Saved evidence · NLSS 2.0.1
Reproduce and inspect
Material 1.0 · NLSS 2.0.1 · R 4.5.2 · 18 September 2026.
Keep the starter and reference folders separate. The starter contains the study materials without answers. The reference includes the finished report, scripts, unrounded checks and original run evidence. No NLSS runtime is included.
Check the numbers independently
From an untouched extracted reference folder, run:
Rscript scripts/check-reference.R intervention . ../intervention-check-new.json
The checker reads the frozen CSV (or the power assumptions), fits the stated methods directly in R and compares unrounded results. It does not call NLSS's calculation code. The methods share their underlying R libraries; this is independent verification of the NLSS workflow, not an independently implemented mixed-model solver. Package versions are recorded in the saved requests and reference-values.json.
Noniterative tolerance is 1e-8 + 1e-7 * abs(reference); iterative estimates use 1e-6 + 1e-5 * abs(reference). Categories and row membership must match exactly. A check output path must be new. scripts/check-reference.R contains branches for the four example types; only the named case is executed.
| Evidence | Original run |
|---|---|
| contrast | 20260918T102709-4c4502c2 |
| numeric-time | 20260918T102955-4612799b |
| figure | 20260918T103002-11b1d4c62c8 |
runs.json maps each operation to its actual run. material-manifest.json and SHA256SUMS identify the delivered files. They describe the original package before any new runs. The manifest is a material list, not a substitute for NLSS's evidence.
Regenerate the data or replay an analysis
Rscript scripts/generate-data.R ../intervention-regenerated.csv
Rscript scripts/check-figures.R intervention . ../intervention-figure-check-new.json
The fixed generator refuses to replace an existing file. Compare the regenerated SHA-256 with input-version.json. Do not select a new seed to obtain a more convenient result.
For an exact NLSS replay, work in a copy of the complete reference project, including hidden .nlss/ files. Ask your agent:
Use NLSS replay-run on .nlss/runs/20260918T102709-4c4502c2/request.json in this copied project. Preserve the original evidence and current working data. Explain any environment or integrity mismatch without weakening the verification.The supported technical command is Rscript <installed-nlss>/scripts/R/run_nlss.R replay-run --request <copied-project>/.nlss/runs/20260918T102709-4c4502c2/request.json. Replay adds a new run; it does not rewrite the original evidence or reproduce the agent's prose. Numerical checks, replay and report integrity answer different questions.
The authored report.md was saved through project-report with selected evidence and its revision explicitly verified. That integrity check does not certify scientific correctness. Original protocol presentation is retained: NLSS 2.0.1 can use a figure caption as its heading and number the alt text inconsistently. The report gives an explicit caption; the numerical graphical data are checked separately.
Figure presentation
The report's figures/trajectory.png re-renders the four saved NLSS means with clearer pre/post labels, stronger contrast and different shapes/line styles. Recreate it with Rscript scripts/render-trajectory.R . ../trajectory-new.png. No new model or smoothing is applied; original NLSS plot evidence remains unchanged. Whole-number axis preparation is recorded as measurement = 1*(time=="post"). An earlier recode added a character occasion column, retained in the reference but unused in the final figure. The model uses the original time variable.
Original material: Mike Hammes. Licenses.