Helm Meaning Layer turns enterprise vocabulary into testable controls, governed publication, and audit-ready evidence. Across models, agents, communications, and data lake content.
Three vocabulary failures are already happening in your enterprise. None of them require monitoring a single employee.
Your standards define terms one way. Your people, agents, and data lake classify them another way. Most enterprises have 50 to 90 percent drift between decreed and observed meaning on at least one term that matters.
Sales calls them "partners." Finance calls them "accounts." Customer Success calls them "customers." Your enterprise-wide AI summarizer picks one and confuses everyone, or invents a fourth term.
Your LLM agents are inventing terms, redefining policy language, and using vocabulary that is not in any glossary. Agents then quote each other's invented terms and the meaning silently rots.
Helm Meaning Layer implements six operating-model functions, each producing audit-ready artifacts the next consumes. Frontier model assists; human stewards decide; the audit log explains every decision.
Approved corpus ──▶ Harvester ──▶ Term Miner ──▶ Decoder ──▶ Ontologist
(policies, │
transcripts, ▼
agent logs) Steward
(human-in-loop)
│
┌──────────────────┴──────────────────┐
▼ ▼
Governed lexicon Data-lake classification
(canonical terms, rules (sensitivity,
ontology, folksonomy) retention, access)
| Function | Operating job | Audit artifact produced |
|---|---|---|
| Policy and standards | Authoritative definitions ingested verbatim from compliance and governance documents. | canonical-term.json (origin: decreed) |
| Risk classification | Sensitivity, tone, regulatory relevance tagged at the term level. | risk_tags, security_class fields |
| Control design | Per-consumer publication scope: which fleet agents and lake partitions receive each term. | review-packet-publication.json |
| Evidence capture | Append-only event log; every Steward decision recorded with channel and authority tier. | lexicon-event.json (JSONL) |
| Monitoring and exceptions | Reconciler surfaces decreed-versus-observed drift; flags policy non-compliance without altering the lexicon. | governance/decisions/ findings |
| Audit and board reporting | Full derivation chains from data-lake tag → classification rule → term version → Steward approval → evidence. | derived_from lineage, exportable |
Helm Meaning Layer serves several roles simultaneously. Each receives audit-ready outputs tuned to their accountabilities.
Quarterly governance outlook on AI vocabulary drift, Steward approval rates, and policy-versus-practice gaps that have escalated.
View board-reporting model →Real-time view of fleet vocabulary, agent-invented term audit, autophagy detection on canonical entries.
View CAIO operating model →Cross-business-unit term inconsistency surfaced as disputed, never silently reconciled. Data-lake classification rules derived from approved vocabulary.
View CDO operating model →Every classification rule traces back to a human-approved term version. Defensible derivation chains for regulator review.
View compliance operating model →Append-only event log of every state transition. Steward approval rate, non-uptake aggregations, exception tracking.
View audit operating model →Schema-validated context packs for fleet agents. JSON Schema 2020-12 contracts. Validation harness exit-code-1 on regressions.
View platform operating model →Helm Meaning Layer's design and operation are aligned to the following frameworks. Mappings are auditable and traceable, not marketing claims.
| Framework | Mapped functions | Status |
|---|---|---|
| NIST AI RMF | Govern, Map, Measure, Manage as lifecycle disciplines across the five-agent cell. | Mapped |
| ISO/IEC 42001 | AI management system controls; documented operating model, evidence capture, continual improvement. | Mapped |
| OWASP LLM Top 10 | Autophagy mitigation (LLM03), confidence inflation (LLM09), insecure output handling (LLM02), excessive agency (LLM08). | Validated |
| SOC 2 Trust Services | Security and confidentiality criteria; audit trail, access control. | In Progress |
| Model Risk Management (SR 11-7) | Conceptual soundness, ongoing monitoring, outcomes analysis applied to LLM-assisted decoder. | Draft |
Fixed-fee engagement against one approved corpus. Five audit-ready artifacts delivered. You retain everything produced, regardless of continuation.
| Scope | One approved corpus: policy library, call or meeting transcripts, existing AI agent prompts and outputs, or governance decision records. |
| Runtime | HELM (Herb Brain + Bot Village cell). Included. No infrastructure spend on your side. |
| Approvals | Your Stewards approve every canonical promotion. We provide the workflow; you provide the decisions and authority. |
| Exit | You retain the lexicon, drift report, agent audit, classification rules, and full audit trail. Standard JSON export. |
| Security | Your AWS region. Encrypted at rest and in transit. SOC 2 in progress; pilot scope designed accordingly. No broad employee monitoring. |