diff --git a/docs/README.instructions.md b/docs/README.instructions.md
index 2b142e6f9f..d4a645b27f 100644
--- a/docs/README.instructions.md
+++ b/docs/README.instructions.md
@@ -174,6 +174,7 @@ See [CONTRIBUTING.md](../CONTRIBUTING.md#adding-instructions) for guidelines on
| [Quarkus MCP Server](../instructions/quarkus-mcp-server-sse.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fquarkus-mcp-server-sse.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode-insiders%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fquarkus-mcp-server-sse.instructions.md) | Quarkus and MCP Server with HTTP SSE transport development standards and instructions |
| [R Programming Language Instructions](../instructions/r.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fr.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode-insiders%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fr.instructions.md) | R language and document formats (R, Rmd, Quarto): coding standards and Copilot guidance for idiomatic, safe, and consistent code generation. |
| [React Controls & Platform Libraries](../instructions/pcf-react-platform-libraries.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fpcf-react-platform-libraries.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode-insiders%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fpcf-react-platform-libraries.instructions.md) | React controls and platform libraries for PCF components |
+| [Research Experiment Discipline](../instructions/research-experiment-discipline.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fresearch-experiment-discipline.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode-insiders%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fresearch-experiment-discipline.instructions.md) | Discipline for research and experiment code: score null models first, keep calibration and held-out data physically separate, reproduce a published baseline before claiming improvements, report gains with paired error bars, and verify every guard by deliberately breaking it. Apply when writing evaluation harnesses, benchmark sweeps, or any code that reports a performance number. |
| [Ruby MCP Server Development Guidelines](../instructions/ruby-mcp-server.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fruby-mcp-server.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode-insiders%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fruby-mcp-server.instructions.md) | Best practices and patterns for building Model Context Protocol (MCP) servers in Ruby using the official MCP Ruby SDK gem. |
| [Ruby on Rails](../instructions/ruby-on-rails.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fruby-on-rails.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode-insiders%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Fruby-on-rails.instructions.md) | Ruby on Rails coding conventions and guidelines |
| [Rust Coding Conventions and Best Practices](../instructions/rust.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Frust.instructions.md)
[](https://aka.ms/awesome-copilot/install/instructions?url=vscode-insiders%3Achat-instructions%2Finstall%3Furl%3Dhttps%3A%2F%2Fraw.githubusercontent.com%2Fgithub%2Fawesome-copilot%2Fmain%2Finstructions%2Frust.instructions.md) | Rust programming language coding conventions and best practices |
diff --git a/instructions/research-experiment-discipline.instructions.md b/instructions/research-experiment-discipline.instructions.md
new file mode 100644
index 0000000000..57eba606fd
--- /dev/null
+++ b/instructions/research-experiment-discipline.instructions.md
@@ -0,0 +1,57 @@
+---
+description: 'Discipline for research and experiment code: score null models first, keep calibration and held-out data physically separate, reproduce a published baseline before claiming improvements, report gains with paired error bars, and verify every guard by deliberately breaking it. Apply when writing evaluation harnesses, benchmark sweeps, or any code that reports a performance number.'
+applyTo: '**'
+---
+
+# Research Experiment Discipline
+
+In research code the failure mode is rarely a crash; it is a number that
+looks great and is wrong. These instructions make evaluation code hard to
+fool - starting with its own author.
+
+## Scoring and evaluation
+
+- Route all method scoring through a single evaluation entry point module;
+ never compute metrics inline in experiment scripts.
+- Score null-model baselines first (constant output, untrained model, input
+ copy). A well-scoring null model means the harness is broken: stop and fix
+ it before evaluating anything else.
+- Keep one positive control - a signal the pipeline must detect. If it stops
+ detecting it, freeze conclusions; every null result since is unusable.
+- Pin the metric convention (data range, averaging order, aggregation) in
+ one place. If a published convention differs from yours, report both,
+ labelled.
+
+## Data separation
+
+- Calibration/tuning data and evaluation data live in different files and
+ are loaded by different code paths; hyperparameter tuning reads
+ calibration data only.
+- Split on the unit of independence (patient, user, site, time period), not
+ on files - two records from the same entity in different splits is group
+ leakage, and it inflates cross-validated gains.
+- Budget accesses to the evaluation set; keep one final untouched split that
+ is scored exactly once for the headline number.
+
+## Baselines and claims
+
+- Reproduce at least one published baseline number before trusting your own.
+ A baseline you cannot match means the recipe has unread layers (optimizer,
+ loss, metric convention, operator) - do not "improve" an unmatched
+ baseline.
+- A gain that does not reproduce on held-out data is not reported.
+- Report confirmed gains as paired differences with an interval across
+ instances or seeds; a sub-point gain whose interval crosses zero is noise,
+ not a result.
+- Persist numbers to files and commit them before quoting them in text.
+
+## Guards and tests
+
+- Algorithm implementations exist exactly once, in a module; experiment
+ scripts import them, never re-implement inline.
+- Every new check or test must be demonstrated to fail on a deliberately
+ broken input before it counts - and fail for the right reason, not because
+ an unrelated assertion tripped first.
+- When a hyperparameter sweep looks flat, measure the gradient force balance
+ between loss terms before concluding the parameter is inert; a flat sweep
+ often means every tested value sat on one side of the balance point.