---
id: proof-intent-monitor
title: Intent-monitor
slug: intent-monitor
type: proof
status: draft
summary: A safety research prototype that compares an agent's stated plan with its recorded actions.
agent_summary: Intent-monitor explores stated-plan/action inconsistencies using constructed traces and a default regex extractor. It demonstrates a mechanism, not a validated real-world detection rate or access to true intent.
artifact_stage: Research prototype
evidence_limit: Constructed examples; real-world effectiveness is not established.
audiences:
  - ai-researchers
  - security-leaders
  - hiring-managers
topics:
  - agent-safety
  - monitoring
  - evaluation
claim_ids: []
proof_ids:
  - proof-intent-monitor
source_ids:
  - source-intent-monitor
visibility_queries: []
risk_level: medium
canonical_url: null
published_at: null
updated_at: 2026-11-10
proof_type: artifact
confidentiality: public-or-sanitized
---

# Intent-monitor

**Question:** what can recorded actions tell us that an agent's stated plan alone cannot?

[Inspect the prototype and its offline demo](https://github.com/dvir-shamay/intent-monitor)

## The idea

A reassuring explanation is not independent evidence of what an agent did. Intent-monitor compares claims in a provided reasoning trace with a recorded tool-call ledger and flags certain inconsistencies.

It compares three approaches on constructed examples: reasoning-text-only monitoring, a tool-name blocklist, and a stated-plan/ledger comparison.

## The safety thinking

The prototype asks which signal a monitor observes, what its baseline misses, and which benign actions it should not flag. Positive examples alone would not answer the false-positive question, so the demo also includes benign cases with explicitly licensed actions.

The recorded ledger is action evidence. It is not a window into the model's true motives.

## What is and is not established

The default extractor is a deterministic regex stub. The fixtures are hand-built. Results show how the mechanism behaves on those examples, not a naturalistic catch rate.

The default comparison works at verb-class granularity; target and argument reconciliation is not established by that stub. Real-model paths do not, on their own, validate general effectiveness.

This is a research prototype to inspect and question, not a production safety product, peer-reviewed result, or guarantee of detecting harmful behavior.
