Status: implemented
Version: latest
Review: source-backed
Last scanned: 2026-06-25T00:00:00Z
Review required: false
NVIDIA And Nemotron
Nemotron routes, safety review, ModelRouter, ModelDecision proof, and fail-closed model behavior.
NVIDIA role in Trinity
NVIDIA is the model intelligence and safety layer. Trinity routes structured tasks through ModelRouter and stores ModelDecision evidence for scoring, outbound copy safety, classification, deliverable QA, high-risk action review, and sponsor-visible model proof.
The product treats model calls as governed operations. A model response is not enough. The route, provider, schema, decision status, safety outcome, evidence hashes, and links to the relevant run/draft/tool/approval records are what make the result operationally trustworthy.
| Route class | Typical use | Proof stored |
|---|---|---|
| Frugal Nemotron route | Lead scoring, classification, low-cost checks | Route name, model label, schema, status, hash |
| Showcase Nemotron route | High-value planning or final QA | Decision summary and validation state |
| Safety route | Outreach draft risk review and policy checks | Safety category, approval recommendation, fail-closed status |
| Fallback route | Explicit fallback-only behavior when official route is unavailable | Labeled fallback evidence, never live proof |
ModelRouter responsibilities
| Responsibility | Implementation posture | Failure behavior |
|---|---|---|
| Route selection | Prefer configured official NVIDIA/NIM routes for showcase and safety work | Missing official route blocks official-proof claims |
| Schema validation | Require structured output for scoring, safety, and QA tasks | Invalid output becomes failed ModelDecision |
| Provider labeling | Store provider, model, route, and fallback flags | Aggregator fallback is never relabeled as NVIDIA proof |
| Prompt secrecy | Persist summaries and hashes rather than private model context | Public docs and proof never expose hidden reasoning |
| Cost posture | Keep frugal routes for routine checks and showcase routes for high-value decisions | Expensive work remains explicit and auditable |
ModelDecision proof
ModelDecision records are designed for auditability:
- Store provider, route, operation, schema version, status, summary, and hashes.
- Exclude credential values, private customer content, and hidden reasoning.
- Link decisions to run events, drafts, approvals, ToolCalls, and proof packets where available.
- Fail closed when output cannot be parsed, validated, or classified safely.
Safety policy
- Drafts and high-risk copy must pass safety review before Gmail execution.
- Unsupported or malformed model output is not treated as approval.
- OpenRouter or other aggregators are fallback-only and disabled by default.
- Public proof surfaces show summaries and IDs, not private prompt material.
NVIDIA surfaces in Trinity
| Surface | Model purpose | Operator value |
|---|---|---|
| CRM/deal scoring | Score lead fit, confidence, urgency, and next action | Prioritizes real acquisition work |
| Draft safety review | Classify risky copy before Gmail draft/send | Prevents unsafe or non-compliant outreach |
| Deliverable QA | Review generated snapshot deliverables before release | Improves quality of paid output |
| Runtime status | Show configured NVIDIA readiness and route posture | Makes model dependency visible |
| Ledger/proof | Link ModelDecision evidence to runs and artifacts | Supports explainability and review |
Primary source links
Official references
| System | Use in Trinity | Official docs |
|---|---|---|
| NVIDIA NIM | Official model route and API surface | NIM LLM API reference |
| NVIDIA NIM | LLM deployment and route posture | NIM LLM overview |
| NeMo Guardrails | Guardrail posture and rail types | NeMo Guardrails |
Source paths
lib/autonomous_agency/ai/model_router.exlib/autonomous_agency/ai/model_decision.exlib/autonomous_agency/tools/nemotron.exlib/autonomous_agency/email/draft_safety_review.extest/autonomous_agency/email/draft_safety_review_test.exs