diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
new file mode 100644
index 0000000..18b109d
--- /dev/null
+++ b/CONTRIBUTING.md
@@ -0,0 +1,31 @@
+# Contributing to PSR-Net
+
+Thanks for looking at this. PSR-Net has two halves and they are contributed to differently.
+
+## The browser pipeline
+
+Everything under the page itself runs client side with no build step. Open the page, scroll, and the deterministic chain executes on the real 12-bit values. To change a restoration stage, edit its module and reload.
+
+## The trained network
+
+The model is 714,401 parameters, trained offline and exported to ONNX. If you change the architecture you must re-export:
+
+    python psr/train/export_onnx.py
+
+## Reproducing the results table
+
+Before proposing a change that claims a metric improvement, reproduce the baseline:
+
+    python psr/train/evaluate.py
+
+The table in the README is measured on 24 held-out scenes with fixed degradation seeds, with both methods scored after the same affine fit to the target. A pull request that reports different numbers should say which of those conditions it changed.
+
+## What makes a good pull request
+
+- State which metric moved and on what data.
+- Do not replace a measured claim with an unmeasured one.
+- If a change improves SSIM but costs PSNR, say so. Both numbers belong in the table.
+
+## Reporting problems
+
+Open an issue with the scene, the stage, and what you expected. For questions about method rather than defects, use Discussions.
\ No newline at end of file