Integrity and Coherence for Environmental Compliance Agents

Regulatory framework

Environmental compliance operates across at least four overlapping regulatory layers. At the federal U.S. layer, the Clean Air Act (42 U.S.C. Sec. 7401 et seq.) governs criteria pollutants, hazardous air pollutants, New Source Review, and Title V operating permits. The Clean Water Act (33 U.S.C. Sec. 1251 et seq.) governs NPDES discharge permits, stormwater, pretreatment, and Section 404 dredge-and-fill jurisdiction. RCRA (42 U.S.C. Sec. 6901 et seq.) governs hazardous waste generation, transportation, treatment, storage, and disposal under Subtitle C and solid waste under Subtitle D. CERCLA (42 U.S.C. Sec. 9601 et seq.) governs cleanup and joint-and-several liability for releases of hazardous substances. EPCRA (42 U.S.C. Sec. 11001 et seq.) governs Tier II inventory, Toxic Release Inventory (TRI) Form R reporting, and emergency notifications. TSCA (15 U.S.C. Sec. 2601 et seq.) governs new and existing chemical substances and PMN review.

State and tribal authorities implement most of these statutes under EPA-delegated programs, frequently adopting more stringent thresholds, lower trigger quantities, or additional reporting categories. Local air districts, water boards, and waste authorities layer further obligations. A facility operating in California, Texas, and Pennsylvania may face the same federal floor implemented under three different state regulatory cultures with three different enforcement postures.

The European layer adds REACH (EC 1907/2006) for chemical registration, evaluation, authorization, and restriction; CLP (EC 1272/2008) for classification, labeling, and packaging of hazardous substances; RoHS (Directive 2011/65/EU) for restriction of hazardous substances in electrical equipment; and WEEE (Directive 2012/19/EU) for end-of-life electronics. Authorization decisions under REACH Annex XIV depend on socioeconomic analysis whose interpretation must remain consistent across applicants.

The disclosure layer has expanded dramatically. The SEC's climate disclosure rule and the EU's CSRD with its mandatory European Sustainability Reporting Standards require quantified Scope 1, 2, and material Scope 3 emissions, climate-related risk, transition plans, and assurance. The GHG Protocol Corporate Standard and Scope 3 Standard provide measurement methodology. IFRS S1 and S2, issued by the ISSB, define general sustainability and climate-specific disclosure with a forward-looking risk lens. A single facility's emissions number may simultaneously serve EPA Greenhouse Gas Reporting Program 40 CFR Part 98 obligations, CSRD assurance, IFRS S2 disclosure, and an internal-carbon-price decision. Inconsistency between those uses is itself a violation.

Architectural requirement

The architectural requirement that emerges from this stack is not better natural language understanding of regulations. It is structural memory of every interpretation an agent has issued, structural comparison of new interpretations against the historical record, and structural justification when treatment legitimately differs. Compliance agents must operate as systems of record rather than systems of recommendation.

Three domains of integrity must be tracked simultaneously. Normative integrity tracks the interpretive positions the agent has taken: how a particular Subpart W flaring rule is applied, how a borderline Bevill exclusion is read, how the de minimis threshold for a TRI listed chemical is computed when the facility manufactures, processes, and otherwise uses the same substance in different streams. Relational integrity tracks how the agent has treated different regulated entities for similar fact patterns: whether two refineries with comparable upset events received comparable enforcement recommendations, whether two REACH applicants with comparable substitution analyses received comparable authorization recommendations. Temporal integrity tracks how interpretations have evolved across rulemakings, court decisions, guidance memoranda, and consent decrees, and ensures that the assessment applied to a specific reporting year corresponds to the regulatory regime that was actually in force at the time.

The architectural requirement is that these three domains be enforced before output, not audited after. An interpretation that contradicts a prior position must be flagged, justified, or rejected before it reaches the regulated entity, the inspector, or the disclosure document. A retrospective audit that catches inconsistency after a determination has been issued cannot undo the legal exposure created by the determination itself.

Why procedural compliance fails

Most enterprise compliance AI today is procedural. A regulation is loaded into a retrieval index, a prompt template instructs the model to apply the regulation to the facts of a query, and the answer is returned. The procedural pipeline has no memory of prior queries, no comparison against prior interpretations, and no temporal anchor to a specific regulatory snapshot. Two facilities asking the same question on consecutive days may receive different answers because the retrieval ranking shifted, the model temperature varied, or an intervening conversation re-anchored the prompt.

Procedural compliance fails on Clean Air Act applicability determinations, where a single threshold judgment about Potential to Emit cascades into Title V major-source status. It fails on RCRA hazardous waste determinations, where the difference between a listed waste, a characteristic waste, and a non-hazardous solid waste depends on interpretive choices that must be defensible across the generator's full waste stream over years. It fails on CERCLA arranger liability analyses, where the legal test announced in Burlington Northern requires a fact-pattern comparison the procedural pipeline cannot perform because it has no record of how the test was previously applied.

Procedural compliance fails most visibly in disclosure. CSRD double materiality assessments, IFRS S2 climate scenario analyses, and SEC-style transition risk narratives are produced once per reporting cycle but must be internally consistent across thousands of pages and across years. A Scope 3 boundary defined one way in the inventory cannot be silently redefined in the transition plan. A physical risk assumption used in IFRS S2 disclosure cannot contradict a resilience claim in the CSRD narrative. Procedural systems cannot enforce this consistency because they do not retain the prior commitments as enforceable constraints.

The deeper failure is equity. Inconsistent interpretation of identical fact patterns is the operational definition of arbitrary and capricious agency action under the Administrative Procedure Act, and the operational definition of disparate enforcement under environmental justice review. A compliance agent that cannot demonstrate consistent treatment across regulated entities is generating evidence against the program that deployed it.

What the AQ primitive provides

The Adaptive Query integrity-coherence primitive provides a governed substrate in which every interpretation, every comparison, and every transition is a first-class artifact. The normative domain stores the agent's interpretive commitments as durable, citable positions linked to the underlying statutory, regulatory, and guidance authority. When a Clean Water Act NPDES determination is made for a facility, the interpretation of the relevant effluent guideline, the variance posture, and the monitoring methodology are recorded as a normative commitment with an authority chain.

The relational domain operates as a structural fairness check. Each new determination is compared against the population of prior determinations involving comparable fact patterns. The comparison is not statistical similarity over text; it is structural similarity over the regulatory features that the framework itself identifies as material. If a new enforcement recommendation differs from precedent, the agent is required to surface the differentiating factors explicitly and to record them in the integrity ledger. Permissible reasons such as repeat violation history, severity of release, or good-faith disclosure are recognized; unjustified divergence is blocked.

The temporal domain anchors every determination to a regulatory snapshot. When 40 CFR Part 60 Subpart OOOOb is amended, when EPA issues new PFAS hazardous substance designations under CERCLA, when REACH authorizations are revised, the prior regime is preserved as a versioned artifact. Determinations covering activity that occurred under the prior regime continue to apply the prior regime; determinations covering subsequent activity apply the new regime; transition periods are explicit rather than implicit. This eliminates the silent regulatory drift that procedural systems exhibit when underlying retrieval corpora are updated.

Together the three domains produce an audit ledger that can be presented to an administrative law judge, an EPA region inspector, a CSRD assurance provider, an IFRS S2 auditor, or a citizen-suit plaintiff and demonstrate, structurally rather than narratively, that the program applied regulations consistently, equitably, and with explicit treatment of regulatory change.

Compliance mapping

The integrity-coherence substrate maps directly onto specific regulatory obligations. For Clean Air Act Title V, the normative ledger preserves applicability determinations and netting analyses across the five-year permit cycle, allowing renewal review to be evaluated against the original permitting record rather than reconstructed from memory. For Clean Water Act NPDES, the temporal domain preserves the effluent guideline version, the mixing zone analysis, and the receiving water classification in force at each monitoring period.

For RCRA, the normative domain captures generator status determinations, land disposal restriction analyses, and corrective action milestones with the interpretive basis for each. For CERCLA, relational integrity provides the structural fairness record that supports allocation defenses and contribution claims. For EPCRA TRI, the temporal domain ensures that the chemical list, threshold determinations, and waste management calculations applied to each reporting year correspond to the rule in force for that year.

For TSCA Section 5 PMN and Section 6 risk evaluations, the normative ledger preserves the data adequacy and risk characterization choices. For REACH registration dossiers and Annex XIV authorization applications, relational integrity provides the substitution-analysis comparability record. For RoHS and WEEE conformity, the temporal domain preserves exemption versions and producer responsibility scheme commitments.

For SEC climate disclosure and CSRD double materiality, the substrate enforces consistency between the inventory boundary, the materiality assessment, the scenario analysis, and the transition plan. For IFRS S1 and S2, the substrate provides the connective tissue between general sustainability disclosure and climate-specific disclosure that the standards demand but procedural systems cannot guarantee. For GHG Protocol inventories, the temporal domain preserves emission factor versions, organizational boundary changes, and recalculation triggers as auditable events.

Adoption pathway

Adoption proceeds in three phases. The first phase is shadow operation. The integrity-coherence substrate is deployed alongside an existing compliance workflow and observes determinations without gating them. The ledger accumulates the program's actual interpretive commitments, surfaces inconsistencies that previously went undetected, and produces a baseline integrity report that quantifies how often the program currently treats comparable fact patterns differently.

The second phase is gating on high-consequence determinations. Title V applicability, RCRA generator status, CERCLA potentially-responsible-party identification, REACH authorization recommendations, and disclosure boundary commitments are routed through the substrate and require explicit normative, relational, and temporal coherence before issuance. Lower-consequence determinations remain in observation mode.

The third phase is full programmatic integration. The substrate becomes the system of record for compliance interpretation across the enterprise or agency. Regulatory change management, disclosure cycle preparation, enforcement docket review, and external assurance all draw from the same integrity ledger. New AI capabilities, whether for permit drafting, monitoring data interpretation, or sustainability narrative generation, are introduced as governed contributors to the ledger rather than as independent systems whose outputs must be reconciled after the fact.

Programs that complete this pathway gain three durable benefits: defensibility against challenge, equity in enforcement, and predictability for regulated entities. None of these benefits is achievable by improving the underlying language model. They are properties of the substrate within which the model operates, and integrity-coherence is the substrate environmental governance now requires.