "tell people not to do that." The remedy is not a disclaimer at the bottom of the output. The remedy is structural: the agent that drafts the brief must be constitutionally incapable of shipping a brief without citation verification. The verification cannot be optional. It cannot be a heuristic the agent runs when it remembers to. It has to be a hard-coded gate the agent cannot override.
This is what we mean when we say an agent has a constitution.
A constitution, in the Fidelic sense, is a written set of rules the agent cannot override. It is not guidance to the model. It is not a system prompt that the model interprets as a suggestion. It is a hard-coded gate in the agent's tool-call layer that determines whether the agent can ship its output.
For a legal-research agent, the constitution names what the agent will and will not do. The agent will draft research memos for a supervising attorney. The agent will not give legal advice. The agent will not finalize redlines. The agent will not invent controls. Every citation in the agent's output is verified against an authoritative source — Westlaw, CourtListener, the case's actual docket — before output is allowed to ship. Unverifiable citations trigger an output block, not a flag-and-ship warning.
The verification step matters. A heuristic that says "check the citations" produces drafts that include unverified citations because the heuristic is interpreted, not enforced. A hard-coded gate produces drafts that do not contain unverifiable citations because the agent could not ship them in the first place. This is the entire structural difference between an AI tool that is a liability magnet and an AI agent that is an employee.
The supervising attorney still owns the judgment. The agent is not the lawyer. The agent does the daily preparation — the research, the structured first drafts, the citation lookup, the contract redlining at the playbook level — and stops at the work that requires human judgment. The Sullivan & Cromwell partner is still the partner; the agent is the senior associate that would have refused to ship the brief without first verifying every citation it contained.
Two questions follow. The first: is citation verification reliable enough to be a hard-coded gate? The answer is conditional. Citation verification against an authoritative source set has a known false-negative rate — sometimes the citation is real but the verifier can't find it. The way to handle this is human review of flagged citations, not human review of every citation. The supervising attorney looks at the agent's flagged-cannot-verify list and resolves those manually. This reduces attorney review time, not just the agent's hallucination rate.
The second question: can you actually build this? Yes. Fidelic's pre-launch legal cohort — VRAX for research, PRAX for contract redlining, ETHA for security questionnaires, VELA for compliance — has citation verification as a hard-coded gate. The verifier looks up case names, parties, and citations against an authoritative source set. Output is blocked when citations cannot be authoritatively verified, not flagged-and-shipped. The Roster legal cohort opens to the public in Q2 2026.
The edge
The Sullivan & Cromwell partner who signed the brief is not a junior attorney. The firm is not a junior firm. The model that generated the citations is not a primitive tool. The failure mode does not require any of those things to be true. It requires only one thing: that the agent that drafted the brief had no constitutional discipline preventing it from shipping unverified citations.
Until that changes — and it is starting to change — the failure mode will keep happening at the top of the legal profession. Eugene Volokh documented seventeen US court decisions in a single day, March 31, 2026, noting suspected AI hallucinations in filings. The next one is already being drafted somewhere.
Honest take
Constitutions do not prevent all hallucination. They prevent shipping hallucinated output that violates a hard-coded gate. A constitution that says "the agent will not advise" does not prevent the agent from drafting bad analysis; it prevents the agent from delivering that analysis as advice. A constitution that says "the agent will not finalize redlines" does not prevent typos; it prevents the agent from committing to an unreviewed redline. The supervising attorney still owns the judgment, the strategy, and the final word.
The constitution plus the supervising attorney plus the citation-verification gate together prevent the failure mode Sullivan & Cromwell ran into. The constitution alone does not. The supervising attorney alone does not — that's the Sullivan & Cromwell case in a sentence. The verification gate alone does not. The three together do.
We are also not claiming Fidelic is the only platform working on this. Harvey, the category-leading legal-AI vendor, has published a Trust Layer and has iterated against well-documented hallucination incidents. The Fidelic difference is publication discipline — the per-agent constitution and capabilities-and-safeguards list are published before purchase, not learned in deployment.
The Sullivan & Cromwell incident is not a story about AI being unreliable. It is a story about the absence of structure that would have caught the unreliability before it shipped. The failure mode is listed, dated, and well-understood. The remedy is structural. The Roster legal cohort ships Q2 2026.