Adam Governance For AI Workloads

Use Adam governance when model, agent, or tool traffic must be checked before it reaches a model provider, saved template, or endpoint tool.

Overview

Adam uses Eve as the governed request path. Applications can send OpenAI-compatible model traffic through Eve, and operators can apply reusable policies for PII, restricted content, route boundaries, promptless audit evidence, and cost attribution.

The same controls support regulated workflows such as sensitive document summarization, customer operations, internal copilots, and tool-using agents. Domain-specific behavior is expressed as configuration; the public product surface stays generic.

Application or agent
    |
    v
Eve LLM gateway key
    |
    +-- Built-in PII detection
    +-- Organization and key dictionaries
    +-- Template-backed PII lookup
    +-- Content policy profiles
    +-- Policy bindings
    +-- Access boundaries
    +-- Feature audit and promptless evidence
    |
    v
Model provider, saved template, or tool

Core Controls

Sensitive Document Workflow Example

An application can submit a document packet to an LLM while masking customer and account data found in internal systems.

Configure a gateway key with the existing PII controls plus a template lookup:

{
  "name": "document-summary-gateway",
  "endpoint_id": "local-ollama",
  "custom_pii_terms": [
    {
      "value": "ESC-4419",
      "label": "internal_case_id",
      "action": "redact"
    }
  ],
  "pii_lookup_templates": [
    {
      "template_id": "lookup-customer-account-pii",
      "input_fields": [
        {
          "name": "account_id",
          "source": "user_message_regex",
          "pattern": "(?i)account(?: id)?[:# ]+(?P<value>[A-Z0-9-]+)",
          "required": true,
          "value_type": "string"
        }
      ],
      "records_path": "",
      "redaction_fields": [
        {
          "path": "/full_name",
          "label": "customer",
          "action": "redact"
        },
        {
          "path": "/email",
          "label": "email",
          "action": "redact"
        },
        {
          "path": "/phone",
          "label": "phone",
          "action": "redact"
        }
      ],
      "max_rows": 1,
      "error_mode": "block"
    }
  ],
  "work_metadata": {
    "workflow": "document_summary",
    "department": "operations"
  }
}

Expected behavior:

• Eve extracts account_id from the original client request.
• Eve runs the saved Read template in sensitive execution mode.
• Template output values become request-scoped PII terms.
• Built-in PII, organization dictionaries, key custom terms, and lookup terms merge together.
• The model provider receives only the masked request.
• Telemetry records lookup status, counts, safe identifiers, route, model, and work metadata. It does not record prompt text or raw lookup values.

Restricted Project Controls Example

A team can separate access to restricted projects without copying sensitive terms into every key.

1. Create auth groups for desks and approved agents.
2. Create content profiles for restricted project names, internal identifiers, credential patterns, and payment instructions.
3. Bind profiles to auth groups, users, agents, routes, gateway keys, or work metadata.
4. Add access boundaries that block disallowed providers, model routes, templates, tools, or endpoints.
5. Use governance feature records for promptless simulations, approvals, evidence packs, entitlement sync jobs, model evaluations, data quality checks, dashboard views, or DLP scans.

Example content policy profile:

{
  "profile_key": "restricted_project_terms",
  "display_name": "Restricted project terms",
  "terms": [
    {
      "value": "Project Atlas",
      "pattern_kind": "literal",
      "label": "restricted_project",
      "action": "redact"
    },
    {
      "value": "PROJECT-[0-9]{4}",
      "pattern_kind": "regex",
      "label": "internal_project_identifier",
      "action": "block"
    }
  ],
  "enabled": true
}

Example policy binding:

{
  "binding": {
    "policy_kind": "content_policy_profile",
    "policy_key": "restricted_project_terms",
    "target_kind": "auth_group",
    "target_uuid": "33333333-3333-4333-8333-333333333333",
    "enabled": true
  }
}

Example access boundary:

{
  "boundary": {
    "boundary_key": "deny-restricted-project-to-public-model",
    "subject_selectors": {
      "auth_group_uuid": "33333333-3333-4333-8333-333333333333"
    },
    "resource_selectors": {
      "provider": "PublicProvider",
      "route_class": "external"
    },
    "action": "block",
    "enabled": true
  }
}

Promptless Evidence

Governance exports, lineage events, replay jobs, and feature records accept only safe evidence. Use UUIDs, hashes, trace identifiers, counts, route names, model names, provider names, status codes, and decision labels.

Do not store:

• prompt text,
• completion text,
• raw documents,
• raw query bodies,
• raw lookup values,
• matched PII terms,
• credentials,
• or unredacted user input.

Example policy simulation record:

{
  "record_key": "restricted-route-check-2026-07",
  "subject_kind": "agent",
  "subject_uuid": "22222222-2222-4222-8222-222222222222",
  "resource_kind": "llm_provider",
  "spec_json": {
    "route_class": "external",
    "provider": "PublicProvider",
    "work_metadata": {
      "project_class": "restricted"
    }
  }
}

Response shape:

{
  "record_uuid": "aaaaaaaa-aaaa-4aaa-8aaa-aaaaaaaaaaaa",
  "feature_kind": "policy_simulation",
  "status": "completed",
  "result_json": {
    "simulation_version": "v1",
    "promptless": true,
    "safe_metadata_only": true,
    "access_boundaries": {
      "decision": "block",
      "block_count": 1
    }
  }
}

Safety Guarantees

• Template-backed PII lookup is additive; it does not replace built-in detection or dictionaries.
• Lookup-enabled keys require a valid owning agent identity and read access to the referenced template.
• Saved lookup templates must be Read templates.
• Sensitive template execution suppresses rendered request and response payload logging.
• Standalone gateway snapshots can enforce hydrated content policies and access boundaries, but service-only template lookup fails closed when configured outside Eve service.
• Stale selectors, malformed persisted policy, missing lookup inputs, revoked template access, and unsupported actions fail closed according to the relevant policy mode.

Validation Artifacts

Adam governance validation covers regulated AI workloads with synthetic data. Validation artifacts record before/after masked requests, content profiles, policy bindings, access boundaries, template lookup summaries, feature audit rows, exports, lineage, replay artifacts, and governance feature records.

Validation artifacts are deterministic and intended to prove product behavior: policy configuration, extracted inputs, matched safe categories, policy actions, route decisions, approval status, and promptless evidence. They should not be treated as production customer data.

Related Docs

• Capabilities By Product
• LLM Endpoint
• LLM And Agent APIs
• Adam Model Routing And Orchestration
• Adam Tool Calling
• Eve Access And Policy
• Analytics And Observability APIs