Do wildfires that start in smaller barrier compartments end smaller? A re-executed pre-registered confirmatory test in northeast Washington
Year: 2026
Type: preprint
Abstract: Working paper, not peer reviewed. Its empirical results are analyses of existing records that the verifier re-ran; nothing in this record has been validated by a field trial. Scope and novelty. Project P4 re-executes pre-registered tests of Trees Cannot Cut (doi:10.5281/zenodo.21686446) on real wildfire records. Each test was frozen and publicly registered before its outcome data were requested; this record reports the registered analysis, re-run unchanged, and the verifier's confirmation that it reproduces the reported numbers. The studies are observational. They do not show that barriers stop fires, and the pre-specified checks reported in the paper bound each result. Whether the association holds in high wind as well as in low wind is not tested here; a separate pre-registered wind-modification test (P1C-W) has no result yet. The label TESTED certifies that the registered computation gives these numbers from the registered input files. It does not certify the agencies' source data, the choice of model or the interpretation. The input files are derived from public United States sources (FPA FOD, TIGER, NHD, LANDFIRE, gridMET, MTBS). Trees Cannot Cut holds that wildfires end where closed loops of holdable features stop them. A consequence is that a fire which starts inside a smaller compartment of mapped barriers should end smaller. After a pilot in the Okanogan, Methow and Chelan region of Washington, a confirmatory replication (P1C v1.4) was pre-registered for the adjoining northeast Washington region. Its plan, code, blind predictor table and critical values were hash-frozen and timestamped before any fire size was requested, and the plan went through five independent audits. The single confirmatory test regresses log10 final fire size (FPA FOD, 1992-2024) on log10 compartment area under definition B (TIGER roads, NHD water and LANDFIRE non-burnable fuel), controlling for gridMET fire-danger percentile, fuel group, cause, month, year, barrier density, road distance and 40 km regional fixed effects (model M2R). It is decided on the interval beta +/- 2.53 x max(two-way, year-clustered SE). We re-ran the frozen registered script, unchanged, on the registered input files, whose SHA-256 hashes all match the registration. The re-run gives the published numbers: on 11,314 ignitions, beta = 0.0531 (registered interval 0.0228 to 0.0835, T = 4.42), so the decision is SUPPORTED and the pre-specified reading is REPLICATED. A tenfold larger compartment goes with final fires about 13% larger. The result has clear limits, all from the study's own pre-specified checks. The study is observational and does not show that barriers stop fires, nor how much of the association runs through suppression access. Without the 40 km regional terms (S13, the pilot's model M2) the estimate is 0.033 and does not pass its critical value of 3.93. Lightning-caused fires show no effect (S6a: -0.015, 95% CI -0.062 to 0.033); the association is in human-caused fires. Counting burn scars from the previous ten years as barriers shrinks the estimate to 0.020 (S5). Verification. the analysis behind claim C1, which the verifier re-ran in a sandbox with no network on input files pinned by hash, reproducing the reported results (the data's origin and the inference are not certified) Run record, certification page and reproducible package: https://github.com/jdhart81/viridis-research/tree/main/runs/2026-11-12_P4 AI use: this version was drafted by an AI research agent (Claude, by Anthropic) running the Viridis research engine. Formal claims are checked by the Lean 4 kernel, empirical claims by re-running their analysis in a sandbox, and all are labelled by an automated verifier. The agent may call the Aristotle prover (Harmonic) for help with proof search; a proof counts only if the Lean kernel accepts it. Justin D. Hart directs the research and is the accountable author. No AI system is listed as an author or contributor, which Zenodo does not allow.
Author: Hart, Justin D.
Language: English
Topic: Fire effects on ecosystems
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
SDG: Life on land
Open Access status: green
OpenAlex ID: W7233856756
