Imported from timothyjgardner/frozen_noise_discrimination (AGENTS.md). Install upstream with npx skills add timothyjgardner/frozen_noise_discrimination. Copyright stays with the author.

Agent guide

Objective

Maintain and reproduce the frozen-noise discrimination experiments. A complete successful run must regenerate both numeric sweeps and SVG plots plus the PDF report, then confirm that the numeric values match both reference CSVs.

Start here

From the repository root:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements-lock.txt
.venv/bin/python -m pip install -e . --no-deps --no-build-isolation
.venv/bin/python -m unittest discover -s tests -v
.venv/bin/python scripts/reproduce_all.py

Do not require network access after dependency installation. Do not replace the checked-in reference artifacts during an ordinary verification run.

Ground truth

• Reference numbers: results/accuracy.csv
• Source-statistics numbers: results/source-statistics.csv
• Reference plot: results/accuracy.svg
• Source-statistics plot: results/source-statistics.svg
• Reference write-up: docs/frozen-noise-cortex-writeup.pdf
• End-to-end runner: scripts/reproduce_all.py
• Model defaults: ExperimentConfig and CochleaConfig

The reference write-up has 12 pages. It opens with the McDermott et al. (2013) psychophysical motivation. Its final four pages form a plain-language Appendix A, with vector-typeset equations and explicit definitions of the noise terms.

The expected 1.6 s accuracies are 100.0%, 88.0%, 72.3%, and 59.8% at 0, 10, 20, and 40 spikes/s readout-noise SD. The clean rate distance should decline from about 69.3 spikes/s at 25 ms to 8.2 spikes/s at 1.6 s.

At 20 spikes/s readout noise, the source-statistics extension should show the same-source exemplar mean declining from about 99.8% at 25 ms to 70.9% at 1.6 s. White versus light-pink source accuracy should rise from about 86.8% to 99.8% and cross the exemplar curve between 100 and 200 ms.

Scientific invariant

The current result depends on a whole-trace rate code. The decoder receives one average spike rate per cochlear channel and does not receive precise spike timing. Do not describe a duration-related loss as a general consequence of all neural decoders. A time-resolved decoder is an explicit alternative hypothesis.

In the source task, training and test waveforms must remain independent. Do not replace that task with repeated presentations of fixed waveforms: doing so would test exemplar identity rather than generalization to source statistics.

Safe change workflow

1. Read README.md, src/frozen_noise_cortex/model.py, and src/frozen_noise_cortex/experiment.py.
2. Make the smallest scoped change.
3. Run the unit tests.
4. Run python scripts/reproduce_all.py.
5. If reference numbers intentionally changed, explain why before updating the affected CSV, SVG, PDF, and ARTIFACTS.sha256 together.
6. Render and inspect every PDF page after changing the report builder.

Output policy

Ordinary reproduction writes only to build/reproduced/. Reference artifacts are versioned evidence and should remain unchanged unless a model or report revision is explicitly intended.