orx-reports

Write reports, figures, CSVs, PDFs, and other outputs directly into the artifacts directory shown in the session playbook. Written files become project artifacts immediately.

Before writing, inspect the existing artifacts directory and follow its organization. Reuse the relevant folder for follow-up outputs. Group related outputs by research topic or deliverable instead of accumulating files at the artifacts root. Add subfolders such as figures/, data/, or models/ only when they help; do not create empty scaffolding.

Use a descriptive filename for each output. For example:

<artifacts-dir>/
  transformer-sweep/
    transformer-sweep-report.md
    figures/
      orx_figstyle.py
      patch-size.pdf
      patch-size.svg
      patch-size.py
      learning-rate.pdf
      learning-rate.svg
      learning-rate.py
    data/
      sweep-results.csv

Keep reproducibility scripts beside their figures. Keep temporary logs and scratch files outside the artifacts directory.

Keep already-published outputs at their existing paths. Reorganize other older outputs only when requested, updating affected document links and verifying they still resolve.

Read the orx-figures module before writing any figure; a default matplotlib plot does not meet the bar the reports are held to.

Use relative links within reports, for example ![Patch size](figures/patch-size.svg) from transformer-sweep-report.md. In the chat handoff, link every finished output using the session playbook's evidence-and-links contract, including the full nested path, for example <file path="artifacts/transformer-sweep/figures/patch-size.svg" />. When an output is an image, also show it inline in the chat handoff using ![Description](artifacts/<relative-path>), while keeping its file link. Put related image previews in consecutive paragraphs so they render as a carousel, with file links after the group. Load orx-results when the report requires interpreting or comparing run results.