47-eva-mti ? EVA Machine Trust Index

EVA Ecosystem Integration

Tool	Purpose	How to Use
37-data-model	Single source of truth for all project entities	GET http://localhost:8010/model/projects/47-eva-mti
29-foundry	Agentic capabilities (search, RAG, eval, observability)	C:\eva-foundry\eva-foundation\29-foundry
48-eva-veritas	Trust score and coverage audit	MCP tool: audit_repo / get_trust_score
07-foundation-layer	Copilot instructions primer + governance templates	MCP tool: apply_primer / audit_project

Agent rule: Query the data model API before reading source files.

Invoke-RestMethod "http://localhost:8010/model/agent-guide"   # complete protocol
Invoke-RestMethod "http://localhost:8010/model/agent-summary" # all layer counts

---

Status: Design complete ? Trust Service specification ready for implementing teams
Owner: AICOE
Created: 2026-02-23 16:33 ET
Maturity: poc

---

What This Project Is

47-eva-mti is the trust computation specification and design authority for the EVA Machine Trust Index (MTI).

It does not ship running code. It ships the complete trust computation design: 6 subscore definitions, computation formulas and weights, trust band mapping, allowed-actions matrix, and the full OpenAPI 3.0 contract for the Trust Service API. Implementing teams consume these specs and build the Trust Service microservice.

Origin: MTI specs were developed as part of 19-ai-gov (EVA AI Governance Plane) and separated into this project on Feb 23, 2026 because trust computation is architecturally distinct from governance policy design. The governance Decision Engine calls the Trust Service at step 5 (resolve_or_compute_trust) via POST /trust/evaluateTrust.

---

Core Concept

Trust is not binary. It is computed.
Six independent domain signals combine into a Composite MTI score (0-100).
The score maps to a Trust Band. The band determines what the actor is allowed to do.

This is policy in motion: instead of static role checks, every actor's autonomy level is continuously evaluated based on evidence from their identity posture, behaviour history, compliance record, evidence quality, security signals, and AI reliability.

---

The 6 Subscores

Subscore	Measures	Primary Signal Sources
ITI -- Identity Trust Index	Authentication strength, MFA, ToU, role maturity	Entra ID, RBAC, ToU logs
BTI -- Behaviour Trust Index	Policy violation history, anomalous patterns, denied actions	audit_events, telemetry (App Insights, APIM)
CTI -- Compliance Trust Index	% of applicable controls satisfied, % obligations fulfilled	governance_evaluations, obligation_instances
ETI -- Evidence Trust Index	Evidence pack completeness, schema compliance, hash validation	evidence_artifacts
STI -- Security Trust Index	Prompt injection signals, data exfiltration indicators, anomaly detection	APIM, App Insights, threat telemetry
ARI -- Agent Reliability Index	Hallucination rate, grounding quality, red team evaluation results	EVA Brain telemetry, 36-red-teaming

Composite MTI = weighted sum of 6 subscores, clamped to [0, 100].

---

Trust Bands

Band	Score Range	Autonomy Level
HIGH TRUST	85-100	Fully autonomous -- all action categories allowed
TRUSTED	70-84	Allowed with monitoring
GUARDED	50-69	Human approval required for sensitive actions
LOW TRUST	30-49	Heavily restricted -- read-only or informational only
UNTRUSTED	0-29	Blocked -- no execution allowed

---

Trust Service API (4 Endpoints)

Endpoint	Purpose
POST /trust/evaluateTrust	Compute all 6 subscores + Composite MTI for a given context envelope
GET /trust/getActorTrust/{actorId}	Retrieve current and historical MTI scores for an actor
POST /trust/getDecision	Execute full 9-step governance Decision Engine (calls governance layer); returns decision + obligations
POST /trust/signal	Ingest a trust signal event (BTI delta, ARI update, ETI change) and trigger async MTI recompute

Note: /getDecision is a convenience endpoint that chains the Trust Service into the Decision Engine. The canonical Decision Engine spec is in 19-ai-gov. The Trust Service is the compute layer; 19-ai-gov is the policy layer.

Full OpenAPI 3.0 contract: eva-mti-trust-service-api.md (reference copy from 19-ai-gov).

---

MTI Computation Model

Formula

CompositeScore = clamp(
  w_iti * ITI + w_bti * BTI + w_cti * CTI +
  w_eti * ETI + w_sti * STI + w_ari * ARI
  - context_risk_adjustment
, 0, 100)

Where weights (w_*) are defined per actor type (HUMAN vs AGENT vs SERVICE vs SYSTEM) in eva-mti-compute-specs.md.

Decay Model

Negative events (violations, anomalies) decay in severity over time:

weight(t_days) = exp(-ln(2) * t_days / half_life_days)
half_life = 14 days (configurable)

This means a violation 28 days ago has 1/4 the impact of the same violation today.

Missing Signal Penalty

If a required signal is unavailable (data source offline, no telemetry), a configurable penalty is subtracted rather than silently defaulting to 0. This prevents trust inflation from data gaps.

---

Source Material

These spec files originated in 19-ai-gov and are the authoritative source for this project:

File	What it defines
eva-mti-scope.md	ITI / BTI / CTI / ETI / STI / ARI -- definitions, data sources, formulas for each subscore
eva-mti-compute-specs.md	Computation spec (YAML) -- inputs, weights, decay, normalization; full actor-type weight tables
eva-mti-actions-matrix.md	Trust band to allowed-action mapping across 9 action categories
EVA-Machine-trust-Index.md	MTI concept, rationale, graduated autonomy model
eva-mti-trust-service-api.md	OpenAPI 3.0 contract for the Trust Service API

Spec quality note: The YAML blocks in eva-mti-compute-specs.md and eva-mti-trust-service-api.md contain   HTML entities. They are reference documents -- not parse-ready YAML. Implementing teams must strip HTML entities before machine processing.

---

Cosmos Container

Container	Purpose	Owner
actor_trust_scores	MTI subscores + composite per actor, with full timestamp history	47-eva-mti

All other governance containers are owned by 19-ai-gov.

---

Relationship to 19-ai-gov

19-ai-gov (Governance Policy)           47-eva-mti (Trust Computation)
      |                                         |
      | Decision Engine                        | Trust Service
      | Steps 1-4: policy, hard-stops          | evaluateTrust: compute 6 subscores
      |                                         |
      | Step 5: POST /trust/evaluateTrust ----> |
      |         <---- TrustEvaluateResponse --- |
      |                                         |
      | Steps 6-11: controls, thresholds,      |
      |             aggregate, obligations,     |
      |             audit, respond              |

The interface is a single HTTP call with a well-defined request/response contract. Neither project has the other as a build dependency at design time.

---

Related Projects

Project	Relationship
19-ai-gov	Governance policy layer; calls Trust Service at Decision Engine step 5
33-eva-brain-v2	Emits BTI and ARI signals via POST /trust/signal; subject to MTI evaluation
36-red-teaming	Produces ARI evaluation results consumed by the Trust Service
29-foundry	Candidate runtime host for the Trust Service API
17-apim	Injects STI signals (prompt injection, anomaly) via trust signal endpoint
40-eva-control-plane	Emits ETI signals when evidence packs are validated
28-rbac	Source of role maturity signals for ITI computation