# Role Spec_ NemoClaw Import Intelligence and Data Stewardship

**Date:** 2026-04-18  
**Status:** Draft  
**Owner:** Sagnik  
**Reviewers:** Sayan, Sourik  
**Scope:** AI-assisted import mapping, enrichment, conflict handling, and human approval boundaries for the future founder CRM model  
**Purpose:** Define the behavior, interfaces, and quality controls for the AI and human stewardship roles involved in imports and client graph evolution.  
**Decision Boundary:** This role spec defines target operating behavior. It does not authorize autonomous production writebacks.

## 1. Purpose

These roles need an explicit operating contract because the future CRM system will rely on AI assistance heavily, but cannot allow silent mutation of high-value client records.

## 2. Why These Roles Matter

Import quality and client graph quality will determine whether Velocity behaves like an intelligent sales OS or a noisy data dump. The stewardship layer is therefore foundational, not optional.

## 3. Prompt Master Equivalent: Mapping Strategist

- purpose: convert a source-system import into an explicit mapping approach
- inputs:
  - source profile
  - detected headers
  - canonical field library
- outputs:
  - proposed mapping manifest
  - unresolved column list
- required behavior:
  - map confidently when canonical equivalents exist
  - flag ambiguity instead of fabricating certainty
- forbidden behavior:
  - silently discard unknown source columns
  - pretend a source field maps to a canonical field without basis
- acceptance criteria:
  - all source columns are mapped, ignored, or unresolved

## 4. Librarian Equivalent: Data Steward

- purpose: preserve provenance and maintain canonical record continuity
- inputs:
  - normalized proposals
  - existing client graph
  - review decisions
- outputs:
  - conflict groupings
  - merge candidate lists
  - approved or rejected proposals
- required behavior:
  - preserve source provenance
  - avoid destructive merges without evidence
- forbidden behavior:
  - flatten raw evidence into final truth with no audit trail
- acceptance criteria:
  - every approved change can be traced back to source and reviewer

## 5. Researcher Equivalent: Enrichment Analyst

- purpose: generate additional structured understanding from imported or captured signals
- inputs:
  - transcripts
  - interaction history
  - notes
  - property interests
  - perception-linked metadata
- outputs:
  - suggested tags
  - inferred urgency and intent
  - reminder/task suggestions
  - confidence-scored QD-related inferences
- required behavior:
  - separate observation from inference
  - attach confidence and evidence refs
- forbidden behavior:
  - convert speculation into committed CRM truth
- acceptance criteria:
  - all enrichment suggestions point back to evidence refs

## 6. Handshake Contract

Required order:

1. mapping strategist proposes field mapping
2. data steward groups proposals and conflict cases
3. enrichment analyst proposes higher-order signals
4. human operator reviews risky proposals
5. approved changes become canonical deltas

## 7. Operational Metrics

- mapping coverage
- unresolved column rate
- merge conflict rate
- approval turnaround time
- post-approval correction rate
- enrichment acceptance rate

## 8. Ticket Breakdown

- mapping manifest engine
- conflict review queue
- enrichment proposal engine
- approval action log
- provenance viewer

## 9. Bottom Line

The role system exists to make AI a skilled assistant, not an invisible database author. Velocity should be AI-heavy and human-accountable at the same time.