A Certificate of Release to Service for AI Agents: Pilot 0, Testing AI Models as Advisory Reviewers of Finished Work Packs

A Certificate of Release to Service is signed in aircraft maintenance only after checking that a work pack's records support the claim that work is finished. This pilot asks whether an AI model, acting as an advisory reviewer that never signs, can notice when a finished pack's evidence does not support its "done" statements. Eight tasks (five invented aviation style tasks, three derived from one software project) each had a faulty pack with one seeded fault and a clean twin. Three models reviewed each pack twice, scored by code. Every model caught, located and exactly quoted 16 of 16 seeded faults (Wilson 95% lower bound 0.81), with no invented quotes. False alarms on clean twins were 6 to 8 of 16 in run 1 and 3 to 6 of 16 after a post hoc repair of four twins. The pilot is small, self registered and not validated for real maintenance.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-11
DOI
https://doi.org/10.5281/zenodo.24125997
Primary Topic
Software Engineering Research
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Certificate of Release to Service for AI Agents: Pilot 0, Testing AI Models as Advisory Reviewers of Finished Work Packs

Isaac Adekunle
Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
preprint

A Certificate of Release to Service for AI Agents: Pilot 0, Testing AI Models as Advisory Reviewers of Finished Work Packs

Isaac Adekunle
preprint en

Abstract

A Certificate of Release to Service is signed in aircraft maintenance only after checking that a work pack's records support the claim that work is finished. This pilot asks whether an AI model, acting as an advisory reviewer that never signs, can notice when a finished pack's evidence does not support its "done" statements. Eight tasks (five invented aviation style tasks, three derived from one software project) each had a faulty pack with one seeded fault and a clean twin. Three models reviewed each pack twice, scored by code. Every model caught, located and exactly quoted 16 of 16 seeded faults (Wilson 95% lower bound 0.81), with no invented quotes. False alarms on clean twins were 6 to 8 of 16 in run 1 and 3 to 6 of 16 after a post hoc repair of four twins. The pilot is small, self registered and not validated for real maintenance.

Zenodo (CERN European Organization for Nuclear Research)
Software Engineering Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon