Two Engines, One Plan: OpenTofu and Terraform Produce Identical Plans and Reproducible Rebuilds, While a Refresh Plan Misses Every Permission Change and Every Unmanaged File

Infrastructure code for AI platforms is usually reviewed by reading it, not by measuring what it does. This study pre-registered four hypotheses about one fixed local AI platform stack (13 local_file resources, hashicorp/local provider 2.5.2) and measured them at zero spend, with no cloud account. H1, plan equivalence, is Supported: OpenTofu v1.13.9 and Terraform v1.17.1 produce the same normalised plan, 13 of 13 addresses matching on type and planned actions, a match rate of 1.0000. H2, rebuild reproducibility, is Supported: over 20 destroy and apply cycles per engine, every run reproduces the run 1 normalised plan hash and state hash for its engine, 20 of 20 for each engine, and the recorded run 1 plan and state hashes are identical across the two engines. H3, drift detection recall, is Partially supported: a refresh plan detects 38 of 54 seeded out-of-band changes, a recall of 0.7037, and 0.8261 over managed file mutations only. Every miss falls in two pre-registered classes: all 8 permission changes and all 8 unmanaged file additions are missed, while every content edit, append, delete and truncate is detected. A naive file-exists baseline reaches only 0.1667. H4, policy precision and recall, is Supported: five Rego deny rules run with conftest 0.71.1 against a labelled corpus of 60 plan JSON files reach precision 1.0000 and recall 0.9333 (28 true positives, 0 false positives, 30 true negatives, 2 false negatives), and the 2 false negatives are exactly the two forms the pre-registration said the policy cannot parse. The claims are deliberately narrow: one provider, one machine class, local files only. Within those bounds, every threshold was frozen before any experiment code was written, and every number below is recomputed from the recorded trial, run and file records.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24328126
Primary Topic
Software System Performance and Reliability
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Two Engines, One Plan: OpenTofu and Terraform Produce Identical Plans and Reproducible Rebuilds, While a Refresh Plan Misses Every Permission Change and Every Unmanaged File

Jesun Ahmad Ushno
3 links
Zenodo (CERN European Organization for Nuclear Research)
Software System Performance and Reliability
preprint

Two Engines, One Plan: OpenTofu and Terraform Produce Identical Plans and Reproducible Rebuilds, While a Refresh Plan Misses Every Permission Change and Every Unmanaged File

Jesun Ahmad Ushno
preprint en

Abstract

Infrastructure code for AI platforms is usually reviewed by reading it, not by measuring what it does. This study pre-registered four hypotheses about one fixed local AI platform stack (13 local_file resources, hashicorp/local provider 2.5.2) and measured them at zero spend, with no cloud account. H1, plan equivalence, is Supported: OpenTofu v1.13.9 and Terraform v1.17.1 produce the same normalised plan, 13 of 13 addresses matching on type and planned actions, a match rate of 1.0000. H2, rebuild reproducibility, is Supported: over 20 destroy and apply cycles per engine, every run reproduces the run 1 normalised plan hash and state hash for its engine, 20 of 20 for each engine, and the recorded run 1 plan and state hashes are identical across the two engines. H3, drift detection recall, is Partially supported: a refresh plan detects 38 of 54 seeded out-of-band changes, a recall of 0.7037, and 0.8261 over managed file mutations only. Every miss falls in two pre-registered classes: all 8 permission changes and all 8 unmanaged file additions are missed, while every content edit, append, delete and truncate is detected. A naive file-exists baseline reaches only 0.1667. H4, policy precision and recall, is Supported: five Rego deny rules run with conftest 0.71.1 against a labelled corpus of 60 plan JSON files reach precision 1.0000 and recall 0.9333 (28 true positives, 0 false positives, 30 true negatives, 2 false negatives), and the 2 false negatives are exactly the two forms the pre-registration said the policy cannot parse. The claims are deliberately narrow: one provider, one machine class, local files only. Within those bounds, every threshold was frozen before any experiment code was written, and every number below is recomputed from the recorded trial, run and file records.

Zenodo (CERN European Organization for Nuclear Research)
Software System Performance and Reliability
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