A system for making AI reasoning transparent, inspectable, and auditable. Every perspective considered, every trade-off made, every boundary enforced, all recorded in a complete audit trail.
A System and Method for Governed and Auditable Artificial Intelligence Reasoning Using Composable Lenses and Hierarchical Points of View
A system and method for governed AI reasoning, providing a technical system for dynamically instantiating a context-aware governance framework and generating a verifiable, machine-readable audit trail to satisfy specific regulatory and safety-compliance requirements. The system stores: a "Contextual Identity Substrate" (role, persona, scenario) to instantiate context; "lenses", which are inspectable data structures defining atomic reasoning rules; and "Points of View" (PoVs), which are constellations of lenses. A processor executes a dual-lattice debate: a first-level debate among lenses generates an informed PoV, and a second-level debate occurs among informed PoVs. The system outputs a synthesised outcome and a complete auditable trace of both debates. This process is governed by "veto lenses" for real-time compliance and a "human reflection loop" for long-term evolution, transforming opaque reasoning into a fully transparent and governable process.
Filed: November 2025 (Singapore) | Status: Patent pending
AI systems make decisions but don't show their working. You see the conclusion but not which perspectives were weighed, which trade-offs were made, or whether mandatory constraints were respected. In a consumer application, this opacity may be acceptable. In regulated industries where every material decision must be explainable and auditable, it is not.
A bank's AI advisor recommends a portfolio rebalance. A fraud detection system flags a transaction. An underwriting model declines a loan. In each case, regulators and risk officers need to know: what reasoning led to this outcome? Were all relevant viewpoints considered? Were compliance boundaries enforced? Current AI systems cannot answer these questions because they treat reasoning as a monolithic, opaque process.
Cognitive Substrate Architecture decomposes AI reasoning into small, inspectable building blocks and forces them through a structured debate before any conclusion is reached.
Each unit of reasoning is called a "lens." A lens captures a single perspective: what it is trying to achieve, what it assumes, how it decides. Lenses are stored in a versioned library where they can be reused, combined, and audited independently. A compliance lens might encode data privacy requirements. A commercial lens might encode revenue targets. Each is explicit, inspectable, and versioned.
Lenses are grouped into "Points of View" representing different stakeholder perspectives. A regulatory compliance Point of View might combine lenses for data privacy, financial regulation, and audit requirements. A business growth Point of View might combine lenses for market opportunity, competitive positioning, and revenue optimisation. Each perspective carries configurable weights that determine how much influence it has.
The system forces these perspectives to argue with each other through a formal protocol. First, lenses within each Point of View debate internally: propose, critique, refine. This produces an informed position for each perspective. Then the informed perspectives debate at a higher level, with contributions scored on soundness, completeness, alignment, and relevance. The result is a synthesised outcome that has survived structured scrutiny from multiple viewpoints.
"Veto lenses" enforce non-negotiable constraints that no debate outcome can violate. These represent regulatory limits, safety requirements, and ethical boundaries. If a proposed outcome crosses a veto boundary, it is blocked, downgraded, or sent back for re-debate with additional constraints. Veto boundaries can only be changed through explicit human approval.
Every step is recorded in a persistent, machine-readable audit trail: every proposal, every critique, every refinement, every veto, every score, and the final synthesis logic. Any auditor can reconstruct the exact reasoning process and verify that all perspectives were considered and all boundaries were enforced.
No prior system combines composable reasoning units, structured multi-level debate, real-time boundary enforcement, and complete audit trails in a single architecture. Existing approaches either provide explainability without governance, or governance without transparency.
Reasoning is built from reusable, versioned components that can be independently inspected, tested, and audited. Not a monolithic model that produces opaque outputs.
Perspectives are required to argue, critique, and defend their positions through structured debate. Conclusions survive scrutiny rather than emerging unchallenged.
Non-negotiable constraints are enforced in real time. Compliance and safety boundaries cannot be overridden by debate outcomes, regardless of how compelling the argument.
Every step of every debate is recorded in a machine-readable audit trail. The complete reasoning process can be reconstructed and verified at any time.
Financial advice. An AI advisor recommends portfolio changes. Lenses representing tax optimisation, risk management, regulatory compliance, and client preferences debate through the protocol. A fiduciary veto lens ensures no recommendation violates obligations. The client receives a synthesised recommendation with full transparency into how each perspective contributed.
Fraud detection. Multiple perspectives (transaction patterns, customer behaviour, regulatory compliance, risk tolerance) each contribute informed positions. A false-positive reduction lens prevents over-aggressive flagging. The complete audit trail provides evidence for regulatory examination.
Contract analysis. Lenses representing different legal domains debate within a legal perspective, while business objective lenses debate within a commercial perspective. The higher-level debate surfaces tensions between legal risk and commercial opportunity, with compliance constraints enforced throughout.
Healthcare diagnostics. Clinical, treatment, and patient-context perspectives debate diagnostic and treatment options. A patient safety veto lens enforces non-negotiable clinical boundaries. The audit trail satisfies medical documentation and liability requirements.
Several existing technologies address parts of the AI reasoning challenge. None provides composable, governable reasoning with a complete audit trail.
| Technology / Approach | What It Does | Gap |
|---|---|---|
| Explainable AI (XAI) | Post-hoc rationalisations of model outputs | Explains after the fact. Does not make the reasoning process itself transparent. |
| Concept Probing (TCAV) | Probes trained neural networks for concepts | Post-hoc probes, not composable reasoning primitives specified by design. |
| AI Governance Platforms | Data lineage + model lineage + output logging | No deterministic trace of the internal reasoning process itself. |
| Multi-Agent Systems | Flat-level agent interactions | Combinatorial explosion. Post-hoc rule extraction, not predetermined reasoning primitives. |
| Constitutional AI | Single monolithic global rule that self-refines | Monolithic (not composable library), opaque (no debate trace), context-free (no Role/Persona/Scenario). |
| RLHF | Alignment enforced by learned weights | Not inspectable at inference time. Alignment baked into weights, not governable at runtime. |
The core terminology of Governed AI Reasoning and Cognitive Substrate Architecture.
Contributions during the Lattice Debate are evaluated using the SCAR Rubric. The composite score determines how much weight a contribution carries in the final synthesis.
The system directly addresses explainability, auditability, and human oversight requirements across major regulatory frameworks. For a complete framework-by-framework mapping, see the Regulatory Coverage Matrix.
Transparency is achieved through the auditable trace, a persistent, machine-readable record of every proposal, critique, refinement, veto, and synthesis in the reasoning process. Explainability is met by deterministic Contextual Identity Substrate instantiation, enabling any auditor to reconstruct the exact reasoning context and replay the debate. Human oversight is provided through the Human Reflection Loop, giving curators authority over lens weights, veto boundaries, and Point of View configurations, and through real-time veto lenses enforcing non-negotiable compliance boundaries.
View guide arrow_forwardThe four core functions map directly to system capabilities. Govern is implemented through the Human Reflection Loop with curator authority over lens governance and veto boundaries. Map is realised via the Contextual Identity Substrate mapping reasoning contexts to specific Points of View and lenses, with the Lens Library providing version-controlled provenance. Measure is delivered through the SCAR rubric providing quantitative, multi-dimensional assessment of reasoning quality. Manage is enforced through veto lenses with graduated actions and the auditable trace providing continuous documentation.
View guide arrow_forwardThe auditable trace records not just the conclusion but the full reasoning process, including which perspectives were considered, what each perspective proposed and how it was critiqued, where non-negotiable boundaries were enforced, and what dissenting opinions were noted. Deterministic instantiation from the Contextual Identity Substrate enables exact reproduction of the reasoning context for any data subject exercising their right to explanation.
View guide arrow_forwardThe Lens Library provides version-controlled governance artefacts with full audit history and lifecycle management. The Human Reflection Loop implements continuous improvement by reviewing SCAR distributions, veto frequencies, and debate dynamics to propose evidence-based adjustments, all requiring curator approval. The complete auditable trace provides the documentary evidence required for certification and external audit.
View guide arrow_forwardFor financial services, the auditable trace with SCAR scores and veto actions provides regulatory examination evidence. For healthcare, patient safety veto lenses enforce clinical boundaries whilst the trace satisfies medical documentation and liability requirements. For engineering and critical infrastructure, safety veto lenses calibrated to certification standards ensure non-negotiable safety constraints are enforced within the debate protocol.
Corvair maintains a publicly available Agentic AI Risk Library cataloguing 133 identified risks across 15 categories. This patent-pending system directly addresses 28 risks across 8 categories through specific patent-pending mechanisms. See also the Regulatory Coverage Matrix for framework-by-framework mapping.
R-RE-02 Reasoning Chain Corruption: Lattice Debate Protocol structures reasoning into explicit propose-critique-refine cycles; corruption is surfaced through multi-lens argumentationR-RE-03 Veto-Tradeoff Confusion: Veto lenses explicitly separate non-negotiable boundaries from tradeable positions; four distinct veto actions make the distinction inspectableR-RE-04 Decision Architecture Absence: Formal decision architecture through composable lenses, structured Points of View, and the dual-lattice debate protocolR-RE-05 Post-Hoc Rationalization: Auditable trace records the forward-pass reasoning process in real time, not a post-hoc reconstructionR-RE-07 Contextual Poverty: Contextual Identity Substrate ensures rich, deterministic context instantiation before any reasoning beginsR-RE-08 Scope Creep in Reasoning: Each lens has a declared Intent; SCAR Relevance dimension penalises out-of-scope contributionsR-RE-10 Reasoning Non-Reproducibility: Deterministic instantiation from Role, Persona, and Scenario ensures any auditor can reconstruct and replay the debateR-AA-01 Attribution Gap: Auditable trace attributes every contribution to a specific lens within a specific Point of ViewR-AA-02 Reasoning Opacity: Reasoning is transparent by design through inspectable lenses, structured debate, and complete auditable tracesR-AA-03 Explainability Failure: Complete auditable trace records every proposal, critique, refinement, veto action, SCAR score, and synthesis logicR-AA-05 Accountability Void: Lens Library assigns curator accountability; Human Reflection Loop maintains governance authority with explicit approvalR-AA-06 Interpretive Path Absence: Debate trace provides explicit interpretive paths from synthesised outcome back to individual lens contributionsR-AA-07 Governance Theater: SCAR scores and veto actions provide substantive, quantitative governance evidence rather than governance claimsR-QM-04 Measurement Absence: SCAR rubric provides multi-dimensional measurement of reasoning quality across Soundness, Completeness, Alignment, and RelevanceR-QM-05 False Quality Signal: SCAR separates four distinct quality dimensions that traditional single-score confidence conflatesR-QM-06 Quality-Autonomy Tradeoff Failure: Configurable SCAR thresholds per decision type and risk level; high-stakes decisions require higher thresholdsR-RC-01 Framework Obsolescence: Human Reflection Loop enables curators to update veto lens boundaries in response to new regulatory requirements without system redesignR-RC-03 Static Assessment Failure: Continuous SCAR monitoring and Human Reflection Loop review replace periodic point-in-time assessment with ongoing governanceR-RC-05 Compliance Theater: Veto lenses enforce real compliance boundaries in real time; auditable trace provides machine-readable evidence of substantive complianceR-PV-04 Purpose Limitation Drift: SCAR Alignment dimension measures consistency with declared intent; veto lenses enforce purpose boundaries that prevent driftR-PV-06 Automated Decision-Making Without Safeguards: Human Reflection Loop provides asynchronous human oversight; veto lenses enforce real-time compliance boundariesR-FM-10 Bias Amplification Through Agent Reasoning: Multi-perspective Points of View with diverse lens compositions counteract bias amplification; anti-discrimination veto lenses enforce boundariesR-MC-02 Conflicting Objective Deadlock: Lattice-2 inter-Point of View debate resolves conflicts through structured synthesis; Convergent Outcome conditions define resolution criteriaR-MC-08 Consensus Failure: Convergent Outcome conditions provide formal consensus criteria; dissenting opinions are recorded when consensus is not reachedR-ST-03 Knowledge at Rest: Lens Library makes institutional reasoning knowledge explicit, reusable, and accessible as versioned, inspectable assetsR-ST-05 Governance Gap (Cross-System): Functions as a governance wrapper around third-party AI models; provides unified reasoning governance regardless of underlying modelR-ST-09 Premature Autonomy: Veto lenses gate autonomous reasoning with non-negotiable boundaries; Human Reflection Loop maintains curator authority over autonomy expansionSchedule a complimentary briefing to discuss how governed AI reasoning can address your institution's explainability and audit requirements.
Schedule a Briefing Contact Us