The Agentic Bank

Regulatory Change Parser

⬡ Marginalia Parses incoming regulation into structured, mapped obligations.
◆ Autonomous Specialist

Ingests every rule, consultation and enforcement action on publication, extracts the discrete obligations, and maps each to the policies and controls it touches. Persists an obligation register outside context and emits a change pack gated by an interpretation-judge agent before it commits.

Memory

Working The rule under review + the obligations extracted so far.
Episodic How prior, similar rules were interpreted and mapped.
Semantic The bank's policy/control inventory and the regulatory taxonomy.
Procedural Extraction patterns refined from judge-agent corrections.
Store Vector + knowledge-graph (obligation → policy → control) hybrid

Orchestration

pipeline MCP

Harness · Managed Agents … long-running session; structured note-taking (an obligation register persisted outside context) across multi-document rule sets.

Tools

{ } Regulator publication feeds API { } Policy & control inventory API Legal & regulatory corpus Retrieval ›_ Obligation register Code exec Interpretation judge agent A2A

Evals & guardrails

  • Gold-set precision/recall on obligation extraction vs. the curated label set.
  • Citation discipline … every mapped obligation links to the exact rule clause.
  • An independent interpretation-judge agent re-derives the mapping before it commits.
  • Drift detection on extraction quality as regulators change drafting style.

Offline reflection

Offline replay of judge-agent corrections as Reflexion-style lessons, sharpening the extraction and mapping patterns (procedural memory) rather than just storing them.

Frontier edge

  • Eval-gated continual learning (SEAL-style): self-edits its extraction rules from each judge-agent correction, gated on the gold-set before they go live.
  • Causal mapping, not keyword overlap: reasons counterfactually about which controls a clause actually binds, surfacing second-order obligations a string match misses.
  • Self-improving fleet: newly-learned drafting conventions for one regulator propagate to sibling parsers between runs.

A sample run

Trigger A 240-page final rule on operational-resilience publishes.
  1. 1Segment the rule; extract each discrete obligation with its effective date.
  2. 2Retrieve interpretations of the related prior consultation for continuity.
  3. 3Map each obligation to affected policies, controls and likely owners.
  4. 4Draft a change pack with a gap assessment and citations.
Output A change pack mapping 47 obligations to 19 controls and 6 policy owners, with gaps flagged … gated by the interpretation-judge agent before it commits.

In numbers

120+
Rules parsed / week
52 min
Median parse turnaround
94%
Obligation extraction recall

Handoffs

More on the Regulatory Compliance & Change Management desk