Verified metricHistorical
76 / 76
Core tests replayed
Deterministic core suite covered by the audit in question (vitest.config.core-only.js).
KOREV /PRISM/ 02
Critical Cognitive Engine
When a decision becomes critical, don't let it depend on a single model.
PRISM brings several independent intelligences to bear on critical processing, cross-checks their conclusions, and applies the defined rules before a result can be used.
Reviewing a case file01 / 04
Model A
Complete file
Model B
Complete file
Model C
Missing document
Policy in this example
Divergence → suspension → human review
Conclusions available
Three analyses. One case file.
Several models examine the same question. Their conclusions stay separate so they can be compared.
Where it fits in the architecture
- Evidence — agents and orchestration
- PRISM — control and approval
- Atlas — Enterprise LLM and reasoning
A model can be wrong. A critical decision calls for reinforced control.
The problem
A single model cannot guarantee the reliability of a sensitive decision.
For critical processing, an isolated generative answer is not enough: the organization must be able to account for the intelligences used, the points of convergence and divergence, the compliance policies applied, and the reason for approval or suspension. A single model, however advanced, provides none of these safeguards on its own.
- 01Misleading confidence: a probabilistic model can state a wrong answer in an assertive tone, with seemingly perfect coherence.
- 02Opaque dependence: changing providers or model versions silently alters judgment calls, with no operational alert.
- 03Superficial review: without a formal control policy, review comes down to a quick skim by a user under pressure.
- 04No recourse: in the event of a challenge or audit, there is no record of the hypotheses compared, accepted, or rejected.
Use
What the critical cognitive engine brings to your operations.
PRISM is not a simple filter: it is a cognitive orchestration engine that maintains the critical context, cross-checks independent intelligences, and applies your decision policy.
Cross-check several intelligences
Critical processing is submitted simultaneously to several independent models. Their conclusions are compared, agreements confirmed, and divergences made explicit.
Apply your decision policy
Documented quorum, validation thresholds, business rules, and regulatory constraints: PRISM assesses whether the criteria required by the process are met.
Suspend and escalate
When the criteria are not met, PRISM switches to protective behavior: the result is suspended and the complete case file is passed to an authorized owner.
Business outcome
What the organization gains with PRISM.
- Greatly reduced dependence on the conclusions, drift, or biases of a single model provider
- Divergences detected and made explicit before any use, instead of errors discovered after the fact
- Formalized approval policies, applied reproducibly to sensitive processing
- Predictable, defined behavior in case of doubt or dispute: secure suspension and escalation
- Actionable information given to the authorized owner to inform the final human arbitration
- Freedom to swap or update the underlying models without altering the governance framework
How it works
From critical processing to an arbitrated decision.
PRISM organizes processing into several coordinated steps. When the conditions are not met, processing is not approved: it is suspended for arbitration.
Step 01 / 06
Critical processing
Criticality assessment & input context
Engineering note: the path has a structural fork — if the quorum or business-rule conditions are not met, the flow emits no direct result and switches to secure suspension, with hand-off to the authorized reviewer.
Operating doctrine
Not every answer carries the same weight.
A document search, an information summary, or drafting assistance does not need the same latency, compute cost, or level of verification as a financial commitment, a medical diagnosis, or a contractual judgment call that involves the company's liability. KOREV therefore matches the level of control to the actual criticality of the processing.
Standard flow · Speed
Direct processing
Routine requests, source searches, summaries, and exploratory analyses run directly through Evidence agents and Atlas, with no computational overhead or unnecessary multi-model latency.
Standard level · Cost & latency optimization
Critical flow · PRISM triggered
Reinforced cognitive control
As soon as a task involves a regulatory judgment call, a contractual commitment, a medical diagnosis, or a sensitive financial decision, the orchestrator engages PRISM to bring in several independent intelligences and apply your approval policy.
Critical level · Cross-checking, 2/3 quorum & conditional suspension
Decision architecture
How PRISM handles critical processing.
From maintaining the contextual state to adversarial cross-checking and strict application of your policy: consensus is not the decision; it is the signal that feeds into it.
PRISM receives critical processing along with its context and policy. Several models produce distinct conclusions. They are cross-checked, and a consensus signal can be computed. The process policy then determines whether processing can continue or must be suspended. A suspension can be passed to an authorized owner for arbitration.
Input · Processing classified as critical
Context preparationOverall case environment
Critical context
PRISM maintains the processing state, the contextual scope of the case, and the policies defined for the process, without relying on an isolated prompt.
Progress stateContextual memoryApplicable policy
Distinct inference pathsAdversarial evaluation
Path A
Model A
Independent intelligence
Isolated conclusion
Path B
Model B
Independent intelligence
Isolated conclusion
Path C
Model C
Independent intelligence
Isolated conclusion
Distinct inference paths within the cross-checking mechanism. The models used depend on the deployment's architecture and policy (local instances, open-weight models, or authorized external services).
Cross-checking
Cross-check matrix
Conclusions are compared point by point. Convergences are isolated, and divergences remain flagged as signals for analysis.
AgreementsDivergencesMissing information
Consensus signal
2 / 3*
of valid votes
* Documented quorum of valid votes in the mechanism concerned.
Core principle: Consensus does not prove truth. Several models can converge on the same error.
Decision junction
Policy EngineDecision policy
Consensus is only one signal among others: the policy configured for the process weighs all the business rules, quorum thresholds, regulatory constraints, and the sensitivity of the operation to authorize or suspend processing.
Formal quorumBusiness rulesCompliance thresholdsRegulatory constraintsReview required
Outcome A · ConformingProcessing continues
Authorized by the policy
When all of the process's quorum, consistency, and compliance requirements are met, processing is validated by the rule and continues its normal execution.
Return to the workflow
Depending on the workflow, an authorized critical output can then be passed to the finalization and integrity pipeline.
Outcome B · Suspension
ConditionalSuspended
When the conditions required by the policy are not met, processing can be suspended rather than continued automatically.
Human arbitrationAuthorized owner
The complete divergence record is passed to the expert for final review.
Add informationRejectAuthorize
Back to the workflow after arbitration
Cryptographic integrity: mechanisms documented and tested within the relevant scope.View the registry in the Trust Center
Distinctions
What PRISM is — and what it is not.
Four key clarifications to pin down exactly what kind of technology the critical cognitive engine is, as opposed to reductive portrayals.
Not a simple majority voter
A 2/3 consensus is not mathematical truth. Several models can converge on the same error. PRISM uses cross-checking as a signal for a strict decision policy, not as a blind vote.
Not a passive proxy or a prompt router
PRISM is a cognitive orchestration engine: it maintains state, keeps the case's contextual memory and applies arbitration rules before anything is passed on.
Not a universal anti-hallucination filter
No probabilistic system fully removes hallucinations. PRISM reduces dependence on a single model and makes divergences visible instead of masking them.
Not a systematic cost on every prompt
PRISM steps in conditionally on critical flows. Simple operational requests run without added latency or unnecessary multi-model computation.
Engineering & proof
Mechanisms backed by a suite of audited tests and contracts.
Claims about PRISM rest on contract tests, adversarial scenarios and documented measurements, presented with their limits of validity.
Verified metricHistorical
31
Property tests & fuzz/adversarial scenarios
24 property-based tests + 7 fuzz / adversarial scenarios in the audited proof suite.
Documented mechanismDocumented
2/3
Documented quorum
Canonical run_consensus() API: a 2/3 quorum of the valid votes cast by the cross-checked models is required.
Technical qualificationQualified
Fail-closed tested
Conditional fail-closed behavior
Fail-closed behavior replayed on the trust contracts, the log, and the providers covered by the proof suite (TrustContext, adversarial tests).
Ed25519 cryptographic verification implemented and covered by property-based tests in TrustContext.See the full protocol in the Trust Center →
Architecture
Atlas understands the enterprise, PRISM protects critical processing.
Evidence runs the operational workflows and calls on Atlas to understand the business context. When processing is assessed as critical, PRISM takes over to maintain the decision state, cross-check several independent intelligences (local or remote), and apply the process policy. The result can then be finalized and sealed for human review.
Evidence organizes the workflow and interacts with Atlas for context and reasoning. After the criticality assessment, processing either follows the standard path or branches to PRISM when critical control is required. The result then returns to Evidence.
Input · Case file or business process
Cognitive coordinationContinuous context & reasoning exchange
Evidence calls on Atlas to contextualize every action
Criticality assessment
Routing by process and policy
The workflow can classify processing by its sensitivity and the applicable policy in order to adjust the level of control.
Standard path
Direct processing
Searches, information summaries, and routine analyses run directly, without systematically triggering multi-model control.
Direct, smooth execution
Delivery in Evidence
Result · Sources · Required approvals
Deliverables are returned to the operational workspace with their justifications and the configured human review points.
Critical control
PRISM
Critical Cognitive Engine
The cognitive engine called on for critical processing to cross-check several intelligences and apply a stricter decision policy.
Key attributes:
Multi-model cross-checking
Policy engine
Fail-closed suspension
Arbitration
Cross-cutting framework · Built into the architecture:
- Human oversight
- Sources & trace
- Governance
- Data control
Capabilities
Capabilities of the critical cognitive engine.
Critical context & state memory
Keeps and uses the state, contextual memory, and history needed to handle a decision within its overall environment, rather than as an isolated request.
Multi-model orchestration
Parallel use of several independent intelligences (local models, suitable open-weight models, or authorized external APIs).
Structured cross-checking
Rigorous comparative analysis of the generated conclusions, isolating points of convergence, nuances of interpretation, and outright divergences.
Consensus & documented quorum
Application of a documented 2/3 quorum of valid votes to characterize the level of agreement, without claiming that consensus is absolute proof of truth.
Decision policy engine
Application of strict business rules, validation thresholds, compliance constraints, and exclusion lists specific to the process concerned.
Conditional fail-closed suspension
Tested protective behavior: if the quorum or business rules are not met, processing is suspended to prevent a non-conforming result from propagating.
Escalation and human arbitration
Structured hand-off of the divergence record to an authorized reviewer, with the points of agreement and disagreement needed for an informed arbitration.
Decision retention and traceability
Timestamped logging of model versions, compared answers, rules applied, and the decision, to allow a later audit.
Modular integration at sensitive points
Triggered directly from Evidence or integrated via API at the critical decision points of your information system.
Governance
An auditable chain of control for sensitive decisions.
PRISM keeps what is needed for later review by an owner, an ethics committee, or an auditor: models used, divergences observed, rules applied, and the reason for the arbitration.
Regulatory compliance depends on the context, the system, its use, and the organization's role. KOREV provides technical governance capabilities; on its own, it does not constitute a legal certification of compliance.
- Precise identification of the models, versions, and execution environments used
- Recording of the compared conclusions, with minority answers made explicit
- Logging of the policies applied, validation thresholds, and business rules
- Formal attribution of suspension decisions and recorded human arbitrations
- Decision history that can be consulted and exported for compliance reviews
- Supporting evidence that can be passed to the finalization pipeline for sealing and integrity checks
Deployment
Three deployment architectures to match your requirements.
PRISM adapts to your security policy and can orchestrate intelligences running locally, in a hybrid environment, or through authorized external services.
On-premises / private infrastructure
A configuration can be designed to run PRISM and the selected models locally within your private infrastructure, with no mandatory dependency on a remote API.
Hybrid architecture
Orchestration combining internal sovereign models for highly sensitive data with complementary remote models, filtered according to your policy.
Dedicated European cloud
A dedicated European environment, with flows and dependencies defined in an environment under European jurisdiction, depending on the chosen architecture.
Use cases
Where critical control makes the difference.
Industry & engineering
- Arbitrating actions on critical nonconformities
- Approving emergency maintenance procedures
- Assessing strategic supplier risks
Healthcare & research
- Quality control of complex regulatory dossiers
- Adversarial verification of health alerts
- Preparing analyses that carry medical liability
Finance & risk
- Arbitration for sensitive credit-approval committees
- Counterparty and fraud risk analysis
- Regulatory controls linked to a verifiable audit trail
Legal & public sector
- Detecting document contradictions in litigation
- Source checks on major contract analyses
- Processing administrative cases with integrity signatures
Don't let your critical decisions depend on a single model.
Let's discuss your most sensitive processes, the associated risks, and the approval policies suited to your organization.
KOREV — AI that shows its work.