Purpose: documents ContrarianAI LLC's segregation between analytical and operational tooling, protecting Rule 702 admissibility under adversarial cross-examination.
Author: Kevin T. Luddy, Principal
First published: 2026-08-25
Status: governing protocol for all engagements
No AI touches the measurement. AI participates in the interpretation. I own the opinion and can derive every part of it without reference to any model.
That is the whole protocol in three sentences. The rest of this page is what each one means in practice and how you can test it.
Where a finding rests on statistical inference, that inference uses classical estimators only, each with published provenance predating by decades any application of machine learning to this kind of problem:
Many findings rest on no statistics at all. They rest on what a document says, how a system was built, and what records a system of that design necessarily produces. Those are established by direct examination of the source and by forty years of building the systems in question, and they are reconstructible by anyone who reads the same documents.
This is the standard the work is held to now, in consulting engagements, and it is the standard any Rule 26(a)(2)(B) engagement would be held to. To be exact about the record: ContrarianAI has completed two paid engagements to date, both consulting-only preliminary case assessments, with no testimony given.
No machine learning is applied to case data at any step, and no conclusion stands because a model produced it. Every element of every opinion is reconstructible by the classical methods above, on demand, in deposition or at trial.
Federal Rule 702, as amended in 2023, puts the burden on the proponent to establish reliability by a preponderance of the evidence. In Kohls v. Ellison (D. Minn., Jan. 10, 2025) the court struck an expert declaration whose AI-drafted citations pointed to papers that do not exist. Kohls is a holding about fabricated content and the nondelegable duty to verify a filing. It is not a holding that machine learning inside an expert's analytical chain is unreliable, and this protocol does not claim one exists.
A second decision sharpens what "verify" has to mean. In N.Z. v. Fenix International Limited, No. 8:24-cv-01655-FWS-SSC (C.D. Cal. Dec. 12, 2025) (ECF 207), Judge Fred W. Slaughter imposed Rule 11 sanctions on Hagens Berman Sobol Shapiro LLP and two attorneys over four opposition briefs drafted with ChatGPT: $10,000 jointly and severally against the firm and lead counsel, $3,000 against co-counsel, service of the order on the California and Arizona State Bars, and a standing requirement that every subsequent brief in the case carry a certification that all citations were checked. The court expressly declined to sanction under its inherent authority, which requires bad faith. Rule 11 requires none.
The detail that matters is what the errors actually were. The court sorted them into four categories: hallucinated authority; real authority quoted incorrectly; real authority cited for a proposition it does not support or otherwise mischaracterized; and incorrect record cites, including a mischaracterization of the court's own prior order. Only the first is a fabrication. Lead counsel, already suspecting AI use, ran the reporter citations to confirm the cases existed and matched by name and subject matter. They did. He did not check the parentheticals. Three of the four error categories survived that check.
That is the operative lesson for expert work, and it is why the verification rule in this protocol is written the way it is. Confirming that a citation resolves is not verification. Verification is reading the source and confirming it says what the filing claims it says. An existence check catches one failure mode out of four.
The risk being managed here is narrower and more practical. An opposing expert who can characterize a methodology as machine-generated gets to litigate admissibility instead of merits. That fight is expensive, it happens before anyone reaches the substance, and it is avoidable by not supplying the hook.
Any expert who uses AI to help draft a report, run an analysis, or form an opinion has created a discoverable, exploitable weakness at the Daubert gatekeeping step.
This is the absolute line and it has not moved. Case-confidential material — discovery production, deposition transcripts, protected-order documents, raw log exports — never passes through AI-assisted tooling of any kind, and no machine-learning inference appears anywhere in the pipeline that measures it. Every number is produced by scripted classical estimators running in an isolated environment with an independent audit trail. No model reads the evidence, scores it, classifies it, summarizes it or decides what it means. That is what protects confidentiality, and it is what removes any question of model contamination of the measurement itself.
These produce the numbers, and nothing in them is AI-assisted at any step:
Tools used: scripted, published-provenance classical statistical tools (Python with numpy and scipy; R notebooks), verifiable by any third party. No AI inference is applied to case data at any step, and no figure in any deliverable is generated by a model.
Once the measurement exists, the work of saying what it means is AI-assisted, and I state that plainly rather than leave it to be discovered. A distribution has to be read, explained, checked for alternative readings, and written in language that survives cross-examination. On that work I use AI the way a careful writer uses a demanding editor and a second reader:
Why I do it this way rather than alone: a single analyst misses things, and the cost of a missed alternative explanation in an expert report is borne by the client. A second reader that never tires improves completeness and precision. I judge that to produce better work than I produce unaided, and I am not willing to claim otherwise for marketing reasons.
The constraints are firm. Nothing reaches a deliverable that I have not independently verified and agreed with. No conclusion stands because a model proposed it. Every element of every opinion is one I can derive, explain and defend without reference to any model, and if I cannot, it does not go in. What is discussed is the shape of a result and the reasoning about it, described so that the matter cannot be reconstructed from it — not the party, not the produced documents, and not the combination of specifics that would identify either. The test is not whether a name was used. A sample size, a date range, a metric and a domain travelling together will identify a defendant with every proper noun removed; the kind of system is enough to reason about a distribution, and the kind of system is what is given.
Two controls sit behind that. These exchanges run in a configuration where inputs are not retained or used for training. And every prompt is written on the assumption that it will be read back to me in a deposition — Conservation Law Foundation v. Shell is the reason, and the AI-use log exists so that the assumption can be tested rather than asserted.
Named explicitly because v1.0 of this protocol did not name it, and silence there invited a reading of total abstinence that I do not practice.
AI is a capable research instrument and I use it as one, on general technical questions, never on case data:
What it produces is a proposal, not a finding. Each proposal is tested against the evidence and against forty years of building the systems in question. Sound proposals are then established independently by the classical means above. Unsound ones are discarded, and discarding them is frequently the more valuable step. The process is iterative by design — proposed, tested, refined, tested again — until what remains is something I can derive and defend myself.
Two rules govern it, and they are written rather than improvised:
This use is logged and disclosed. I expect the treatment ordered in Conservation Law Foundation v. Shell Oil Co., No. 3:21-cv-00933 (D. Conn. May 18, 2026), where Magistrate Judge Farrish held that "an expert witness's methodology is fair ground for discovery" and ordered production of an expert's generative-AI prompts, finding generic protective stipulations insufficient to shield them. The exposure in that matter was delegation — document filtering had been handed to the tool and its output went into the work unexamined. Mine does not go unexamined, and the log is how that is demonstrated rather than asserted.
The following administrative and business activities are AI-assisted. That use is substantial and I do not minimize it; the discipline here is segregation and disclosure, not abstinence:
Every AI-drafted email or artifact that ships to a case-adjacent party is reviewed and approved by Kevin personally before sending.
Some business-side publication is automated and runs without per-item review — the nightly SEC EDGAR AI-risk digest on this site is generated, committed and published on a schedule. It is disclosed here rather than left to be discovered: none of it is case-analytical, none of it touches case data, and none of it ships to a case-adjacent party.
Any retaining party may request:
Applied to every citation in every deliverable, and reusable by anyone. The categories are the
court's own, from N.Z. v. Fenix (C.D. Cal. Dec. 12, 2025); the checks are what each one requires.
| # | Failure | What catches it | What does NOT catch it |
|---|---|---|---|
| 1 | Authority does not exist | Reporter lookup; does the case resolve | Nothing weaker |
| 2 | Authority exists, quotation or parenthetical is wrong | Open the source, match the quoted words character for character | Confirming the case is real |
| 3 | Authority exists, does not support the proposition | Read the holding and confirm it reaches this issue | Name and subject-matter match |
| 4 | Record cite is wrong, including mischaracterizing a prior order | Open the docket entry and read the paragraph cited | Any check performed outside the record |
Categories 2 through 4 involve real authority. Every one of them survives a check that only asks
whether the citation resolves. In Fenix, that is precisely the check that was run, and precisely
why it failed.
Standing rule: any claim that cannot be verified against a primary source is marked and held
out of the deliverable. Not softened. Held.
Standard cross-exam question-and-answer, kept as governing script. Versions 1.0 and 1.1 scripted these answers around the analysis of case data, where they hold exactly. They are restated below so that they cannot be read past that subject, and so that the use of a research instrument on general technical questions is disclosed in the same breath rather than surfacing later as if withheld.
Q: Do you use AI in your work?
A: Yes, and I will tell you exactly where. It is nowhere in the measurement. No model reads the produced data, computes a figure, or decides what the data shows. It is present in the interpretation — testing my reading of a result against alternatives, checking my argument for gaps, and in the writing itself. It has no authority in my conclusions. Nothing enters an opinion of mine because a model produced it.
Q: So artificial intelligence contributed to your analysis.
A: It contributed to the analysis of what the measurements mean. It contributed nothing to the measurements. Each proposal was tested against the evidence and against forty years of building these systems; what survived I established independently, what did not I discarded. The opinion is mine and I can derive every part of it, here, without reference to any model.
Q: Did the case data go through an AI system?
A: No. Case-confidential material never passes through AI tooling and no machine learning appears in the pipeline that measures it. That line is absolute and the isolated-environment audit trail demonstrates it. What I discuss with a model is the shape of a result and the reasoning about it, put in terms from which this matter could not be reconstructed. Not the party, not the produced documents, and not the combination of details that would identify either. My prompts are logged and you may read every one of them.
Q: Did an AI draft this report?
A: No. I wrote it, and the commit history shows it being written. AI was used on it the way an editor and a second reader are used — checking the argument for gaps, testing a reading against alternatives, and correcting language, structure and precision. Every suggestion was mine to accept or reject, nothing went in that I had not independently verified and agreed with, and no figure in it came from a model.
Q: How do we know your conclusions are not contaminated by a model's error?
A: Because no conclusion rests on a model being correct. The figures come from estimators with published provenance — Fisher, Pearson, Kullback-Leibler, Cohen — through scripted tools you can re-run on the same inputs and obtain the same results. A model that was wrong at the proposal stage produces a discarded proposal, not a downstream error, because nothing reaches the deliverable that I have not independently established. Ask me any element of this opinion and I will derive it for you now.
Q: Produce your prompts.
A: They are logged and I will produce them. They show background reading on general technical questions and the testing of my own reasoning about results. You will not find produced case data in them, you will not find a figure that came out of one, and you will not find a conclusion that rests on one.
Q: Your operational log shows AI-assisted approvals. Explain.
A: Those are administrative operations — email, calendar, scheduling — auto-executed within a hard deny-list, the same as any professional using automation. Nothing analytical, nothing case-touching.
This protocol is:
contrarianai-landing.onrender.com/independence-moat.htmlVersion: 3.0 (2026-10-20)
3.0 — 2026-10-20. A correction of description, not of practice, and the most important one on this page. Versions 1.0 through 2.1 described analytical work as performed without AI at any step. That described work I have not done and would not do. A Rule 26(a)(2)(B) analysis is an enormous amount of work for one person, and I judge that AI participation in the interpretive part of it produces more complete and more precise work than I produce unaided. Publishing a protocol that promised otherwise would have handed an opposing expert the only thing that actually threatens an opinion, which is a demonstrable gap between what a witness says he does and what he does.
The firewall is not abandoned; it is stated where it actually sits. No case data and no measurement step passes through AI tooling — no model reads produced evidence, computes a figure, or decides what the data shows, and every number in a deliverable comes from scripted classical estimators in an isolated environment. The interpretation of those measurements is AI-assisted — testing a reading against alternatives, finding gaps in an argument, and language. Nothing reaches a deliverable unverified and unagreed, no conclusion rests on a model being right, and every element is one I can derive unaided on demand. The cross-examination script, the attestation at Verification item 4, and the segregation tiers are all restated to this line.
2.1 — 2026-10-20. Accuracy corrections. No commitment is changed or withdrawn. Six statements were testable by an opposing expert and three of them would not have held. The operational tier claimed review of every AI-drafted artifact while the nightly EDGAR digest on this same site publishes unreviewed; that automation is now disclosed and bounded. The estimators were described as predating "the AI era by 50-125 years", a span that is arithmetically wrong and rests on an undefined date — on the 1956 Dartmouth reading, Cohen's kappa (1960) postdates it; the dated citations now stand alone. The case-data absolute said "client records", broad enough to be crossed by the business system's own record of who the client is; it now names discovery, transcripts, protected-order documents, log exports and analytical work product, which is the line that matters at the gatekeeping step and the one that holds without qualification. "The same outputs" became "the same results", since bit-identical output is not guaranteed across library versions. The comparison to Word autocomplete was removed as minimization. Review cadence is stated as a schedule rather than as a history, the first quarterly review being still ahead.
Corrected in the same amendment, and the most consequential of the set: the claim opened "Every finding in a ContrarianAI Rule 26(a)(2)(B) engagement is independently established using only classical statistical estimators." Both halves failed. ContrarianAI has not performed a Rule 26(a)(2)(B) engagement, so the sentence described an empty set and would not have survived the question of how many such reports exist. And "only classical statistical estimators" misdescribed the work: much of what I find rests on documents and system architecture rather than on any estimator. The claim is now stated as a standard rather than a history, covers both kinds of finding, and the engagement record is stated plainly.
2.0 — 2026-09-18. Clarified in scope. Nothing in the underlying commitment has changed or is withdrawn.
Versions 1.0 and 1.1 stated "no AI, ever" under the ANALYTICAL heading, whose first listed item is statistical analyses on case data. That was the referent, and as to that referent it was exact then and is exact now: no machine learning is applied to the analysis of logged or produced case data, no case-confidential material passes through AI tooling, and no report, methodology selection or opinion is generated by a model. Every one of those commitments survives this amendment unaltered.
What the earlier wording did not do was distinguish that scope from a separate activity it was never meant to cover: consulting a research instrument on general technical questions — how a particular legacy platform or obsolete format behaved, what lines of inquiry an architecture of a given shape invites — and testing what it proposes against the evidence and against forty years of building such systems. That activity touches no case data and produces no finding; its output is a proposal that is verified against a primary source and independently re-established before anything is relied on, or else discarded.
A reader could have taken the earlier phrasing more broadly than its heading, and a protocol published under my own name should not depend on a reader reading carefully. This version therefore names the BACKGROUND AND ORIENTATION tier explicitly, states the case-data firewall as its own absolute line rather than leaving it implied, replaces the attestation at Verification item 4 with one framed to the actual scope, and restates the cross-examination script so the answers cannot be read past their subject. The practice is unchanged; only its description is now unambiguous.
1.1 — 2026-09-07. Added the N.Z. v. Fenix analysis and the four-category citation check. Published to the website; the markdown footer was not updated to match and continued to read 1.0 until this amendment. That drift is itself the argument for generating the published page from this file rather than maintaining two copies by hand.
1.0 — 2026-08-25. First published.
Contact: kevinluddy39@gmail.com