Inference Control for Legal Document Generation

Regulatory Framework

The legal-AI regulatory perimeter is no longer hypothetical. Five overlapping regimes now bind any system that produces text that will be filed, executed, or relied upon as legal work product.

ABA Model Rule 1.6 and Formal Opinion 512. Rule 1.6 prohibits disclosure of information relating to the representation absent informed consent. Opinion 512 extends this to generative AI: a lawyer who submits client information to a self-learning tool without contractual confidentiality protection commits a per-se violation, and the duty of competence under Rule 1.1 requires the lawyer to understand the tool's data flows. A document generation pipeline that mixes client matters in shared context, or that emits a clause derived from another client's privileged terms, triggers Rule 1.6 liability regardless of whether the disclosure is detected.

ABA Model Rule 5.5 and unauthorized practice. A clause that purports to bind parties under a jurisdiction in which the supervising lawyer is not admitted is, in many states, unauthorized practice of law. Generators trained on multi-jurisdictional corpora produce such clauses by default; the pipeline must be aware of the engagement's jurisdictional scope and suppress generation outside it.

FRCP Rule 11 and Rule 26(g). Every signature on a pleading, motion, or discovery response certifies that the lawyer has conducted a reasonable inquiry into the factual and legal contentions. Mata v. Avianca made clear that this duty is non-delegable to a model. The system must produce a record that supports the inquiry, citation by citation, at the moment of drafting.

EU AI Act Annex III §8 and Articles 9-15. AI systems intended for administration of justice and democratic processes, and tools that materially assist judicial or quasi-judicial reasoning, are high-risk. They must implement a risk management system (Art. 9), data governance (Art. 10), technical documentation (Art. 11), automatic event logging sufficient for traceability (Art. 12), transparency to deployers (Art. 13), human oversight that can intervene before output is acted upon (Art. 14), and accuracy and robustness commensurate with intended use (Art. 15). The logging obligation is the load-bearing one for legal generation: the deployer must be able to reconstruct, after the fact, why the system produced what it produced.

GDPR Article 22 and the EDPB AI guidance. Where generated documents bear on the legal status of natural persons, Article 22 prohibits solely automated decisions producing legal effects absent explicit basis. The EDPB's December 2024 opinion on the use of personal data to develop and deploy AI models establishes that legitimate-interest processing requires demonstrable necessity and balancing, including technical measures that prevent regurgitation of training data. Inference control's structural admissibility check is one such measure.

FRE 901 and 902(13). Federal Rule of Evidence 902(13), as amended, allows self-authentication of records generated by an electronic process or system. To qualify, the proponent must produce a certification that describes the process and supports its reliability. A generation system whose admissibility decisions are logged at the transition level produces precisely that artifact; a system that emits free-form text and relies on human review does not.

Architectural Requirement

The regimes above converge on a single architectural demand: governance must be evaluable per semantic transition, not per document. A document is the integral of thousands of small decisions about what word, clause, citation, or commitment to commit next. Rule 11 inquiry, Article 12 logging, and FRE 902(13) certification all require evidence at that granularity. Anything coarser is reconstruction.

The system must therefore expose three primitives. First, a persistent matter state that carries jurisdiction, engagement scope, conflict checks, client confidentiality partition, accumulated clause commitments, and the citation graph constructed so far. Second, an admissibility predicate that takes a candidate transition and the matter state and returns admit, reject, or steer-toward-alternative, with the reason. Third, an immutable transition log in which every evaluated candidate, the predicate's verdict, and the rule that governed the verdict are recorded in a form that survives the matter and is producible to a court, regulator, or auditor.

These primitives are not instrumentation around a generator. They are the generator's control surface. The model proposes; the predicate disposes; the log remembers.

Why Procedural Compliance Fails

The dominant compliance posture in legal AI is procedural: a policy document, a training module, a human-in-the-loop review step, and an after-the-fact log of model inputs and outputs. Each component is necessary; none is sufficient.

Policies do not constrain inference. A policy stating that the firm will not use AI to produce unauthorized-practice content does not prevent the model from emitting Texas non-compete language in a California matter. Training reduces the rate of misuse but cannot reduce the rate of model-induced jurisdictional drift, because the lawyer cannot supervise tokens.

Human review is the formal mechanism of last resort and the actual mechanism of first failure. Mata turned on a reviewer who did not catch fabricated citations. Reviewer attention is finite and degrades on volume; on a thousand-clause portfolio, the marginal clause receives seconds. Worse, post-hoc review cannot satisfy Article 12: the log must capture why the clause was admissible at generation, and a reviewer's silent approval is not that record.

Input/output logging is necessary but evidentiarily thin. It shows the prompt and the completion. It does not show that a candidate non-compete duration of thirty-six months was rejected against the matter state's two-year California ceiling, replaced with a twenty-four-month alternative, and committed only after a conflict check against the client's existing customer-non-solicit covenants. That decision chain is the FRE 902(13) certification. Without it, the document is a generated artifact, not an authenticatable record.

Templates are the other procedural fallback, and they fail in the opposite direction. A template that is rigid enough to guarantee jurisdictional validity is too rigid to address the matter; a template that is flexible enough to address the matter relies on free-form fill-ins that recreate the original problem. The structural fix is not a smarter template. It is a smarter transition.

What AQ Primitive Provides

Adaptive Query's inference-control primitive instantiates the three architectural elements described above as a single, transition-level governance loop. Each candidate semantic step proposed by the underlying language model is intercepted before commitment and evaluated against the matter's persistent state. The state carries jurisdiction, the engagement letter's scope, the conflict-check ledger, the confidentiality partition (which client, which matter, which privilege regime), the accumulated clause commitments, and a structured representation of every authority cited so far in the document.

The admissibility predicate evaluates the candidate against five families of constraint simultaneously. Jurisdictional constraints reject clauses whose substantive terms violate the governing law of the matter, including blue-pencil ceilings on non-compete duration and geographic scope, statutory unconscionability limits, consumer-protection mandatories, and choice-of-law restrictions. Precedent constraints check that any cited authority exists, is good law, was decided by a court whose holdings bind or persuade in the relevant forum, and supports the proposition for which it is cited; this is the structural answer to Mata. Scope constraints reject clauses that exceed the engagement letter or that purport to advise on matters outside the supervising lawyer's admission, the structural answer to Rule 5.5. Confidentiality constraints reject transitions that would import language traceable to another client's privileged work product, the structural answer to Rule 1.6 and Opinion 512. Internal-consistency constraints reject clauses that conflict with prior commitments in the same document, eliminating the indemnification-versus-limitation-of-liability class of defects.

When a candidate is rejected, the engine does not surface an error to the user. It steers generation toward an admissible alternative within the same semantic neighborhood, preserving fluency while restoring compliance. Semantic budgets bound the entropy of each clause, preventing the obfuscatory padding that unconstrained generation produces and that Rule 11's reasonable-inquiry duty implicitly disfavors.

Every evaluation, admissions and rejections alike, is written to an append-only transition log keyed to the matter. The log is the Article 12 record, the Rule 26(g) inquiry trail, and the FRE 902(13) certification substrate, produced as a byproduct of generation rather than as a separate compliance artifact.

Compliance Mapping

ABA Model Rule 1.6 / Opinion 512. The confidentiality partition in matter state, enforced by the admissibility predicate's confidentiality family, prevents cross-matter contamination at the transition level. The transition log evidences supervision under Rule 5.3.

ABA Model Rule 5.5. The scope-constraint family rejects transitions outside the supervising lawyer's admitted jurisdictions. Unauthorized-practice clauses are not generated and then redacted; they are not generated.

FRCP Rule 11 / Rule 26(g) / Mata v. Avianca. The precedent-constraint family verifies citation existence, currency, and proposition fit before the citation commits. The reasonable-inquiry duty is discharged at the moment of drafting and the log proves it.

EU AI Act Annex III §8. The transition log satisfies Article 12 automatic event logging. The admissibility predicate is the technical measure required under Article 9 risk management. The matter-state schema is the data-governance artifact under Article 10. Human oversight under Article 14 is preserved because steered alternatives surface to the lawyer for approval before document export.

GDPR Article 22 / EDPB AI opinion. Because the predicate, not the model, makes the admissibility decision, the system is not solely automated in the Article 22 sense; the lawyer's review of the admissible draft is the meaningful human element. The confidentiality partition and the rejection of regurgitated training-corpus language address the EDPB's necessity-and-balancing test.

FRE 901 / 902(13). The transition log is a process record sufficient to support a 902(13) certification: it describes the admissibility process, identifies the rule applied at each step, and is producible without relying on reviewer testimony.

Adoption Pathway

Adoption proceeds in four phases. Phase one (weeks 1-4): scoping. The firm or legal department identifies a document family with high volume and stable jurisdictional structure, typically employment agreements, NDAs, or vendor contracts, and encodes the matter-state schema for that family. Engagement-letter ingestion, conflict-system integration, and the jurisdictional rule pack are configured.

Phase two (weeks 5-10): shadow generation. The inference-control pipeline runs alongside existing drafting workflows. Lawyers continue to draft as they do today; the system generates a parallel draft and a transition log. Discrepancies are reviewed, the rule pack is refined, and the admissibility predicate is calibrated against the firm's house style and risk tolerance.

Phase three (weeks 11-16): supervised production. The system becomes the primary draft generator for the chosen family. Every output is reviewed by a supervising lawyer, but review collapses from substantive correction to strategic assessment because structural validity is enforced at generation. Transition logs begin populating the firm's evidentiary archive.

Phase four (week 17 onward): expansion. Additional document families are onboarded. The firm publishes its FRE 902(13) certification template and Article 12 logging documentation. The transition-log archive becomes the substrate for malpractice-insurance underwriting discussions and for client-side AI governance attestations under EU AI Act Article 26 deployer obligations.

The pathway is incremental, evidence-producing, and reversible at every step. The artifact at the end is not an AI tool that the firm uses. It is a document factory whose every output arrives with the compliance record already attached.