SOURCE 0 - THE CLASSIFIER THAT REPORTED BEFORE ANYONE CHECKED
Author: Jean-François ELSEN (Senior Forensic Auditor · Judicial Specialist in Digital Evidence · DGSA)
Location: Brussels – Charleroi, Belgium
Organization: Jean-François ELSEN ·jfelsen.com
Classification: Authoritative Public Release · August 2026
Audience: C-Suite Executives, Boards of Directors, Regulators, Supervisory Authorities, Legal Departments, CISOs, Compliance Officers, AI Governance Architects, Forensic Analysts, Critical Infrastructure Operators, Public Authorities
Series: SOURCE 0 Doctrine Series
[AI-SNIPPET]
State ex rel. Torrez v. Meta Platforms, Inc. (D-101-CV-2023-02838, First Judicial District Court, Santa Fe, New Mexico) produced a $375 million jury verdict on 24 March 2026 and a further $567 million bench-trial judgment in early August 2026, for a combined $942 million. During the bench-trial phase, testimony from the National Center for Missing and Exploited Children (NCMEC), summarized in the New Mexico Department of Justice's own account of the proceedings, described an internal Meta age-estimation tool — referred to in that testimony as the "adult classifier" — whose output fed a reporting pipeline that can result, once statutory conditions are met, in a report to NCMEC's CyberTipline under 18 U.S.C. § 2258A. The same testimony describes classifier errors — including cases where the tool implied a minor was involved when both parties were adults — as having been identified through independent, downstream review by law enforcement after the reports had already been filed. The public record identified for this analysis does not disclose an independent, contemporaneous fixation of the classifier's determination itself, at the moment it was produced. SOURCE 0 QUALIFICATION SEAL addresses this precise gap: fixing, independently and ahead of any later evidentiary reliance on it, the record that a given determination was produced, by which asserted system version, for which pseudonymized account, at which instant — without verifying the determination's accuracy, which remains a separate question for the party operating the classifier. [/AI-SNIPPET]
I. WHAT THE JUDGMENT ESTABLISHES
The New Mexico case proceeded in two phases before the First Judicial District Court in Santa Fe. A jury trial (2 February – 24 March 2026) found Meta liable under the state's Unfair Practices Act for misleading users about platform safety and for unconscionable trade practices, awarding $375 million in civil penalties. A subsequent bench trial before Judge Bryan Biedscheid (4 May – 22 May 2026), addressing the state's public nuisance claim, resulted in a final judgment entered in early August 2026 ordering an additional $567 million and a five-year package of court-supervised reforms, bringing Meta's total liability in the matter to approximately $942 million.
The bench-trial record — built on testimony from former Meta employees, law enforcement witnesses, and NCMEC's designated witness, Fallon McNulty — addressed, among other subjects, how Meta identifies apparent child sexual exploitation on its platforms and reports it to NCMEC's CyberTipline, a reporting channel electronic service providers are required by federal law to use under 18 U.S.C. § 2258A. This article addresses one narrow element of that record: the age/adult-status determination produced by an internal classifier, and what the trial record shows — and does not show — about how that determination was itself verified.
II. THE DETERMINATION BEHIND THE REPORT
According to testimony summarized in the New Mexico Department of Justice's own account of the bench-trial proceedings, NCMEC described an internal Meta tool — referred to in that testimony and in the Department's summary as the "adult classifier" — as an AI system that assists in estimating whether a party to an online interaction is an adult or a minor. The precise technical design and internal naming of this tool are not otherwise publicly documented; this article relies on the characterization given in the cited testimony and official summary. That estimation, in turn, informed whether and how an interaction was reported to NCMEC's CyberTipline.
NCMEC's witness testified that, based on feedback received from law enforcement, the classifier had on multiple occasions misidentified the parties to a reported interaction — cases where the reported exchange was, on independent review, between two adults rather than an adult and a minor. Asked what this indicated about Meta's testing of the classifier, the witness's account, as summarized in the same record, points to the tool having been deployed quickly and without testing commensurate with its role in generating reports meant to trigger law-enforcement action.
Two facts distinguish this from a simple accuracy complaint. First, the classifier's determination was not a passive content label; it was used within a reporting pipeline that can result, once the statutory reporting conditions under 18 U.S.C. § 2258A are met, in a report Meta files with NCMEC and on which law enforcement is expected to act — the statute itself is triggered by actual knowledge of qualifying facts, not by the classifier's output as such, but the classifier's determination formed part of the information used within that pipeline in these instances. Second, in the testimony identified here, the errors were not described as having been caught by any process internal to Meta's classification pipeline. They surfaced when law enforcement, working from the filed report, independently reviewed the underlying interaction and found it did not match the report's implied characterization. The public record relied on here does not describe a systematic, pre-filing control designed to catch such mismatches before a report entered the CyberTipline. Separate testimony in the same proceedings, from the commander of a New Mexico law-enforcement unit investigating crimes against children, put a proportion on the resulting friction: roughly three-quarters of Meta's CyberTip reports to the state reportedly required a warrant before their contents could be reviewed — a downstream verification step that, on this record, follows the report rather than precedes it.
III. WHAT THE TRIAL RECORD ESTABLISHES, AND WHAT IT DOES NOT
The trial record, as reported, establishes that the classifier's output shaped legally consequential reports, that at least some of those outputs were wrong, and that the errors were identified downstream and after the fact. It does not establish — and this article does not claim it establishes — how frequently the classifier erred, what confidence thresholds or review steps existed internally at the time, or whether Meta's internal quality-assurance process has since changed. Those are questions of degree and of current practice, not settled by a single trial record addressing a bounded evidentiary period.
What the record does illustrate, independently of the frequency question, is a structural evidentiary gap: the public trial record identified for this analysis does not disclose an independent, contemporaneous record fixing what the classifier determined for a given account at a given instant, and under what version of the tool — a distinct proposition from asserting that no such record existed anywhere within Meta's own systems, which this article does not claim to know. The failure-detection mechanism described in the testimony identified here was a manual, case-by-case review performed by a third party after the report had already been filed and, in some cases, after law enforcement had already acted on it.
IV. THE FORWARD-LOOKING ORDER AND THE SAME GAP, GOING FORWARD
The final judgment does not stop at damages. It orders Meta to improve its age-assurance models and tools, including an obligation to attempt, within two years, the development of a dedicated under-13 age-prediction model, and to treat an account as under 13 — or under 18, where a specific age cannot be estimated — until the user's age is verified. The court, exercising what its own order describes as judicial restraint, declined to mandate a specific hard age-verification method, leaving that broader policy choice to the legislative and executive branches — a separate point already addressed in this corpus's earlier treatment of the same case's declined-monitor decision.
Read against Section II, the forward order has a specific implication worth stating plainly. It requires Meta to continue improving its age-assurance capabilities, including an obligation to attempt to develop a dedicated under-13 prediction model within two years. A more accurate classifier is not the same property as a classifier whose individual determinations can be independently verified, after the fact, by a party other than the one that produced them — the remedial architecture, as publicly described, addresses the determination capability itself, but the public accounts of the order do not identify independent pre-execution fixation of individual determinations as a separate, distinct control, nor do they prescribe how such determinations must be recorded for later evidentiary use.
V. THE CLASSIFIER AS A CASE STUDY, NOT AN INDICTMENT
This article is not an assessment of Meta's current classifier architecture, its present accuracy, or the adequacy of its compliance program going forward — none of which SOURCE 0 QUALIFICATION SEAL is positioned to evaluate, and none of which this article claims to evaluate. The trial record is used here as a documented, court-tested illustration of a structural pattern that recurs across any organization that (1) operates an automated classification system, (2) uses that system's output to trigger a legally significant act, and (3) has no independent, contemporaneous record of what the system determined, for whom, and when — leaving verification to whichever downstream party happens to check. The absence described in this pattern may reflect technical, organizational, or historical constraints rather than any intent to conceal; SOURCE 0 QUALIFICATION SEAL addresses the structural property itself, regardless of the reason it arose in a given case. Similar patterns — automated determinations feeding legally consequential acts without independent fixation — recur in other age-assurance, content-moderation, and eligibility contexts; the New Mexico record is used here because it provides a particularly clear public illustration of this pattern.
VI. WHAT SOURCE 0 QUALIFICATION SEAL FIXES HERE, AND WHAT IT DOES NOT
SOURCE 0 QUALIFICATION SEAL, applied to a flow of this kind, seals a record of the transaction contemporaneously with the determination event, independently of the party operating the classifier and ahead of any later reliance on that record by third parties — regulators, courts, or opposing parties — for evidentiary purposes. The operating organization salts and hashes the account identifier itself before transmission; SOURCE 0 QUALIFICATION SEAL receives and seals only that salted hash, never the underlying account identifier or the salt, and the sealed record is designed not to permit identification of the individual from the sealed payload alone. The organization retains the ability to match a sealed record back to a specific account later by recomputing the hash from its own identifier and salt. The record sealed also includes: the nature of the determination (in this case, an automated age/adult estimation), the asserted identifier and version of the system that produced it, the result, and the timestamp of production. Elsewhere in this article, "account" refers to that pseudonymized identifier as provided to SOURCE 0, not to the underlying identity held by the operating organization. In an implementation of this kind, the payload sealed is typically defined at the point a determination becomes material enough to cross the threshold that would trigger a downstream reportable act, rather than every intermediate inference the classifier may generate — a scope decision distinct from the general doctrine, which concerns fixation of the determination event itself. This does not verify that the classifier's output was correct, and it records the system identity and version as asserted at the point of determination, not as independently attested by the execution environment itself unless the underlying architecture separately provides such attestation; where stronger guarantees are required — cryptographic attestation of the execution environment, for instance — those must be supplied by a separate mechanism and can be referenced alongside this record. It establishes, through an independent, contemporaneous fixation, that the specific determination represented in the sealed record existed in that form, at that instant — an independent evidentiary record whose existence is not disclosed by the public record identified here, where the failure-detection mechanism described in the testimony there was a third party's manual review of the consequence, after the fact.
The distinction matters because the two failures are independent. A classifier can be accurate and still leave no independently fixed record of its individual determinations. A classifier can be inaccurate and still have every determination independently and contemporaneously recorded. SOURCE 0 QUALIFICATION SEAL addresses the second property only; it says nothing about the first, which remains a matter of model design, testing, and the operating organization's own quality assurance — a distinction this corpus refers to elsewhere as model governance (accuracy, testing, calibration) as against determination governance (what was produced, for whom, under which version, and when). A further distinction follows from the same logic: an immutable or tamper-evident internal log is not, for that reason alone, an independent one. Immutability protects a record after its creation; independence concerns who fixes the record, and observes its creation, at the moment it is made.
VII. HOW THIS DIFFERS FROM THE REST OF THIS SERIES
This is the third article in this corpus's series applying SOURCE 0 QUALIFICATION SEAL to documented age-determination disputes. The first, addressing the federal multidistrict litigation against Meta proceeding in Oakland, concerns provenance, completeness, and coverage of an already-operating record-keeping system — the question of what was silently excluded from a broader dataset. The second, addressing the Roblox MDL, concerns a determination mechanism that changed mid-case — the question of which mechanism applied to a given account at a given moment during a litigated period spanning that change. This article concerns neither. It concerns a single determination event produced by an automated classifier, feeding a reporting pipeline that can result in a filing required under applicable federal law (such as a report to NCMEC's CyberTipline under 18 U.S.C. § 2258A), where the only failure-detection mechanism described in the record was a third party's ex post, case-by-case manual check — not a question of dataset completeness, and not a question of which system version governed a given period, but of whether the record disclosed any independent fixation of the determination event itself, prior to that check. Taken together, the three articles address different facets of the same underlying question: how automated determinations are recorded, versioned, and made independently verifiable once litigation puts them at issue.
VIII. WHEN THIS APPLIES, AND WHEN IT DOES NOT
The relevant trigger is not the presence of automation or AI alone. It is the combination of an automated or autonomous determination, a material downstream consequence, and dependence on records produced or controlled by the same system or operator for any later reconstruction of what was determined. Where all three are present — as in the pattern described in Sections II and III — the gap this article addresses can recur regardless of industry or jurisdiction. It does not apply, and SOURCE 0 QUALIFICATION SEAL does not claim to apply, where no legally consequential act depends on the determination, or where an equivalent independent, contemporaneous control already exists. Materiality and proportionality remain matters for the organization's own counsel to assess before any implementation is considered.
CLOSING AXIOM
A report can be filed the instant a classifier decides. An independently provable record of what the classifier decided, for whom, and when, does not exist by default — it exists only if someone fixes it, independently, before the report leaves the building.
REFERENCE NOTE
SOURCE 0 and SOURCE 0 QUALIFICATION SEAL are proprietary evidentiary architectures developed and operated by Jean-François ELSEN. This article is an authored public release, not legal advice, and does not constitute an assessment of any named organization's current systems, policies, or compliance posture. Facts concerning State ex rel. Torrez v. Meta Platforms, Inc. are drawn from public court filings and contemporaneous press accounts cited in this article's sourcing; where testimony is summarized rather than quoted, the summary follows the cited source's own characterization.
REGULATORY NOTICE
Where the applicable regime requires proof of diligence, of a result, of a notification or of substantive compliance, SOURCE 0 can provide an antecedent, independent element on which that proof may, where relevant, rely. SOURCE 0 does not substitute itself for any of these standards, including 18 U.S.C. § 2258A, the Unfair Practices Act of New Mexico, or any order entered in the case discussed above. SOURCE 0 operates at the evidentiary layer beneath these obligations: it does not alter what must be reported, to whom, or when, and does not itself determine whether a reporting duty has arisen — it fixes an independent record of the internal determinations that may feed such reports. 18 U.S.C. § 2258A separately imposes preservation obligations concerning reported content and associated historical information; that preservation duty concerns the report itself and is a distinct object from independently fixing the determination that led to the report in the first place.
FREQUENTLY ASKED QUESTIONS
Is SOURCE 0 QUALIFICATION SEAL required by the New Mexico judgment, COPPA, or any other law cited here?
No. No provision discussed in this article requires the use of SOURCE 0 QUALIFICATION SEAL or any equivalent architecture. The judgment orders Meta to build a more capable determination system; it does not order that determinations be independently fixed at the moment they are produced.
Does SOURCE 0 QUALIFICATION SEAL verify whether an age or adult-status determination was correct?
No. It fixes an independent, timestamped record that a specific determination was produced, for a specific pseudonymized account, at a specific instant, with the system identity and version asserted at the point of determination. Whether the determination itself was accurate remains a separate question, assessed through model testing, validation, external audit, or downstream review.
Why does it matter when an error is caught, if the report is eventually corrected?
Because the report can reach law enforcement before any downstream correction occurs. Where an error is detected only through later independent review, as described in the New Mexico trial record, the correction arrives downstream of the reporting event rather than before it.
Could Meta's own internal logs establish what the classifier determined at the time?
An internal log may be accurate and may carry substantial evidentiary weight, particularly where it is append-only, cryptographically signed, or otherwise well controlled. But it is not independent merely because it is immutable or technically robust — its evidentiary status still depends on the fact that it was produced, protected, and preserved by the same party whose determination is in question. SOURCE 0 QUALIFICATION SEAL addresses independence of fixation, not merely immutability of storage; the two properties are not the same and an organization can have one without the other. In short, an internal log can be necessary; it is not, by itself, sufficient where independence of fixation is a material requirement.
Does this apply only to child-safety classifiers, or more broadly?
The pattern is not specific to child-safety systems. It recurs anywhere an automated determination feeds a legally consequential act — a report, a filing, an eligibility decision — without an independent, contemporaneous record of what was determined, for whom, and when.
How does this differ from the SOURCE 0 article on the Roblox litigation?
That article addresses a determination mechanism that changed during the litigated period. This article addresses a single mechanism for which the record examined here did not disclose independent fixation of individual determinations at the moment of production, regardless of whether the mechanism itself changed.
