KAiM's audit-first framework: AI agents propose, deterministic evaluators enforce. The line is not crossed. Four functional cells, structural privacy protection, and decision-grade audit evidence on every agent action.
Non-crossable boundaries that survive every product decision, every commercial pressure, every architectural pivot. Each line is enforced structurally — in compile-time checks, IAM gates, and type-system constraints — not by policy.
LLM-driven agents may propose meaning, classifications, charter content, or routing recommendations. They may not directly govern runtime enforcement decisions. The Codex Engine compiler refuses to accept LLM output as authoritative input without explicit promotion through deterministic review.
Semantic memory authority (vocabulary, charters, provenance) is structurally separated from runtime enforcement. Neither side can collapse into the other. Provenance is preserved across the boundary.
The Codex Engine intermediate representation refuses to compile rules that emit per-individual scores, classifications, or rankings. This is structural prohibition, not policy. KAiM cannot be used for employee surveillance because the architecture forbids it.
Every deployed agent has a Charter — a typed contract with a named accountable business owner, declared kill switch, declared blast radius, HITL checkpoints, and sunset criteria. The build pipeline structurally rejects any agent submission lacking a valid Charter reference.
Every Charter, canonical term, governance rule, and architectural commitment carries an expires_review_at date. Re-review is continuous discipline, not a one-time event. Nothing is allowed to silently rot.
Each cell does one job and writes to a shared substrate the next cell reads. No cell holds another's authoritative state. The boundary is the architecture.
Business owner
│
▼
┌─────────────────┐ ┌──────────┐ ┌─────────────┐ ┌──────────┐
│ Atelier Cell │───▶│ Forge │───▶│ Bot Village │───▶│ Cobbler │
│ design+charter │ │ build │ │ registry │ │ maintain │
└────────┬────────┘ └──────────┘ └──────┬──────┘ └────┬─────┘
│ │ │
│ binds terms │ runtime │ drift/sunset
▼ ▼ ▼
┌──────────────────────┐ ┌─────────────────┐
│ Lexicographer Cell │◀────────────▶│ Forensics Cell │
│ Helm Meaning Layer │ signals │ continuous int. │
└──────────────────────┘ └────────┬────────┘
│
│ proposals (typed)
▼
┌─────────────────┐
│ Codex Engine │
│ rules + IR + AI │
└────────┬────────┘
│
▼
┌─────────────────┐
│ Helm Runtime │
│ enforcement │
└─────────────────┘
Builds and governs the enterprise's canonical vocabulary, ontology, and taxonomy. Source of truth for both agent context and data-lake classification. Five agents, three load-bearing separations, four-level provenance.
LiveFront-of-house Agent Factory. Takes a business owner from vision to signed Charter through phase-disciplined facilitation. The Critic stress-tests every Charter against ten documented enterprise-AI failure modes before publication.
Architecture completeRules layer with twelve KAiM-native governance primitives: tier elevation, blast-radius, lexicon binding, expires-review-at, dissent flow. Compiles to mature substrates (CEL design-time, Cedar runtime). IR-first; no custom DSL.
ADR-001 ProposedContinuous integrity layer. Aggregates drift, slop, and behavior signals from the other cells. Root-cause inference is typed as PROPOSAL — never FINDING. Corrective actions route through Codex compilation and Helm enforcement.
Intake brief draftedBot Village (registry), Forge (build), Cobbler (maintenance), and Helm Runtime (governance control plane) complete the operating topology.
Every deployed agent has a Charter — the typed contract that crosses Atelier → Forge → Bot Village registry → Cobbler maintenance. No Charter, no deployment. This is the structural enforcement of "no shadow agents."
The Charter is what 95% of failed GenAI pilots in 2025 did not have. It carries seventeen required fields including:
The Critic — an adversarial agent — stress-tests every Charter against these ten documented failure modes before Provost gate. Override requires T:3 + T:0 co-sign with immutable rationale log.
The framework applied to specific operational scenarios. Each walkthrough demonstrates the same audit-first discipline against a different vertical or function.
Six-stage claim workflow with Helm governance attached at every agent decision. Operator and examiner views, tier coordination, deterministic agent communication bus, GRC framework linkage. Demonstrates CEI, data-quality, bias, HITL, provenance, and audit chain on every privileged action.
Enterprise vocabulary governance. Turn organizational language into testable controls. Material-incident walkthrough shows how a single ambiguous term ("material") becomes a governed canonical entry with classification rules propagating to fleet agents and data-lake content tagging.
SR 11-7 model risk governance applied to agentic AI. Charter discipline for trading desk agents, compliance review automation with deterministic enforcement, audit trail mapped to federal examiner expectations.
KAiM's audit-first framework is designed to satisfy or exceed the standards your compliance officer is already reading.
| Standard | How KAiM aligns |
|---|---|
| NIST AI Risk Management Framework | The four functions (Govern, Map, Measure, Manage) map to Atelier (Map), Codex (Govern), Forensics (Measure), and Cobbler (Manage). Full alignment artifact published. |
| EU AI Act — Article 14 (Human Oversight) | HITL checkpoints are a required Charter field. Compliance is structural at design time, not policy at deploy time. Effective August 2026. |
| ISO/IEC 42001 — AI Management System | The KAiM Ethos document maps directly to the management-system clauses. Decision rituals, change discipline, and external vetting are encoded as operating constraints. |
| OWASP Top 10 for LLM Applications | Excessive Agency (LLM08), Overreliance (LLM09), and Insecure Output Handling (LLM02) are mitigated at design time by Charter discipline and Critic stress-testing. |
| GDPR · CCPA · Quebec Law 25 · NLRB guidance | Structural privacy posture: the Codex Engine refuses to compile rules emitting per-individual scores. KAiM cannot be used for employee surveillance because the architecture forbids it. |
A 60-minute working session with KAiM founder and the architecture team. Review the four cells against one of your live operational scenarios. Walk away with a concrete read on whether the framework fits your governance posture.
No sales script. No demo theater. A working session where your compliance officer, AI lead, and a relevant business owner sit at the same table and pressure-test the framework against a real operational question of yours.
What you bring: one operational scenario where AI agents are or will be making decisions that need governance. What we bring: the framework, an honest read on fit, and a written summary you can take to your board.