Trust Center

We don't ask you to take our word for it. We show you how it works.

What KOREV documents, what is partially documented, what remains to be published — and the known limits. Transparency is part of the product.

Technical publication · v1.0.2

From AI logs to verifiable decision records

Protocol, threat model, hostile scenarios, experimental results, and limits, all documented. Verification covers integrity and continuity within declared boundaries; it proves neither truth nor regulatory compliance.

Evidence / Trust controls

Trust is verified, one control at a time.

Sources, uncertainties, publication, and security: every mechanism has a scope. The badges describe controls and technologies, not certifications or a real-time monitoring status.

Analysis available / Separate validation

Understanding a block is part of the result.

An analysis can explain the risks, the missing information, and the next checks while keeping a non-validated decision status. The record's actual status is what counts.

Decision controls

Security and cryptographic mechanisms

Measures described in the technical documentation provided. Their activation and scope must be confirmed for your deployment; this presentation is not an independent audit of the infrastructure.

A record may show "consensus not executed" or "cryptographic integrity not computed." These states must remain visible and never count as a passed control. Proof of integrity does not prove that the content is accurate.

Decision Evidence / Verifiable Decision Records

Documented

KOREV documents a bounded post-admission continuity property over declared, instrumented boundaries: an admission acknowledged by the witness must remain represented and be reconciled before a population can be classified VERIFIED.

  • Three technical states: VERIFIED, INCOMPLETE, FAILED — none of them is a regulatory label or a proof of truth
  • Targeted v2 study: 67 tests and 22 recorded observations on predefined hostile scenarios
  • Ed25519 witness separated by process and by key in the experiment; independent administration not claimed
  • Frozen references: protocol/publication PR #29 and Conversation → Record hardening PR #30

Published limits: instrumented boundaries only, no global completeness, no semantic truth, no certification, no independent third-party reproduction yet.

Technical note EN (PDF)

Operational maturity

Documented

KOREV Evidence is operated as a deployable product: production releases, a single release-to-production procedure, blocking CI checks, health checks, versioning of the deployed commit, and rollback/diagnostic procedures.

  • Controlled release chain: PRs to main, required checks, and a protected branch
  • Deployment verified through /healthz and VERSION.json.build_commit
  • Operations runbooks, drain before restart, rollback, monitoring, and post-deployment validation
  • Live validations kept separate from synthetic tests, with real providers and official sources on some business pipelines

We do not publish a product TRL (technology readiness level) until the integrated system as a whole has been formally reassessed; the earlier TRL assessment of the PRISM component is archived.

Architecture

Documented

Three layers — Atlas (Enterprise LLM and reasoning), Evidence (execution agents), PRISM (control) — between the business process and the client infrastructure, with five cross-cutting mechanisms: human oversight, audit trail, governance, security, data control.

  • Explicit separation between understanding (Atlas), execution (agents), control (PRISM), and proof (Evidence)
  • Each layer can be adopted separately or together
  • Integration with your tools through APIs and connectors
KOREV architecture

Security

Documented

Technical controls — encryption, isolation, access management, logging — within the chosen deployment environment.

  • Encryption in transit; at-rest coverage to be specified per deployment
  • Role management and isolation scope to be verified in the selected environment
  • Access logging and minimization of sensitive data

Security certifications (ISO 27001, HDS, SecNumCloud…): no certification is claimed on the current site; to be published as actual certification steps are undertaken.

Security page

Governance

Documented

Mechanisms built into the architecture: model identification, traces, sources, human approval, rules, divergences, history, review, permissions, responsibilities.

  • Approval rules defined per process
  • Divergences retained, never hidden
  • History and review capabilities
Multi-agent AI governance FR

Models & reasoning (Atlas)

Partially documented

Atlas, KOREV AI's Enterprise LLM, specialized on your enterprise knowledge and managed with a version registry.

  • Registry of Atlas versions and deployed models
  • Progressive specialization on the organization's knowledge
  • Controlled, traceable enrichment loop

Formalization of the public list of supported models by version, and of the update policy.

Atlas

Data

Documented

Data location, flows, and processing defined by the organization; KOREV does not use client data to train public models.

  • European cloud, hybrid, or on-premises
  • No use of client data by KOREV to train public models; external dependencies made explicit in hybrid mode
  • Configurable retention of traces and sources
Sovereignty and external dependencies

AI Act

Documented

Architecture designed to help meet the oversight, traceability, and governance requirements that apply depending on the use. Not a certification of compliance.

  • Model identification and trace retention
  • Human oversight at defined steps
  • Exportable documentation of decisions and rules
AI Act: requirements and implementation

Human oversight

Documented

Approval points defined process by process; automatic escalation when rules are not met; approvals traced and attributed.

  • When the process policy requires a review, the output concerned stays on hold until approved
  • Designated reviewer, provided with the divergence details
  • Human oversight is configured according to the criticality level and the workflow

PRISM

Documented

Control layer: multi-model cross-checking, divergence detection, configurable consensus, thresholds, rules, fail-closed, human oversight.

  • Does not promise the absence of errors: reduces reliance on a single model
  • Suspends the decision when the conditions are not met
  • Retains what was compared, accepted, rejected, or escalated
PRISM

Evaluations & validation

Partially documented

KOREV already publishes targeted results on Decision Evidence and runs business validation campaigns; Atlas benchmarks and end-to-end performance measurements still need to be consolidated before publication.

  • Decision Evidence v1.0.2: targeted suite and hostile scenarios published
  • Business validation and live tests kept separate from performance benchmarks
  • No throughput or overall latency figure is published without a reproducible protocol

Atlas, latency, cost, and load benchmarks to be published once they have been rerun on a stable public protocol.

Methodology

Documented

How a process is assessed, piloted, and then rolled out to production: Pulse (assessment), Pilot (first use case), Production rollout, Multi-process deployment.

  • Validation criteria defined before the pilot
  • Scope, data, and human approvals made explicit
  • Move to production conditional on the pilot's results
KOREV Pulse

Responsible AI

Documented

Principles: people remain accountable for the steps that require it; no marketing absolutes; limits are documented.

  • No promise of zero errors or of automatic compliance
  • Technical claims tied to a scope and to evidence
  • Known limits published below

Deployment models

Documented

European cloud, hybrid, or on-premises; isolated environment depending on the project.

  • Same architecture at different levels of independence
  • Explicit list of accepted external dependencies
  • The level can evolve over the course of the project
Atlas — deployment

Engineering & auditability

Public register of technical evidence

An inventory of KOREV's engineering claims, measurements, and mechanisms. Each public signal is linked to a method, a precise scope, and an explicit methodological limitation.

Documented mechanismDocumented

14

Documented agent profiles

14 active profiles under agents/: compliance, contradictor, default, developer, finance, hacker, infrastructure, legal_drafting_guarded, legal_safe, marketing, medical, multitask, researcher, sales. The _example folder is excluded.

Methodological limitation: Refers to profiles configured in the Evidence orchestrator, not to 14 distinct AIs or proprietary models.

Documented mechanismDocumented

3 levels

Criticality routing (LEVEL 1 / 2 / 3)

CriticalityRouter architecture (ADR-010 / ADR-011): LEVEL 1 (simple request: consensus bypassed by default, source baseline mandatory); LEVEL 2 (professional zone: analysis/advice, consensus can be enabled through explicit user opt-in or caller force_consensus, source baseline); LEVEL 3 (critical decision, liability, dispute, critical action: PRISM consensus mandatory, strict mode if the domain is critical). Fail-closed always takes priority in case of ambiguity.

Methodological limitation: PRISM consensus is not enabled uniformly across all interactions: it is triggered conditionally, according to the policy and the assessed criticality level.

Documented mechanismDocumented

RSA-PSS / HMAC

Cryptographic signature of critical outputs

critical_output layer (ADR-010): v2 signature binding input_hash, output_hash, consensus_result_hash, criticality_level, policy_id/version, timestamp, trace_id. RSA-PSS-SHA256 algorithm in production, HMAC fallback depending on configuration.

Methodological limitation: Does not apply uniformly to every general chat response; specifically covers critical outputs subject to the ADR-010 doctrine and the dedicated finalization pipeline.

Documented mechanismDocumented · 2026-09-20

v1.0.2

Verifiable decision records

Admissions acknowledged by the witness on the configured path; history continuity, admission→terminal reconciliation, checkpoint freshness, and bundle inventory.

Method usedTechnical note v1.0.2 + v2 protocol + verifier + recorded hostile scenarios.

Methodological limitation: Proves neither semantic truth, nor regulatory compliance, nor the capture of all production paths. Witness separated by process and by key in the experiment, but not administratively independent.

Verified metricReplayed · 2026-09-19

67 tests

Decision Evidence v2 targeted suite

Targeted decision_pack / decision_capture / witness suite, not the product's full test set.

Method usedpython -m unittest discover -s tests -p 'test_decision*.py' -v

Methodological limitation: Does not mean that the product as a whole contains 67 tests, nor that these tests prove correctness in production.

Verified metricEvidentiary snapshot · 2026-09-19

22 / 22

v2 experimental observations

Deterministic synthetic study, with no external LLM calls and no external network traffic.

Method usedTwo decision-capture-v2 runs; 11 hostile scenarios per batch.

Methodological limitation: A study of bounded properties; it is neither a production benchmark nor proof against every possible attack.

Documented mechanismDocumented · 2026-09-17

5 required checks

Controlled production deployment chain

Production release chain for the Evidence backend: protected main branch, mandatory PR, 5 required checks, enforce_admins, build/deploy guardrails, drain before restart, /healthz, and VERSION.json.build_commit.

Method usedOperations runbook + branch protection + deploy_prod.sh / stamp_version.py scripts.

Methodological limitation: Proves deployment discipline and release traceability; it is neither a measured SLA, nor a security certification, nor proof that no incident has occurred.

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.

Method usedFormal consensus specification and run_consensus() implementation.

Methodological limitation: Multi-model consensus reduces the risk of error from a single model but is not a mathematical guarantee of absolute truth.

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).

Methodological limitation: Does not mean that the whole system, or any external third-party provider, is proven fail-closed in 100% of cases; the claim must be strictly limited to the audited scope.

Documented mechanismDocumented

Specialized enterprise reasoning engine

Language model and contextualized reasoning capabilities, drawing on internal documents, reference frameworks, procedures, and controlled interactions with KOREV agents.

Methodological limitation: Contextual adaptation and enrichment take place according to the chosen architecture (contextual RAG, specialized corpora, targeted adaptation), without claiming raw injection of all enterprise data into the model weights.

Known limits

  • Generative models can make errors; PRISM reduces reliance on a single model and detects divergences, but it does not eliminate them.
  • The quality of the results depends on the quality, completeness, and structure of the data and corpora provided.
  • Regulatory compliance depends on the context, the system, its use, and the organization's role; KOREV provides technical capabilities, not a certification.
  • Some integrations (ERP, DMS, CRM) require organization-specific work.
  • References and case studies are published with their maturity level stated explicitly, as and when they can be made public.

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.

Technical FAQ