The compliance discovery problem

A compliance officer asked whether a new consumer-data product is permissible faces a question that is structurally larger than a term-matching lookup. The applicable obligations are scattered across statute, implementing regulation, agency guidance, no-action letters, consent decrees, and interpretive bulletins, and the same term means different things in different regimes. "Processing" under GDPR is not "processing" under a state privacy act. A retention duty under SEC Rule 17a-4 does not compose cleanly with an erasure right under CPRA. Worse, the analysis is not finished when the relevant passages are found. The officer has to show, later, to an examiner or a court, that the mapping from activity to obligation to control was complete at the time the activity occurred, and that nothing material was missed or discarded without reason.

Two properties are therefore load-bearing for this task, and both are architectural rather than incidental to ranking quality. The first is a durable record of disposition: which passage was kept, which was dropped, on what authority the source carries legal weight, and how the boundary of the considered set was established. Keyword and vector retrieval are organized to return a ranked list, and a disposition record is a separate structure layered above that ranking. The second is a binding between generated text and admitted sources, so that a narrative determination is constrained to cite what the process actually admitted. For a regulated activity, an unsupported citation is not a minor quality issue; it is a defective compliance record.

Built on Semantic Discovery (US Patent Application 19/647,395)

This application is an enabling implementation of the governed semantic discovery layer disclosed in Chapter 10 of United States Patent Application 19/647,395. The disclosure provides the primitives this use case needs, and the regulatory framing below is a faithful mapping of those primitives onto compliance corpora.

The core construct is the discovery object: a traversal-native semantic agent carrying a typed, persistent state comprising an intent field, a context block, a memory field, a policy reference field, a lineage field, an affective state field, and a confidence field. The discovery object traverses an adaptive index organized into entropy-band-partitioned anchor clusters, advancing through successive anchor evaluations rather than scoring every document in the corpus.

At each anchor boundary the discovery object undergoes the three-in-one traversal step, in which three structurally coupled phases execute as a single atomic transition, and in the described embodiments an advance through the index follows completion of all three in sequence:

  • Search. The anchor evaluates the discovery object's current semantic state against the anchor's published reachable semantic neighborhood and produces a candidate transition set. Search is local to the anchor's neighborhood, not global over the full corpus, which bounds the cost of each step and narrows the search space monotonically as the traversal goes deeper.
  • Inference. A local inference engine scores, ranks, or selects among the candidate transitions by semantic match to intent, information gain to the memory field, and progress toward the resolution criterion, then updates the discovery object's state: the intent field is refined, the memory field is extended with the admitted content, and the confidence field is updated. The inference engine is model-agnostic, and may be an embedding-similarity scorer, a rule-based matcher, a probabilistic model, a neural ranker, or a large language model.
  • Governance (execution). The selected transition is evaluated for admissibility against policy constraints, lineage continuity, entropy bounds, and temporal validity, producing one of three outcomes: admit (advance to the next anchor), reject (discard and try an alternative candidate, or backtrack), or decompose (break the transition into sub-transitions that are individually re-evaluated). Every admissibility determination, including rejections and decompositions, is recorded in the lineage field.

Governance here is not a post-hoc filter applied after results are produced. It is a constituent phase of every step: the model proposes a transition, and the substrate decides whether it is admissible. Relevance is therefore not a precomputed global score but an admissibility-verified traversal history that is query-specific by construction, a post-PageRank notion of relevance in which an object is relevant only if a governed path admitted every transition that reached it. Because the per-step overhead is bounded and the search is local, the substrate scales with traversal depth rather than corpus size. The same substrate supports three operating modes over one set of mechanics: human search (return source-grounded objects with their traversal path), agent reasoning (construct an admissibility-verified reasoning chain), and answer synthesis (generate a natural-language answer in which the generation step is itself an admissibility-evaluated traversal step, so that in the described embodiments content not admitted through the traversal is rejected at the gate rather than emitted).

Mapping the invention onto regulatory corpora

A compliance deployment instantiates these primitives directly.

Anchors as authority-scoped regulatory containers. Each anchor in the index represents a coherent slice of a regulatory corpus: a regulation part, an agency's guidance set, an enforcement docket, a jurisdiction's case reporter. The anchor's governance configuration encodes the source type and therefore the legal weight of what it contains. A statute anchor, an implementing-regulation anchor, a binding-guidance anchor, and a non-binding interpretive-bulletin anchor publish different governance configurations, so authority is scoped at the boundary rather than asserted after the fact.

The discovery object as a compliance query under jurisdictional context. The intent field carries the regulated activity and the question being asked ("is this consumer-data flow permissible for an EU establishment serving California residents"). The context block carries the operative jurisdictional and temporal frame, which persists across every step so that, where the declared jurisdictional bound is in force, a transition into a regime outside that bound is surfaced as a governed rejection rather than a silent crossing. The policy reference field carries the organization's own constraints, for example that only binding sources may ground a control mapping while interpretive material may inform but not establish a duty.

Admit, reject, and decompose as compliance dispositions. When the traversal proposes pulling in an enforcement order, the governance phase checks it against the policy reference field (is this source type permitted to establish a duty), lineage continuity (does it cohere with the obligations already accumulated), temporal validity (is it current, unexpired, and not superseded or stayed), and entropy bounds (does it introduce more uncertainty than the configured threshold allows). A superseded rule is rejected, not silently ranked low. A compound obligation that is admissible in part is decomposed into its sub-duties and each is re-evaluated, so a control mapping can advance through fine-grained requirements even when the coarse obligation is not admissible as a whole.

Lineage as the defensible compliance record. Because every disposition is written to the lineage field, the output of a compliance traversal is not a ranked list but a complete admissibility audit trail: the obligations admitted and the source that grounds each, the candidates rejected and the reason (wrong jurisdiction, superseded, non-authoritative source type, out of temporal scope), and the decomposition paths explored. This is the artifact a supervisor or court asks for. It establishes what was considered and why each item was kept or discarded, addressing the completeness question directly rather than asserting it.

Answer synthesis without fabricated citations. When a compliance officer asks for a narrative determination, answer synthesis mode generates it from the accumulated, admitted state, and the generation step is subject to the same admissibility gate as every prior step. If the generated text introduces an obligation or a citation that is not grounded in the traversal's admitted state, the gate rejects the mutation and the generation fails. In the described embodiments an ungrounded citation is handled as a structurally inadmissible mutation rather than as a statistically improbable output.

Embodiments and deployment options

The application admits multiple embodiments, and is not limited to a single instance.

  • Point-in-time compliance assessment. A human-search-mode traversal scoped to a regulated activity and a jurisdiction returns the set of applicable obligations with the path and admissibility record that reached each, producing a dated, defensible snapshot of the mapping as it stood at assessment time.
  • Regulatory change monitoring. Because temporal validity is a governance criterion, a re-traversal of a prior compliance query against an updated index detects obligations that have become inadmissible (superseded, stayed, sunset) and newly admissible obligations introduced since the last assessment, surfacing the delta as the lineage difference between two traversals rather than as a diff of two flat result lists.
  • Control-mapping for an audit. Agent reasoning mode constructs an admissibility-verified chain from a control to the obligations it satisfies, so each leg of the mapping is an admitted transition rather than an asserted association, and the chain itself is the evidence.
  • Multi-jurisdiction reconciliation. Two discovery objects scoped to different regimes can traverse and then undergo a collaborative merge under policy evaluation and conflict resolution, surfacing where a retention duty in one regime conflicts with an erasure right in another as an explicit governed conflict rather than a silently merged result.
  • Source-type authority tiers. The policy reference field can be configured to admit only binding sources for duty establishment in one deployment, or to admit interpretive guidance as informative-but-non-establishing in another, without changing the traversal mechanics.

Alias addressing makes corpora navigable without a centralized lookup table. Regulatory objects are addressed by structured aliases of the form [email protected]/path (for example a regulation part, a guidance set, or an enforcement docket addressed by its jurisdiction and citation path), and an alias is resolved by walking the index itself one path segment at a time through the same governed traversal. Resolution is therefore mutation-aware and subject to the same governance as discovery, so a citation in a compliance record resolves through the same admissibility-verified mechanism that produced it.

A skilled implementer can build this by adopting the adaptive index as the corpus substrate, modeling each regulatory source as anchors whose governance configuration encodes source-type authority and temporal validity, instantiating compliance queries as discovery objects whose policy reference field encodes the organization's authority rules, and reading the lineage field as the compliance record. The inference engine is a deployment choice and need not be a large language model.

Why this is a technical disclosure

This article is a dated, public, enabling description of applying the Semantic Discovery invention to regulatory compliance search. It ties a concrete, valuable application (defensible, audit-trailed requirement discovery across fragmented regulatory corpora) to specific disclosed mechanisms (the discovery object, the three-in-one traversal step, admit/reject/decompose governance, lineage-as-record, and traversal-based relevance). It documents the enabling path from the underlying invention to this domain and records the date on which that application of the disclosed mechanisms was described publicly.

Disclosure Scope

The technology described in this article (the discovery object, the adaptive index, the three-in-one traversal step, admit/reject/decompose governance, traversal lineage, alias resolution, the three operating modes, and traversal-based relevance) is disclosed in United States Patent Application 19/647,395. The regulatory compliance framing, deployment scenarios, and beneficiaries are application context layered on that disclosed technology and are not themselves claims of the patent. Numbers, mechanisms, and guarantees stated here trace to that application; this article does not introduce ranking metrics or benchmark results beyond what is disclosed there.