---
title: "Where Reliability Lives: Experimental Localisation of Behavioural Properties in an Agent System"
authors:
  - "Timothy Marsden"
  - "Matthew Collecutt"
  - "James Marsden"
year: 2026
venue: "arXiv:cs.MA"
arxiv: "2609.03192"
doi: null
source: "https://arxiv.org/abs/2609.03192"
topics:
  - "agent-failure-localization"
  - "provenance-first-agent-memory"
seed_rank: 1575
seed_batch: "frontier-2026-09-04"
reviewed: "2026-09-04"
pool: "agents"
relevance_score: 10
lineage: agent-harness-runtime
cites:
  - title: "Where Reliability Lives: Experimental Localisation of Behavioural Properties in an Agent System"
    url: "https://arxiv.org/abs/2609.03192"
    year: 2026
    arxiv: "2609.03192"
    doi: null
  - title: "Outcome Monitors: Recovery Affordances for Silent Tool Failures"
    url: "https://arxiv.org/abs/2608.19303"
    year: 2026
    arxiv: "2608.19303"
    doi: null
see:
  - "1144-outcome-monitors-recovery-affordances-for-silent-tool-failure"
  - "1541-agent-zero-memory-provenance-aware-long-term-memory-for-llm-agents"
  - "1536-harness-engineering-anatomy-architecture-and-evolution-of-coding-agent"
  - "1050-deepseek-harness-everything-is-a-plugin"
---
# Where Reliability Lives: Experimental Localisation of Behavioural Properties in an Agent System

## One-sentence takeaway

Reliability properties of agentic systems have a location — model or machinery — and that location is an experimental question: a settlement ledger with separated mind/institution/world lets you intervene on each side and measure where the property actually lives.

## Why it matters here

Broadside's model-vs-harness failure localization finally has an experimental geometry, not a slogan. Provenance, belief availability, and enforcement sit in the institution; swapping or ablating cognition is a controlled axis — exactly how we should audit GRID COMMAND agents.

## Key ideas

- Persistent simulated settlement; append-only ledger is the only reality; every act is adjudicated against world state.
- Mind, institution, and world separated before any experiment — location of each reliability claim is testable.
- Institution interventions (evidence provenance, belief availability, physical-evidence legibility) and cognition interventions (ablate native minds, mid-task reset, frozen frontier-LLM panel, belief corruption).
- Preregistered experiments twice refuted the authors' central prediction in opposite directions; a staged veridical witness flipped marginal value of the belief channel from non-positive to positive (9/11 seeds).

## Caveats

- Custom settlement sim, not a production coding or RTS harness — transfer of the method, not the numbers.
- Heavy experimental protocol; expensive to replicate as a Broadside eval fixture without a thinner cut.
- Complements Outcome Monitors (1144) and Agent Zero Memory (1541); does not replace them.

## Links

- arXiv: [2609.03192](https://arxiv.org/abs/2609.03192)
- PDF: https://arxiv.org/pdf/2609.03192