Most AI in compliance operates as a black box. It produces an output, but the reasoning behind it is opaque. When an auditor asks why an alert was closed, or a regulator questions how a decision was made, teams are left reconstructing logic after the fact.
Sphinx was built around a different assumption: if a decision can't be traced, it shouldn't be made.
The Interpretable Agentic Framework is how we deliver on that. It governs how every Sphinx agent reasons, documents, and reaches a conclusion. Every decision is structured, reproducible, and auditable by default.
Three Agents. One Decision.
Most AI systems assign a single model to produce a single answer. The Interpretable Agentic Framework takes a different approach: every decision is the product of three specialized agents working together.
- The Prosecutor scans the input for red flags. It applies institutional risk rules, watchlists, sanctioned country indicators, and adverse media signals to build the initial risk case.
- The Defender interrogates that case. It looks for mitigating evidence — weak matches, missing context, safe conditions that could reduce the risk score. This is especially valuable for false positives driven by name similarity or common entities.
- The Judge weighs both sides using a structured, versioned decision rubric. It doesn't average scores. It arbitrates based on transparent criteria, with options for escalation, human review, or ambiguity flags. Its output is a final recommendation with visible rationale at every step.
Each agent sees the reasoning of the others. This shared awareness reduces contradictions, suppresses hallucinations, and produces decisions that hold up under internal QA and regulatory review.
Genes: Version-Controlled Logic for Every Decision
Inside the framework, every agent's behavior is governed by what we call a Gene — a versioned, signed bundle that couples the agent's instructions with formal policy constraints.
A Gene contains four components:
(1) The prompt template that defines the agent's task and expected output
(2) The logic policy that enforces institutional thresholds and rules
(3) The fallback graph that specifies what happens in ambiguous or low-confidence scenarios
(4) The signed metadata that provides versioning, authorship, and timestamps
This means every decision made in production can be traced back to the exact logic configuration that produced it. If something goes wrong, the previous version can be restored in seconds. If a regulator asks why a decision was made six months ago, the Gene that governed it is still on record.
Compliance teams get the same change management controls they expect from traditional software: peer review, testing, and immediate rollback.
Built to Catch Its Own Mistakes
During shadow testing on a sanctions workflow, our agents repeatedly underweighted geographic information when address fields were truncated by frontend rendering limits. The result was false dismissals for name matches that should have been escalated.
Analysts caught it during review. But the error was systematic and reproducible. It became a design-level change.
Now, ambiguous or incomplete inputs trigger an automatic fallback to analyst review. Contradictory outputs across agents route the case to manual review. No decision proceeds when internal logic diverges beyond a defined threshold. And rollback is instantaneous — reverting to a prior Gene version takes seconds and is logged automatically.
We share this example because trust in compliance automation isn't built by claiming perfection. It's built by showing what happens when things go wrong and proving the system is designed to handle it.
Human Oversight at Every Stage
The framework is designed so that institutions control how much autonomy agents have, and that control can change over time.
Institutions configure where agents must pause — typically in high-risk scenarios like sanctions clearance, PEP downgrades, or SAR submissions. At these checkpoints, analysts inspect the agent's reasoning and approve or override the recommendation. All overrides are logged with timestamps, analyst IDs, and before-and-after state.
Feedback from these interventions flows back into the system. When an analyst corrects a decision, that signal is structured, tracked, and used to refine agent behavior through versioned Gene updates — not ad hoc model retraining.
As institutional trust builds, organizations can scale back interruptions and allow greater autonomy without modifying core logic or compromising auditability.
The Result
The framework was validated across 10,000 historical cases spanning five compliance workflows — AML screening, transaction monitoring, KYB processes, enhanced due diligence, and SAR filings. The results:
- 91.2% agreement with expert human analysts
- 10x improvement in throughput.
- 83.9% reduction in false positives
- Case backlogs decreased by an average of 89% within 16 days of full deployment
- All participating institutions maintained 100% regulatory examination pass rates with zero material weaknesses identified.
Quality held under volume. There was no correlation between processing speed and accuracy. The system performed consistently across workflows, from sanctions screening at 96.1% agreement to complex SAR determinations at 85.7%.
Designed for Scrutiny
Every decision the framework produces can be replayed, inspected, and reversed. There is no hidden state. No silent drift. The architecture aligns with model risk governance standards including SR 11-7, and requires no backend refactor to deploy.
For institutions evaluating Sphinx, the question isn't whether AI can automate compliance. The question is whether it can do so in a way that survives an audit. The Interpretable Agentic Framework is our answer.