NORAEarly Access

Part I — Foundations · Chapter 10

An Engineer's Tour of the Federal Rules of Evidence

An Engineer's Tour of the Federal Rules of Evidence

This chapter is a working frame, not legal advice. The legal landscape moves; verify the current state before you cite anything from this chapter in a filing. The dossier in research/05_fre707_and_ai_evidence_law.md carries citations and > links to primary sources. You are designing a system whose outputs may be offered into evidence. You will not be the lawyer who offers them, and you will not be the judge who rules. But the design choices you make in the next twenty- five chapters either make the lawyer's job possible or make it impossible. The lawyer is operating against a fixed set of doctrinal questions; you should know what those questions are before you build the system that answers them. There are five questions. This chapter takes each in turn and shows where the answer lives in a Canon attestation. The running case anchors each question: an attorney representing a parent in a termination of parental rights proceeding (Harrow County, case 2024JC000099) has fifteen minutes to locate ten specific text messages. The evidence is scattered across an iPhone backup, a Gmail account, and a DHS case file. The judge will ask all five questions before any of it comes in. Non-technical readers: Appendix C distills this chapter into a bench-ready reference with Wisconsin-specific statutes, FRE citations, and a chain-of-custody checklist for attorneys. ## At a glance - Five questions admissibility doctrine asks of any digital evidence: is it authentic, is it the best available original, is it hearsay, is it reliable as expert opinion, has it been disclosed? - Each question has a hook in the Canon attestation structure: Witness for authentication, content_ref for best-evidence, claims- and-supports for hearsay, refutation+gaps for reliability, declined-coverage for disclosure. - Proposed FRE 707 specifically addresses machine-generated evidence and is not yet law. Earliest possible operative date: December 1, 2027. ## Learning objectives After reading this chapter you should be able to: 1. Name the five admissibility questions and locate the Canon field that addresses each. 2. Distinguish FRE 901(b)(4) authentication from FRE 901(b)(9) process-or-system authentication, and explain why probabilistic AI outputs fit the latter poorly. 3. Explain why the best-evidence rule is satisfied by a content-hashed content_ref rather than by producing the paper original. 4. Identify when an AI-generated claim faces a hearsay challenge and what the Canon gaps array must disclose about that challenge. 5. Articulate the Daubert factors, map each to a Canon field, and describe what a proponent must still supply that the attestation does not. 6. Explain why FRE 502(b) is relevant to a lawyer reviewing AI-generated evidence searches and what inadvertent disclosure of privilege looks like in a personal-data corpus. ## Why an engineer should read this chapter This is the only chapter in the book where the lawyer's frame is the primary frame. If you are a Builder, do not skim it: the design decisions in Part III only make sense once you have internalized what the doctrine asks. If you are a Receiver, read it for the mapping to Canon fields, not for the law. ## Question 1 — Authentication (FRE 901) > § For the Record — FRE 901(a) and 902(13). > > "To satisfy the requirement of authenticating or identifying an item > of evidence, the proponent must produce evidence sufficient to support > a finding that the item is what the proponent claims it is." > — FRE 901(a). > > "A record generated by an electronic process or system that produces > an accurate result, as shown by a certification of a qualified person > that complies with the certification requirements of Rule 902(11) or > (12) or with a statute allowing certification." > — FRE 902(13). Authentication is the threshold question: is this the thing it purports to be? Not "is it true," not "is it admissible." Just: when you offer this email as a 14 March 2024 message from a DHS caseworker to your client, can a reasonable jury conclude that's what it is? The standard is low. Sufficient to support a finding means a jury could rationally find this is what the proponent says it is. Authentication does not require proof beyond a reasonable doubt; it requires only conditional relevance under FRE 104(b). FRE 901(b) provides ten illustrative authentication methods. Two matter for evidence systems: FRE 901(b)(4) — Distinctive characteristics. Internal indicators — content, substance, internal patterns — that, taken with the circumstances, identify the item. A text message can be authenticated by reference to facts only the purported sender would know. An email can be authenticated by its consistency with prior correspondence in the same thread. FRE 901(b)(9) — Process or system. "Evidence describing a process or system and showing that it produces an accurate result." This is the rule that controls the authentication of machine-generated evidence — security camera footage, GPS logs, breathalyzer readings, and now AI outputs. Rule 902(13) permits self-authentication of electronic records by qualified-person certification, removing the need to call a live custodian at trial. Where 901(b)(9) requires foundation testimony to establish the system produced accurate results, 902(13) allows that showing to happen via pre-trial certification. For a Canon-emitting system, a pre-trial declaration by the system's operator attaching the conformance test results serves as that certification. > ◆ Going Deeper — Why 901(b)(9) fits poorly with probabilistic AI. > > The 901(b)(9) standard was designed for closed, deterministic > systems. A breathalyzer is calibrated; you can show it was > calibrated; you can show its operational range; you can show the > specific reading came from the calibrated device. AI systems do not > fit this paradigm cleanly. > > They are probabilistic; their outputs vary with implementation > details (batch size, hardware, BF16 vs FP32; see Chapter 12 on > determinism); their reliability depends on whether the input is > in-distribution. The 2025 Georgetown Law Technology Review article > "AI Is Coming, But the Rules Aren't Ready" makes this argument at > length. > > Professor Rebecca Delfino's April 2025 submission to the Advisory > Committee on Evidence Rules proposes adding a new Rule 901(c) > specifically for AI-generated media, on the theory that a generative > model does not "produce an accurate result" in the traditional > sense — it produces statistically plausible outputs. As of May > 2026 the proposal has not been adopted. It is on the agenda. > > For a Canon-emitting system, the mitigation is the Refutation > block's replay challenge. A replay challenge re-runs the same > extraction pipeline against the same hashed bytes. If the output > is deterministic (an iMessage extractor reading a SQLite dump is > deterministic), the replay challenge demonstrates a known, > quantifiable error rate: zero variance across replays, given the > same input. This is not a general solution for probabilistic > inference, but it covers the authentication layer precisely. For our purposes, what matters is the mechanism by which authentication is supported in a Canon-emitting system. Each authentication factor a court asks about has a specific field in the attestation the proponent can point to: | 901 element | Canon hook | |---|---| | Identity of the system that produced the output | The findings.method string ("imessage_extractor.py v0.4.2"). | | Validation that the system produces accurate results | The Refutation block's challenges and outcomes; the Admissibility Auditor (Chapter 26) summarizes these. | | Evidence that the version used is the validated version | The method string is version-pinned; the test suite is hashed and committed. | | Evidence the specific output was unaltered post-generation | The Seal block's signature over the canonical chain hash. | | Chain of custody for input data | The Witness block's content hash and the custody_log table's hash-chained entries (Chapter 24). | When an exhibit comes in as a Canon attestation, the proponent does not need to call a custodian to lay foundation for each field. The fields lay themselves: the Seal authenticates the issuer, the chain hash authenticates the content, the custody chain authenticates the handling. What the proponent must still do is show that the issuer's process — the system that produced the artifact — produces accurate results. Chapter 26's Admissibility Auditor is the mechanized form of that showing. ### Wisconsin parallel (§ 909.01, § 909.015(8)) Wisconsin's authentication rules are functionally identical to the federal rules. The "process or system" illustration appears at § 909.015(8) and is the hook for AI-generated evidence in Wisconsin proceedings. The same doctrinal tensions about probabilistic systems apply. There is no Wisconsin Supreme Court or Court of Appeals opinion on AI-generated evidence as of May 2026; the landscape is malleable both for and against admission. ## Question 2 — Best evidence (FRE 1002, FRE 1003) > § For the Record — FRE 1002. > > "An original writing, recording, or photograph is required in order > to prove its content unless these rules or a federal statute provide > otherwise." > — FRE 1002. The best-evidence rule is misnamed. It does not require the best version of a document; it requires the original (or, under 1003, a duplicate that is not unfairly questioned). For digital evidence, this is largely a question of integrity. A Canon Witness entry stores either the bytes inline or a content_ref URL pointing to the bytes; either way, the bytes are hashed and the hash is signed. A recipient who fetches the bytes and recomputes the hash either matches — in which case this is the original — or doesn't, in which case the integrity has been broken. FRE 1001(c) explicitly recognizes that "original" includes electronic data and any printout that accurately reflects it. The rule is friendlier to digital systems than its name suggests; what it asks is demonstrable integrity, not paper. > ✻ Try This. Open the Chapter 1 worked attestation. Find the > content_hash field. The hash is computed over the original bytes > of the source message. If you are reading the attestation as > opposing counsel, what would you ask of the proponent to verify the > bytes have not been altered? (Answer: retrieve the bytes via > content_ref, recompute SHA-256, compare. The proponent does not > have to be present; the issuer's signature does not need to be > renewed; the test is byte-for-byte deterministic.) Wisconsin parallel: § 910.02 (best evidence rule) and § 910.01 (definitions, including electronic data). ## Question 3 — Hearsay (FRE 801–807) > § For the Record — FRE 801(c) and 803(6). > > "'Hearsay' means a statement that: (1) the declarant does not make > while testifying at the current trial or hearing; and (2) a party > offers in evidence to prove the truth of the matter asserted in the > statement." > — FRE 801(c). The hearsay rule is the doctrine engineers most consistently misunderstand. Three points to fix that: A statement is by a person. Machine-generated outputs — security camera footage, ATM logs, breathalyzer readings — are not hearsay because they are not "statements" in the technical sense (FRE 801(a)– (b)). They may face authentication challenges (Question 1) but not hearsay challenges per se. This treatment is being stress-tested by AI. When an AI output incorporates training data drawn from human-authored documents, courts are increasingly skeptical of the "not a statement" analysis. An AI summary of a document may, depending on how the court conceives of the AI's role, be treated as either a machine output (no hearsay problem) or a derivative of the underlying document (in which case the underlying document's hearsay status passes through). The doctrine is unsettled. > ◆ Going Deeper — A worked hearsay analysis. > > An EnrichmentAttestation's Findings block contains a Claim: > "sender's tone is formal and urgent." This is an induction- > typed claim, generalized by an LLM across multiple emails. Is it > hearsay? > > Under the conventional rule: no, because no person uttered it as a > statement. The LLM is a process, not a declarant. > > Under a strict reading: arguably yes, because the LLM was trained > on human-authored text, and its tone-classification output reflects > patterns extracted from declarants whose statements are not before > the court. The induction is the LLM's; the training is theirs. > > The right answer is currently being worked out in trial-court > opinions. The Canon discipline does not resolve the question; it > exposes it. The inference_type: induction and the gap entry > "subjectivity of tone assessment; potential for cultural nuance" > are exactly the disclosures opposing counsel needs to argue the > hearsay analysis. The proponent cannot quietly pretend the > induction is just a fact. The rule also supplies the business-records exception in FRE 803(6): > "A record of an act, event, condition, opinion, or diagnosis if: > (A) the record was made at or near the time by — or from information > transmitted by — someone with knowledge; (B) the record was kept in > the course of a regularly conducted activity of a business, > organization, occupation, or calling, whether or not for profit; > (C) making the record was a regular practice of that activity; (D) > all these conditions are shown by the testimony of the custodian or > another qualified witness, or by a certification that complies with > Rule 902(11) or (12) or with a statute permitting certification; and > (E) the opponent does not show that the source of information or > the method or circumstances of preparation indicate a lack of > trustworthiness." > — FRE 803(6). The 803(6) business-records exception is the workhorse for most institutional records — DHS case notes, CPS visit logs, medical records, school records. It is broad and applies to most records kept in the ordinary course of an organization's operations. Two limits matter for Canon use: First, records made for litigation are not business records. A summary memo prepared by a party in anticipation of litigation does not qualify under 803(6). An EnrichmentAttestation produced during normal evidence ingestion may qualify; one produced for a particular pleading may not. Ingestion discipline matters. Second, authenticity is a gateway. The 803(6) exception requires a custodian or qualified-witness certification. A Canon Witness block supports both: the system operator can certify at deposition or via FRE 902(11) affidavit that the records were produced in the ordinary course of the ingestion system's operation. For Canon's purposes: the Witness block contains the underlying record (the email, the text message). Its hearsay treatment is a function of the record, not of the attestation. The Findings block contains claims about the record produced by the issuer's system. Those claims may face hearsay-adjacent challenges; the gaps array (R5) is where the issuer surfaces the analytical decisions that produced the claim. Wisconsin parallel: § 908.03(6) is functionally identical. ## Question 4 — Reliability of expert and machine output (FRE 702 and proposed 707) ### FRE 702 > § For the Record — FRE 702 (amended Dec. 1, 2023) and proposed FRE 707. > > "A witness who is qualified as an expert by knowledge, skill, > experience, training, or education may testify in the form of an > opinion or otherwise if: (a) the expert's scientific, technical, or > other specialized knowledge will help the trier of fact to understand > the evidence or to determine a fact in issue; (b) the testimony is > based on sufficient facts or data; (c) the testimony is the product > of reliable principles and methods; and (d) the expert has reliably > applied the principles and methods to the facts of the case." > — FRE 702. > > "When machine-generated evidence is offered without an expert witness > and would be subject to Rule 702 if testified to by a witness, the > court may admit the evidence only if it satisfies the requirements of > Rule 702(a)–(d). This rule does not apply to the output of simple > scientific instruments." > — Proposed FRE 707 (published for public comment Aug. 16, 2025; > not yet in effect). This is the Daubert gate. Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993) and Kumho Tire Co. v. Carmichael, 526 U.S. 137 (1999) construed FRE 702 as imposing a gatekeeping obligation on the trial court: the proponent must show, by a preponderance of the evidence, that the expert's methodology is reliable. The Daubert factors are non-exhaustive but conventionally listed: 1. Has the methodology been tested? 2. Has it been peer-reviewed and published? 3. What is the known or potential error rate? 4. Are there standards controlling the technique's operation? 5. Has it gained general acceptance in the relevant scientific community? ML-based expert testimony is being evaluated under these factors with uneven results. The post-Loomis (Wisconsin 2016) case law on algorithmic risk assessments has settled into a four-question version of the Daubert factors, framed for ML systems: 1. Was the system independently tested by a party other than the developer? 2. Is there a known, quantified error rate? 3. Is the error rate acceptable given the stakes? 4. Was the method applied reliably to the specific facts of this case? A Canon-emitting system has direct hooks for each of these: | Daubert/post-Loomis factor | Canon hook | |---|---| | Independent testing | The standalone nora-canon-verifier runs against any emitted attestation; the conformance test suite is publicly published. | | Quantified error rate | The Refutation block's replay challenge variance and the model-eval results referenced in findings.method provide quantified metrics. | | Acceptable error rate | The recipient evaluates this against the substantive context — Step 7 of the verification protocol. | | Reliable application to specific facts | Per-claim inference_type, supports, and gaps arrays document this for every claim. | The Admissibility Auditor (Chapter 26) walks each of these factor-by-factor and emits the result as a Canon-conformant audit attestation. This is not an accident of the design. ### Proposed FRE 707 The Advisory Committee on Evidence Rules has proposed a new rule specifically for machine-generated evidence offered without an expert sponsor (text quoted in the sidebar above). As of May 2026 the rule is not in effect and will not be effective before December 1, 2027 at the earliest. The Committee Note targets AI outputs that "draw inferences and make predictions," distinguishing them from passive recording instruments. > ☉ In the Wild — The Mata → Park → Johnson v. Dunn line. > > Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). Six > fabricated ChatGPT citations in a brief. $5,000 sanction under > FRCP 11. The opening of the line. > > Park v. Kim, No. 22-2057 (2d Cir. Jan. 30, 2024). Counsel cited > a hallucinated case in a reply brief. The Second Circuit held that > FRCP 11 — without any new rule — already covers AI-hallucinated > citations. Established that the verification duty is a baseline > duty of counsel. > > Kohls v. Ellison, 2025 WL 66514 (D. Minn. Jan. 10, 2025). Expert > declaration with fabricated cites struck. Counsel held to have > "personal, nondelegable responsibility" to verify AI-assisted > expert materials. > > Concord Music Group v. Anthropic, 2025 WL 1482734 (N.D. Cal. May > 23, 2025). Single hallucinated paragraph in an expert declaration > "infected the credibility of the entire declaration." The most > aggressive credibility-contamination ruling to date. > > Johnson v. Dunn, 792 F. Supp. 3d 1241 (N.D. Ala. July 23, 2025). > Three Butler Snow attorneys submitted ChatGPT-hallucinated > citations in two routine motions. The court declined monetary > sanctions in favor of public reprimand, disqualification from the > matter, state-bar referral, and mandatory client/court > notification. "Modest fines have proven insufficient to deter AI > misuse." > > The line is two years old and accelerating. By the time FRE 707 > reaches an effective date, the precedent for verification as > personal-nondelegable-duty will be settled. A Canon-emitting > system, by structure, helps counsel discharge that duty: the > attestation is the verification record. Two practical observations: 1. Until FRE 707 takes effect, FRE 702/Daubert applies anyway if the proponent offers AI-generated output as a substitute for expert opinion. The Daubert framework is already the standard. 2. FRE 707, if and when it takes effect, formalizes the existing best practice. A system designed for Daubert-conformant operation today will be FRE 707-conformant on the day it becomes law. Wisconsin parallel: § 907.02 was amended by 2011 Wis. Act 2 to incorporate the Daubert standard. Wisconsin is a Daubert state. Some Wisconsin practitioners encounter judges who reflexively cite the pre-2011 "general relevancy" standard; the amendment is now well-established but is not always applied with federal-court rigor. ## Question 5 — Disclosure and privilege (FRCP 26, ABA 512, FRE 502(b)) A fifth question that bears on whether evidence is allowed in — procedural rather than strictly admissibility, but equally consequential: - FRCP 26: pre-trial discovery obligations. - FRCP 11: certifications attached to filings; the rule the Mata–Park–Kohls–Concord–Johnson v. Dunn line enforces. - ABA Formal Opinion 512 (July 2024): lawyers using GenAI must verify outputs, supervise tools, assess confidentiality, and disclose AI use as material to the attorney's competence under Model Rule 1.1. - Sedona Conference, "Navigating AI in the Judiciary" (Feb 2025): "judicial authority vests solely in judicial officers, not in AI systems"; human verification of all AI outputs. FRE 502(b) provides a privilege clawback rule when disclosure is inadvertent, the holder took reasonable steps to prevent it, and promptly moved to rectify it. The operative text: "the disclosure does not operate as a waiver in a Federal or State proceeding if: (1) the disclosure is inadvertent; (2) the holder of the privilege or protection took reasonable steps to prevent disclosure; and (3) the holder promptly took reasonable steps to rectify the error." FRE 502(b) matters here because a personal-data corpus — particularly one drawn from Gmail and iMessage — will contain privileged communications alongside non-privileged ones. An AI ingestion system that blindly indexes everything will produce search results surfacing attorney-client communications, work product, and spousal communications. If those results are produced to opposing counsel in discovery — even inadvertently — privilege clawback under 502(b) requires showing that the producing party took reasonable steps to prevent the disclosure. "Reasonable steps" is the operative phrase. For a Canon-emitting system, the declined inventory in the Refutation block provides machine-readable evidence of what the system did not do. A privilege_screen challenge declined for not_implemented is a candid admission that the system did not screen for privilege before producing results. That admission both enables clawback (the party was on notice of the gap) and creates liability exposure (the party knew and didn't fix it). Privilege screening is a coverage_audit challenge type in the Canon taxonomy (Chapter 19). The implementation is in workers/jobs/privilege_filter.py (Chapter 22). This is an area where the engineering choice directly maps to a procedural protection. For a Canon-emitting system, the disclosure question is partly answered by the artifact: an attestation that lists its findings.method, its applied and declined challenges, and its gaps is a more disclosable artifact than an opaque LLM output. But it is not a substitute for the lawyer's verification duty. The verifier, the Admissibility Auditor (Chapter 26), and the conformance test suite are tools; they do not displace human review. If you are the subject of a proceeding in which a Canon attestation is offered against you, the disclosure rules above are working for you. Counsel for the proponent has a personal, nondelegable duty to verify the artifact. ABA Formal Op. 512 requires disclosure of AI use. The audit attestation (Chapter 26) is the structured form of that disclosure. Your counsel can ask for it and, if it is not produced, point to the gap. ## The Canon four-stage chain against the five questions Read the Canon blocks against the admissibility questions: | Question | Block | |---|---| | Authentication (901, 902) | Witness — content hash + custody chain. | | Best evidence (1002) | Witness — content_ref resolves to the original; the hash detects alteration. | | Hearsay (801, 803(6)) | The underlying record's status carries through; the Findings block is the issuer's analysis, distinct from the record. | | Reliability (702 / proposed 707) | Refutation — challenges applied + declined, with reasons. Gaps array on every Claim. | | Disclosure (FRCP 26, 502(b), ABA 512) | The whole artifact is structured for production. Every field has a documented meaning; the declined inventory shows what was not tested. | The spec is an engineering specification, not legal scholarship. Every requirement (R1–R9) maps to a place admissibility doctrine asks the proponent to do work. When an R-numbered requirement seems strict, check the case law: a court has almost certainly demanded exactly that. ## Working example — The rescheduled visit The parent's attorney in case 2024JC000099 needs to prove that a DHS caseworker sent a text message on March 14, 2024 rescheduling a supervised visit. The message is in an iPhone backup. Here is the admissibility analysis, question by question: Authentication. The Witness block holds the iMessage extracted from the chat.db SQLite dump: observation_id: obs-01JABC, source: imessage://chat.db/handle:+17155559876/..., content_hash: sha256:f4c9...a2b3. The Findings claim is a deduction typed claim: "Message from +17155559876 on 2024-03-14 rescheduled the supervised visit from March 14 to March 21." The supports field points to obs-01JABC. Authentication under 901(b)(9): the replay challenge confirms the extractor produces an identical output from the same bytes. The findings.method string is version-pinned. Best evidence. The content_ref resolves to the original SQLite dump, hashed and committed. A party who doubts the extract fetches the dump and reruns the extractor. The hash detects any alteration. Hearsay. The text message is an out-of-court statement by the caseworker. If offered to prove the visit was actually rescheduled (truth of the matter asserted), it is hearsay. The attorney must argue an exception — likely FRE 801(d)(2)(D) (party-opponent admission, if DHS is the opposing party) or, if DHS's records are at issue, FRE 803(6) applied to DHS's own case records. The attestation does not resolve this; it exposes the underlying record with the inference_type: deduction flag and a gap noting "sender identity verified by phone number only; hearsay treatment depends on exception." That gap is the proponent's roadmap. Reliability. The replay challenge demonstrates deterministic output. The adversarial prompt challenge ("could this message have been about something other than rescheduling?") returned survived after Tri-Model Consensus. The counter_evidence challenge was declined with reason no_negation_search_implemented_for_imessage_extractor; the proponent discloses this gap rather than hiding it. Disclosure. The full attestation is a discovery-production-ready artifact. Every claim, every gap, every challenge, every decline is in the JSON. Opposing counsel can run the seven-step verifier against it without the proponent's assistance. ABA 512 is satisfied by the disclosed method. FRE 502(b) privilege screening: the system ran a coverage_audit challenge that reviewed all messages in the conversation thread for attorney-client markers before producing this result. The coverage block records it.

💡Key Takeaways
- FRE 901 authentication and FRE 902 self-authentication address different burdens: 901(b)(9) requires process-or-system foundation testimony showing accurate results, while 902(13) permits pre-trial certification — a Canon Attestation supports both via the Refutation block's replay challenge and a declarant-signed conformance report. - The best-evidence rule (FRE 1002) is satisfied for digital records by a content-hashed content_ref that resolves to the original bytes — any recipient can re-hash and compare without requiring a paper original. - The FRE 803(6) business-records exception is the workhorse for DHS and agency records but fails for records made in anticipation of litigation; an EnrichmentAttestation produced during routine evidence ingestion may qualify, while one produced for a specific pleading likely does not. - Judges applying Daubert (and Wisconsin's § 907.02 equivalent) ask whether the methodology was independently tested, has a quantified error rate, and was reliably applied to this case — each of those questions maps to a specific Canon field (Refutation block, replay-challenge variance, per-claim supports and gaps). - Every forensic exhibit must answer two questions before admission: "is this what it purports to be?" (authentication — Canon Witness block) and "is the method that produced it reliable?" (reliability — Canon Refutation block); a Canon Attestation structures both answers for the judge.
## Exercises ### Warm-up 1. For each of the five admissibility questions, find the field in the Chapter 1 worked attestation where the proponent would point to satisfy that question. 2. Read the text of FRE 901, 902, 1002, 803(6), 702, and 502(b). (Cornell LII.) Note that each fits on a single screen. Note which ones were amended after 2010 and which language drove the amendment. ### Core 3. Read Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023). Identify the specific lapses that produced the sanction. For each, name a Canon mechanism that, had it been in place, would have prevented it. 4. Wisconsin's § 909.015(8) is functionally identical to FRE 901(b)(9). What is the practical consequence for a Canon-emitting system used in Wisconsin proceedings? Which design decisions would change? 5. In the working example above, the caseworker's text message is hearsay. Identify which exception — FRE 801(d)(2)(D), FRE 803(6), FRE 803(1), or FRE 803(3) — is most likely to apply, and explain what the proponent must show for each. For each exception, note whether a field in the Canon attestation supports that showing or whether the attorney must supply it independently. ### Stretch 6. Suppose proposed FRE 707 takes effect on its earliest possible date. What changes for a Canon-emitting system that operates only against personal-data corpora the issuer custodies — i.e., is the issuer's own AI output offered through the issuer (a fact witness) or offered as machine evidence (subject to FRE 707)? Where does the line fall? 7. The declined inventory in the Refutation block shows that the privilege screen was not implemented. Under FRE 502(b), does that fact help or hurt the proponent's clawback argument if a privileged document is inadvertently produced? Construct both arguments. 8. Outline the contents of a one-page "AI use disclosure" you would attach to a filing produced by a Canon-emitting system. What does the attestation give you for free? What do you have to add by hand? ## Build-your-own prompt For your capstone corpus: identify which of the five admissibility questions is most likely to be contested if your output goes into a proceeding. Sketch one paragraph of the disclosure you would prepare in advance — what the attestation gives you automatically, what you must add by human review, and what the declined inventory reveals about the limits of your current implementation. This paragraph will become the "Admissibility Profile" section of the AuditAttestation your system emits at Chapter 26. ## Further reading - Appendix C of this book — Wisconsin & federal evidence law primer — is the bench reference for Canon attestations in Wisconsin circuit court proceedings. Non-technical readers (judges, lawyers, paralegals) should read Appendix C as a companion to this chapter before any hearing. - The text of FRE 901, 902, 1001–1003, 801–807, 702, 703, 502 (Cornell LII). - Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023) (Castel, J.). Read the opinion. - Park v. Kim, No. 22-2057 (2d Cir. Jan. 30, 2024). - Johnson v. Dunn, 792 F. Supp. 3d 1241 (N.D. Ala. July 23, 2025). - Ferlito v. Harbor Freight Tools USA, Inc., No. CV 20-5615, 2025 WL 1181699 (E.D.N.Y. Apr. 23, 2025). - ABA Formal Opinion 512 (July 29, 2024). - The Sedona Conference, "Navigating AI in the Judiciary" (February 2025). - The full citation list in research/05_fre707_and_ai_evidence_law.md.

  • Wisconsin: 2011 Wis. Act 2; § 907.02 (Daubert); ch. 909 (authentication); § 968.31 (one-party consent); ch. 48 (child welfare); §§ 19.31–19.39 (open records); § 908.03(6) (business records); § 910.02 (best evidence).
  • State v. Loomis, 881 N.W.2d 749 (Wis. 2016) (algorithmic risk assessment under Daubert; applied to ML systems in Wisconsin courts).

Next: Chapter 4 — The Canon Standard.