diff --git a/.github/copilot-instructions.md b/.github/copilot-instructions.md
new file mode 100644
index 0000000..cb007b2
--- /dev/null
+++ b/.github/copilot-instructions.md
@@ -0,0 +1,102 @@
+# DSProdModels — instructions for Copilot code review
+
+The physics inputs for custom CMS Monte-Carlo production driven by
+[DSProd](https://github.com/cms-flaf/DSProd): generator cards, gen fragments, per-process plugins
+and the production setups that say what to produce, for which eras, and how many events.
+
+**Read `FLAF/.github/copilot-instructions.md` first** for the shared rules on what a useful review
+comment looks like and what not to flag, and `DSProd/.github/copilot-instructions.md` for the
+production machinery this repository feeds.
+
+## Why review matters more here than anywhere else
+
+**This repository has no CI.** No formatting check, no consistency check, no integration test.
+Review is the only gate between a change and a production that runs for days on the grid.
+
+Nothing here executes at review time, so nothing fails fast. A wrong number in a setup or a wrong
+parameter in a card produces samples that are valid files, pass every downstream task, and are
+simply not the physics that was asked for. The cost of finding that out later is the whole
+production.
+
+Read a diff here the way you would read a configuration change to a running system: assume it
+will be applied exactly as written, and ask what it will produce.
+
+## Invariants
+
+### Numbers must be derivable
+
+Production setups (`*/setups/*.yaml`) carry `events_total` per era, and the existing files explain
+in comments **how those numbers were derived** — a luminosity scaling from the central production,
+then a unification rule across mass points, with the arithmetic written out.
+
+A diff that changes an event count, adds an era, or adds a mass point must keep that derivation
+truthful. Check the arithmetic against the stated rule, and flag a number that no longer follows
+from the comment above it, or a comment left describing the previous numbers. A count changed
+without its justification is the single most expensive kind of error this repository can contain:
+too low and the sample is unusable, too high and the grid time is wasted.
+
+### A card change invalidates what already exists
+
+`cards/` (`run_card.dat`, `proc_card.dat`, `customizecards.dat`, `extramodels.dat`) and
+`fragments/` determine the generated physics. Editing them changes what a **new** gridpack
+produces while any gridpack already built from the old cards stays as it was. Ask whether the
+change is meant to apply retroactively, and if so, whether the affected gridpacks and samples are
+being regenerated. Silent divergence between a card and the samples produced from it is not
+detectable downstream.
+
+### Names are an interface
+
+Points name a production mode and a final state using the **tokens the corresponding central
+dataset uses on DAS**, and the plugin maps those onto a cards directory and a fragment. A renamed
+token, a new production mode without its `PRODUCTION_MODES` entry, or a final state without a
+matching fragment breaks the mapping — and it breaks it at production time, not at review time.
+
+### Era coverage must be complete and intentional
+
+`eras`, `nano_versions` and any per-era mapping must agree with each other and with the conditions
+DSProd declares for those eras. An era listed in one and missing from another is a gap that only
+appears when someone produces it. Note that Summer24 MC is shared across 2024, 2025 and 2026 — a
+count given for 2024 covers all three.
+
+### The gridpack store is not here
+
+Gridpacks live in `cms-flaf/DSProdGridpacks` and are checked out sparsely with Git-LFS downloads
+disabled. Never suggest committing a gridpack, a tarball or any other large binary to this
+repository.
+
+## Documentation must ship with the change
+
+A PR must update the documentation **in the same PR** whenever it changes anything a user can
+observe: a setup key, a point-naming convention, a supported production mode or final state, the
+meaning of an existing field, or what a process requires.
+
+Where it goes:
+
+- the per-process `README.md` files (`<process>/README.md`, `<process>/<generator>/<...>/README.md`)
+  carry the physics and the links to the original sources — keep them accurate;
+- the plugin docstrings describe the naming contract and are read as documentation;
+- anything about how DSProd *consumes* these files belongs in `DSProd/docs/`
+  (`configuration/prod-setups.md`, `configuration/processes.md`), which is a **separate
+  repository** and therefore a companion PR. Flag its absence.
+
+The comments inside a setup file are documentation too: they are the only record of how its
+numbers were obtained.
+
+## Do not flag
+
+- The generator card formats — `run_card.dat` and friends are MadGraph's, not ours to restyle.
+- Verbose comment blocks in setups and plugins; here they are the derivation record, not clutter.
+- Duplication between production modes or final states; they are kept explicit so each can be read
+  on its own.
+- Missing tests — nothing here is executable in isolation.
+
+## Repository facts
+
+Verified 2026-08-27; re-check before relying on any of it.
+
+| | |
+|---|---|
+| Layout | `<process>/` (e.g. `X_HH/`) containing `README.md`, `setups/*.yaml`, and `<generator>/` (e.g. `MadGraph5_aMCatNLO/`) with `plugin.py` and `<comEnergy>/` holding `cards/<production_mode>/` and `fragments/<final_state>.py` |
+| Consumed by | `cms-flaf/DSProd`, as the `models` submodule |
+| Related | gridpacks in `cms-flaf/DSProdGridpacks` (CERN GitLab, sparse LFS); conditions in `DSProd/config/conditions_Run3.yaml` |
+| CI | **none** — no workflows in this repository |