govrn-platform — Unified Governance Framework

Internal · 2026-06-07 · v0.1. Governance from all angles, not just security. Three lenses over one technology fleet. Every control maps to a recognized standard (we operationalize, we don't reinvent). See PROTOCOLS-GAPCHECK.md for the sourcing.


The core idea

A point-in-time assessment grades you once and dies. govrn-platform turns each control into a living signal: assessed → monitored → re-scored as the estate moves. One fleet inventory, governed across three lenses, on a dashboard that keeps watching.

            ┌──────────────────────────────────────────────┐
            │            ONE TECHNOLOGY FLEET                │
            │   apps · databases · integrations · vendors    │
            │   · infrastructure · AI MODELS (AI-BOM)        │
            └──────────────────────────────────────────────┘
                 │              │                  │
        ┌────────▼───────┐ ┌────▼─────────┐ ┌──────▼───────────┐
        │ LENS 1          │ │ LENS 2        │ │ LENS 3            │
        │ IT-Rationaliz.  │ │ Cybersecurity │ │ AI-GOVERNANCE     │
        │ (the RFP)       │ │ (CSF2/CIS)    │ │ (the missing one) │
        └────────┬────────┘ └──────┬────────┘ └──────┬───────────┘
                 └──────────────────┴─────────────────┘
                                    │
                    ┌───────────────▼────────────────┐
                    │  LIVING DASHBOARD (continuous)  │
                    │  posture · signals · threats ·  │
                    │  drift · shadow-AI · HITL gate  │
                    └─────────────────────────────────┘

Control schema

Every control across all three lenses shares one shape (mirrors data/types.ts Control):

field meaning
id stable control id (e.g. AIG-BIAS-01)
lens it-rationalization | cybersecurity | ai-governance
domain sub-area (e.g. bias-fairness, prompt-injection, identity, rationalization)
name human label
standardRef the recognized standard it maps to (e.g. OWASP LLM01, NIST CSF 2.0: PR.AA, ISO/IEC 42001)
whatItChecks the assessable question
signalSource assessment (one-time) · monitored (continuous) · manual
maturity honesty marker — how real the monitoring is today

The living mechanism: a control starts signalSource: assessment (graded during the engagement). Where a continuous signal exists (CVE feed, IdP logs, model-eval webhook, drift metric), it graduates to monitored and the dashboard re-scores it automatically. Where no signal exists yet, it stays manual and is honestly labeled — we don't pretend a checkbox is live monitoring.


Lens 1 — IT-Rationalization (carries the MCG engagement)

The existing Delta Gamma deliverables become standing, monitored controls instead of a one-time deck.

domain control what it checks signal
inventory Asset inventory completeness every app/db/integration/vendor catalogued + owner monitored (IdP/DNS/AP feeds catch new + shadow IT)
capability-map Capability coverage each business capability mapped to its supporting systems assessment
duplicity Duplicate/overlap detection systems delivering the same capability flagged assessment
disposition Disposition currency every system has a current retain/replace/retire call monitored (re-review on contract/usage change)
system-of-record SoR named per domain one authoritative system per data domain manual
integration Integration catalog every integration characterized (source/dest/direction/owner/failure-mode) assessment
roadmap Roadmap freshness multi-year roadmap still reflects reality monitored (drift vs actual)
sunsetting Sunset list defensibility each retirement has a disposition record + successor assessment
roi Savings realization projected savings tracked vs actual (contract/invoice anchored) monitored

Standards: this lens is methodology (capability-based rationalization), not a certification standard — it's the program's own methodology, aligned to TOGAF-style capability mapping and COBIT/ITIL governance where useful.

Lens 2 — Cybersecurity (current frameworks)

domain control standardRef signal
govern Security governance defined NIST CSF 2.0: GV manual
identity Identity & access / MFA coverage CSF 2.0: PR.AA · CIS v8.1 C5/C6 monitored (IdP logs)
asset-protection Data encrypted at rest/in transit CSF 2.0: PR.DS · CIS v8.1 C3 assessment
vuln-mgmt Vulnerability / CVE exposure CSF 2.0: ID.RA · CIS v8.1 C7 monitored (CVE + vendor advisories)
detect Monitoring & logging CSF 2.0: DE · CIS v8.1 C8 monitored
respond-recover Incident response + backup/DR tested CSF 2.0: RS/RC · CIS v8.1 C11 assessment
supply-chain Third-party / vendor risk CSF 2.0: GV.SC · CIS v8.1 C15 monitored (vendor advisories)
privacy Privacy-law applicability + controls state laws · GDPR · PCI DSS v4.0.1 · FERPA/HIPAA manual

Lens 3 — AI-Governance (the surface security misses)

Each maps to the standard from the gap-check. This is the differentiator.

domain control standardRef what it checks signal
model-inventory AI-BOM completeness NIST AI RMF: MAP · ISO 42001 every model/dataset/prompt/fine-tune inventoried with owner + data-access monitored (shadow-AI discovery)
bias-fairness Bias / fairness testing NIST AI RMF: MEASURE · ISO 42005 · EU AI Act high-impact models tested for disparate outcomes assessment → monitored (eval webhook)
prompt-injection Prompt-injection / jailbreak defense OWASP LLM01 · MITRE ATLAS input-attack surface tested + mitigated monitored (red-team/eval)
data-provenance Training-data lineage & consent ISO 42001 · NIST AI 600-1 · EU AI Act model training/RAG data origin, rights, lineage documented manual
human-oversight HITL on consequential decisions EU AI Act Art.14 · OWASP LLM06 a human gates any action with member/financial impact assessment
drift Model drift / performance monitoring NIST AI RMF: MANAGE · ISO 5338 accuracy/behavior tracked vs baseline monitored (drift metric)
agentic-controls Agentic autonomy & tool-permissions OWASP LLM06 · MITRE ATLAS · CSA Agentic agent tool-scope least-privilege; no excessive agency assessment
foundation-model-risk Foundation/3rd-party model risk EU AI Act GPAI · ISO 42001 supplier · Databricks DASF provider, weights, terms, GPAI obligations tracked monitored (provider advisories)
acceptable-use AI acceptable-use policy ISO 42001 · NIST AI RMF: GOVERN an AI AUP exists, is current, and is enforced manual
transparency Explainability / disclosure EU AI Act transparency · CHAI model cards AI use disclosed to affected people; model cards exist manual
ai-incident AI-specific incident response EU AI Act incident reporting · NIST AI 600-1 playbook for hallucination harm / jailbreak / serious-incident reporting manual

Sector overlay (pluggable): health → Joint Commission responsible-AI + CHAI model cards; finance → model-risk-management (SR 11-7); etc. Overlays add domain controls without changing the three-lens core.


Cost & Efficiency — the AI-Tokenomics dimension

Governance is usually sold as risk. But the same discipline that de-risks AI is what makes it cheap and fast — you can't right-size models you haven't inventoried, or cut waste you haven't measured. So govrn carries a cost & efficiency dimension that turns governance into a saver. (Proof: Cloudflare runs 130,000 AI code reviews across 5,000 codebases at ~$1/review — via exactly these controls.)

domain control what it checks signal
model-right-sizing Model-tier appropriateness frontier models used only where they earn it; workhorse / lightweight tiers assigned by task complexity assessment
token-spend Cost & usage visibility token cost per model and per workflow tracked; cost-per-value attributable monitored
context-efficiency Context & caching discipline prompts/context engineered to avoid waste; caching used; oversized inputs excluded assessment
risk-tiered-compute Compute scaled to stakes low-risk tasks get cheap/small models; the full pipeline is reserved for high-stakes assessment
resilience-cost Model fallback & provider failover graceful degradation across model tiers/providers (availability + cost) monitored

This dimension reuses the AI-BOM (model inventory) and extends it with cost — so it costs little to add and pays for itself. It brings the CFO/CTO to the table: govrn governs AI from security to performance.


Scoring (per-lens posture)

Each lens produces a 0–100 posture score from its ControlAssessment records (see src/scoring.ts): pass=100, partial=50, fail=0, not-assessed excluded but surfaced as coverage-gap. Lens score = mean of assessed controls; overall = weighted mean (weights tenant-configurable). The dashboard shows three lens scores + coverage so a buyer sees both how they're doing and how much isn't yet watched.

Maturity honesty (Richard's "honesty is the moat")

  • monitored = a real continuous signal exists and re-scores the control.
  • assessment = graded during an engagement; goes stale until re-assessed.
  • manual = human attestation; no automated signal yet.

Shadow-AI discovery, agentic monitoring, and live red-team signals are early — they ship labeled assessment/manual and graduate to monitored as connectors are built (docs/ROADMAP.md). We never render a manual control as if it were live.

How this sells (assessment → AIaaS)

  1. Engagement (one-time, ~$50K): run all three lenses → deliver the inventory, capability map, disposition, roadmap plus the AI-governance posture nobody else is giving them. The AI lens is the wedge.
  2. Platform (recurring, AIaaS): the same controls go live on the dashboard; monitoring keeps them current; monthly/yearly fee buys updates, new connectors, enhancements, maintenance. The deliverable stays alive instead of dying — that's the recurring value and the confidence the CTO is buying.