Cohler and Soon audit evidence and reproducibility

This supplement supports Haivian’s audit of Mathematical Falsification of the Fossil-Dominant Hypothesis for Rising Atmospheric CO2 Using the Keeling Relation, DOI 10.53234/scc202612/05. The scientific work was completed on 6 November 2026. That is an audit date, not a publication date.

Finding: the claimed general mathematical falsification is not established. The observed modern intercept and several conditional calculations reproduce; interpreting them as universal attribution bounds requires additional assumptions. This audit does not measure actual fossil attribution, validate a competing climate model or certify every official estimate.

Read the validation article, the report PDF, or download the reproducibility supplement. The download includes this account, the report text, scripts, saved aggregate/synthetic results and review records. Its file manifest identifies the included evidence bytes; third-party inputs are linked separately.

Inspect the evidence

• Technical report: exact balances, assumptions, counterexamples, empirical comparisons, nine findings and limitations.
• Review reconciliation: separate first reviews, prior exposure, targeted addendum and challenge corrections.
• Source and claim ledger: primary-source locators and access limits.
• External input manifest: exact input filenames, source URLs, versions, retrieval times where recorded and SHA256 checksums. Third-party files are not bundled.
• Public-copy provenance: original and sanitized-copy identities. Original record hashes identify the retained scientific history; they are not expected to equal sanitized public-copy hashes.
• Rights and exclusions: authorized authored material, third-party references and excluded operational records.

The following are saved outputs of the completed audit, repackaged without another scientific run:

The mathematical and challenge examples are deliberately constructed logical tests. The synthetic CSVs in analysis/math-review/ contain no atmospheric observations. Aggregate fit results are our calculations from identified third-party inputs; raw or paired observational series are not distributed here.

Reproduce the calculations

The original environment was Python 3.12.14 with NumPy 2.3.5 and openpyxl 3.1.5 on Windows. The mathematical script requires only the standard library; the other scripts require the indicated packages. No interpreter, library binary or third-party dependency is distributed. The original scripts are byte-identical copies, with relative project paths. Their source/result bindings are recorded in the saved JSON and provenance record.

1. Work in a separate copy of this supplement, preserving the supplied results. The analysis scripts write their result subdirectories and generated CSVs.
2. Retrieve the 17 listed inputs from their primary providers into the paths in sources/SOURCE-INPUTS.json, under those exact filenames. For the GCB workbook, use the linked ICOS metadata page and its public license/download process. An HTML license page is not the XLSX workbook. Respect provider terms.
3. Compare every input with its recorded SHA256 using python -B verify_inputs.py. If a provider now serves revised bytes, this is a new-vintage analysis rather than exact input replay. The supplement does not promise that historical bytes remain publicly obtainable. Do not silently substitute them or claim a matching replay.
4. From the supplement root, run the scripts in this order:

python -B analysis/math-review/verify_math.py
python -B analysis/scripps_reanalysis.py
python -B analysis/extended_reanalysis.py
python -B analysis/final_checks.py
python -B analysis/challenge_checks.py

These are instructions for external reproducibility, not a claim of a new publication-stage run. The verify_inputs.py helper checks file identity only; it does not establish scientific validity. Scripts 2 and 3 regenerate the observational CSVs needed by later checks. The challenge script also reads the supplied sanitized reviews and the three primary PDFs identified by the input manifest. Its new provenance hashes will identify those public review copies rather than the private originals; the mathematical assertions are unaffected.

Assumptions and limits that must travel with the results

Station fits. CO₂ and isotope values are paired by station and month; annual means require nine paired months. Requested windows and actual coverage differ. Publisher-filled columns are distinguished from raw/adjusted columns. The data vintage is not the original 2023 author download. HAC lags and bootstrap blocks use retained rows across gaps, not guaranteed calendar spacing. Intervals are descriptive, omit calibration and x-error propagation, and do not establish station independence or source-identification uncertainty.

Model reconstruction. GCB historical estimates supply fossil/industry forcing through 2024, starting the modeled perturbation at zero in 1850. Baseline integration treats emissions as constant within each year; endpoint-impulse and conversion-factor alternatives are saved. Model end-year values are compared with annual means. The isotope equation retains the factor in the W-H governing Eq.57 that is missing in its printed Eq.74. This is a disclosed reconstruction, not execution of unavailable author code.

The isotope discrepancy. The annual firn/Bermuda reconstruction gives residual correlation about −0.075 versus the paper’s −0.45. Sampling variants and hypothetical offset tests are saved; they are not measured calibration corrections. Exact author inputs, regional harmonization and firn-age smoothing remain unresolved. The discrepancy is not a fabrication finding or proof the reported number is wrong.

Other conditional calculations. Appending 10.4 GtC/year for 2025 is an explicit paper-assumed scenario, not an observed workbook entry. Short-lifetime inventories test a one-way model and omit return flux. Linear-ramp calculations and historical forcing are different experiments. Squared correlation is distinguished from fixed-prediction skill. Failure to reject an emissions trend is not an equivalence test. Current Law Dome sample matches are diagnostic, not a replay of the older source analysis or a precise preindustrial endmember estimate.

Historical sources. The exact 53-point atmospheric digitization was not located. The cited atmospheric curve was scaled to the sponge record; exact handling of that scaling remains unresolved. The author-hosted Koutsoyiannis v4 includes editorial notes and is not assumed identical to final publisher typesetting. Full methods of several historical papers were inaccessible. None of these gaps was filled with synthetic observations.

Review and correction history

The first mathematical tests were executed successfully, then extended with the stronger net-origin example. The Scripps script ran successfully. The first extended analysis used a wrong Law Dome column name, leaving that diagnostic empty; the corrected run used CO2ppm and supplies the reported results. This was a parser correction, not a change to the observations. The final bounded checks and the separate challenge completed successfully. An initial GCB download returned a license HTML response and a different publisher workbook was inspected; neither supplied the accepted forcing. The accepted ICOS workbook hash is in the input manifest.

Raw execution logs, obsolete intermediate outputs, acquisition headers/cookies, local runtime paths and private coordination are excluded. The public report, original scientific reasoning and sanitized review sequence preserve the substantive corrections and limitations. See the reconciliation for each reviewer’s actual exposure and work; a separate agent is not an independent provider or a human peer reviewer.

Richard Janutka of Qonduit Limited, New Zealand, commissioned the audit. Haivian’s audit service and Validation Journal share organizational control. All scientific reviewers were AI from the same provider. The challenger inherited context and saw the earlier work; the challenge was not blind. No author or provider endorsement is claimed.