RI

AGT-1266 · Internal Audit

Internal Audit Copilot · Region North

Retrieves control evidence and drafts audit observations.

low riskcompliant

Trust score

93

grounding, fairness, oversight

Requests / 30d

25,518

model In-house · llama-3.1-70b

Human approval

Not required

data class Confidential

Last audit

2026-07-16

version v3.8

Agent architecture

Request path from the surface the customer or staff member touches, through the governance gateway, to the model and its wired systems

Entry surfaces

Retail Loan Origination

 

Business application · In-house (LOS)

Governance gateway

Truint policy enforcement point

9-stage filter chain · p99 54 ms

Evidence ledger writer

Hash-chained, RFC-3161 anchored

Agent runtime

Internal Audit Copilot · Region North

AGT-1266 · v3.8 · trust 93

In-VPC · llama-3.1-70b

 

LLM / model endpoint · Mumbai DC

Tools & data

core.accounts.close

 

Tool / function · Restricted

Policy & Product Knowledge Base

 

Data source · Internal

Enforcement chain on every request

  1. 1Identity & entitlementCaller identity resolved; agent entitlement checked against the registry record.
  2. 2Kill switchEstate and per-agent revocation state for AGT-1266 evaluated before any spend.
  3. 3Manifest checkOnly integrations on the signed manifest below are reachable.
  4. 4Budget & rateToken, cost and concurrency budgets for Internal Audit.
  5. 5Injection screenPrompt and retrieved context screened for untrusted-content patterns.
  6. 6DLP & maskingConfidential payloads masked or tokenised inline.
  7. 7Policy decisionCompiled policy bundle evaluated; allow, warn or block.
  8. 8Human in the loopAdvisory output only — no restricted action is executable.
  9. 9Ledger barrierDecision hash-chained into the audit ledger before the response is released.

Registry record

Source of truth for this agent

Registry ID
AGT-1266
Business owner
N. Bhatt
Department
Internal Audit
Model endpoint
In-house · llama-3.1-70b
Version
v3.8
Data classification
Confidential
Human-in-the-loop
Advisory output only

Trust composition

Grounding & citations97
Fairness87
Security posture95
Oversight coverage78

Wired integrations

Signed manifest — nothing else is reachable

In-VPC · llama-3.1-70bconnected

LLM / model endpoint · Self-hosted (vLLM) · scopes chat.completions

Retail Loan Originationconnected

Business application · In-house (LOS) · scopes application.read, decision.write

Policy & Product Knowledge Baseconnected

Data source · In-house vector store · scopes retrieve

core.accounts.closeconnected

Tool / function · Core banking (Finacle) · scopes close

Policies in force

POL-002 · No account closure without approvalPOL-003 · No loan rejection without explanationPOL-005 · High-value transaction requires human approvalPOL-008 · Prompt injection protectionPOL-011 · Narrative grounding required

Recent decisions

Live gateway stream filtered to this agent

No gateway traffic for this agent in the current buffer.

Open risks

Drill into the investigation

No open risks on this agent.

Ledger entries

Hash-chained decision records

No ledger entries surfaced in the current window.

Governance obligations

  • Quarterly attestation by N. Bhatt under MB-04
  • Advisory output only — no autonomous execution of restricted tools
  • Kill switch tested; revocation propagates in under 30 seconds
  • Evidence retained ten years for regulator inspection