Audited confirmation benchmark¶
Status reviewed on 9 September 2026 against merge commit
40d10f5.
The classical-estimator confirmation experiment, independent results audit and manuscript
rebuild are complete. This page is the entry point for that experiment. The older
neural_classical_workstation campaign, observational fixtures and EEG pilot are separate
records; they are not the evidence source for the rebuilt manuscript.
Scope and results¶
The experiment evaluates stationary scalar increments from fGn, ARFIMA(0,d,0) and AR(1). It separates clean accuracy, recovery of the latent clean target after contamination, and clean confidence-interval coverage. It does not validate neural recordings, a calibrated LRD test or a universal estimator ranking.
| Audited quantity | Count |
|---|---|
| Independent clean parents | 66,000 |
| Clean process–length cells | 33 |
| Point-estimation pipelines | 21 |
| Separately applied contamination conditions | 23 |
| Contaminated descendants | 161,000 |
| Clean and contaminated point fits | 4,767,000 |
| Parents in the clean interval study | 15,500 |
| Available interval rows | 248,000 |
| Finite bootstrap-statistic attempts | 278,860,500 |
| Canonical cell/domain summary rows | 118,917 |
The stress and interval studies reuse subsets of the clean parents. Their counts must not be added to the independent-parent total. The planned-contrast and convenience exports overlap the six complete summary tables; they are not additional simulations.
Mean removal eliminates constant-offset sensitivity in the geometric estimators but does not remove internal-step sensitivity. Fitted-fGn intervals improve coverage in persistent fGn examples while failing on strongly persistent AR(1) controls. Availability of every interval therefore does not establish calibration. Read the findings and claim ledger for numerical comparisons and their uncertainty.
Evidence and provenance¶
| Purpose | Authoritative source |
|---|---|
| Frozen design, settings and accounting | Protocol and design lock |
| Producer revision | 9c02d04 |
| Results audit revision | b3d0648 |
| Full summary exports and audit evidence | Audit results |
| Manuscript source revision | a2f4ee2 |
| Prose, equations, figures, tables and verification | Manuscript package |
Run identity:
43d0a7f75c8afab6d32ab6fd1f2ff8db49cca59e36971e2b74404173f4cbfcf2.
Canonical summary SHA-256:
95b8e938fb98a674c7ba2e92b7d59ed0f909ffd6ae2e8585c90a0f23f7c01504.
The audit's audit_evidence.json supplies the workload counts;
export_provenance.json supplies the export hashes. Manuscript insertions and plotted
values are mapped to those exports in manuscript_v3/tables/evidence_map.csv.
Rebuild the manuscript¶
The publication sources and complete summary exports are tracked in Git. Building the manuscript does not require rerunning the simulation or obtaining the full raw archive. From a repository checkout, use separate interpreters with the required packages:
SCIENCE_PYTHON benchmark_experiment/manuscript_v3/prepare_assets.py
DOCUMENT_PYTHON benchmark_experiment/manuscript_v3/build_docx.py --output rebuilt-manuscript/lrdbench-manuscript-v3.docx
DOCUMENT_PYTHON benchmark_experiment/manuscript_v3/verify_manuscript.py --docx rebuilt-manuscript/lrdbench-manuscript-v3.docx
Replace the interpreter placeholders with executable paths. The science environment needs
NumPy, pandas and Matplotlib; use the recorded scientific environment for reproduction.
The document environment needs python-docx, lxml and pypdf. Native equation XML is included;
Microsoft Office is only needed by the optional equation-regeneration utility.
See the package README
for the original build environment and rendering procedure. Check every rendered page
after changing prose, tables, figures or layout. The resolved manuscript.md still contains
placement commands and is a typesetting source, not a standalone reading edition.
Reproduce the experiment or audit¶
The production runner deliberately requires its validated Python 3.14.5 environment, dependency pins and exact source hashes. A current checkout is not interchangeable with the frozen producer: later changes even to recorded test sources change its runtime identity. Use the preserved runtime or the specified producer revision and its matching release evidence. Do not loosen that guard to reproduce the paper.
General library CI uses Python 3.12. Its small rehearsal tests use a scoped test runtime; separate cases exercise the production environment guard. A passing CI rehearsal does not authorize a new production run or replace validation of the frozen runtime.
The full audit requires the preserved raw production directory as well as the frozen runtime. Follow the audit reproduction instructions and use a fresh output directory to retain earlier evidence.
Release and publication status¶
These sources were merged into main in PR #3.
Package release 2.0.0 includes the repaired implementations; 1.2.1 predates them.
Installing the package does not install the research archive or recreate the frozen
confirmation runtime. Use the producer revision and archive described above to reproduce
the paper, and record the installed version for new evaluations.
The full raw bootstrap archive has not been deposited publicly. The software concept DOI does not identify this confirmation archive. Author review and declarations, journal formatting, archive deposition and final availability wording remain publication work. The delivered draft dated 8 September predates the GitHub merge; its availability wording must be refreshed during the next editorial revision.
For subsequent work, see Current research next steps.