diff --git a/docs/README.instructions.md b/docs/README.instructions.md
index 99855f4e7c..b4ba54c5ae 100644
--- a/docs/README.instructions.md
+++ b/docs/README.instructions.md
@@ -104,6 +104,7 @@ See [CONTRIBUTING.md](../CONTRIBUTING.md#adding-instructions) for guidelines on
| [Dotnet Wpf](../instructions/dotnet-wpf.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%2Fdotnet-wpf.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%2Fdotnet-wpf.instructions.md) | .NET WPF component and application patterns |
| [draw.io Diagram Standards](../instructions/draw-io.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%2Fdraw-io.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%2Fdraw-io.instructions.md) | Use when creating, editing, or reviewing draw.io diagrams and mxGraph XML in .drawio, .drawio.svg, or .drawio.png files. |
| [Exclude Prompt Data](../instructions/exclude-prompt-data.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%2Fexclude-prompt-data.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%2Fexclude-prompt-data.instructions.md) | Write only the resulting content into files. Never echo prompt instructions, rationale, or meta-commentary into documentation, comments, or code being produced from a prompt. |
+| [Falsifiable Code Review Findings](../instructions/code-review-falsifiable-findings.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%2Fcode-review-falsifiable-findings.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%2Fcode-review-falsifiable-findings.instructions.md) | Code review instructions that require every finding to be falsifiable: file and line, triggering input, expected vs. predicted result, and the smallest failing test. |
| [Fedora Administration Guidelines](../instructions/fedora-linux.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%2Ffedora-linux.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%2Ffedora-linux.instructions.md) | Guidance for Fedora (Red Hat family) systems, dnf workflows, SELinux, and modern systemd practices. |
| [Genaiscript](../instructions/genaiscript.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%2Fgenaiscript.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%2Fgenaiscript.instructions.md) | AI-powered script generation guidelines |
| [Generate Modern Terraform Code For Azure](../instructions/generate-modern-terraform-code-for-azure.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%2Fgenerate-modern-terraform-code-for-azure.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%2Fgenerate-modern-terraform-code-for-azure.instructions.md) | Guidelines for generating modern Terraform code for Azure |
diff --git a/instructions/code-review-falsifiable-findings.instructions.md b/instructions/code-review-falsifiable-findings.instructions.md
new file mode 100644
index 0000000000..9fdd728518
--- /dev/null
+++ b/instructions/code-review-falsifiable-findings.instructions.md
@@ -0,0 +1,69 @@
+---
+description: 'Code review instructions that require every finding to be falsifiable: file and line, triggering input, expected vs. predicted result, and the smallest failing test.'
+applyTo: '**'
+excludeAgent: 'cloud-agent'
+---
+
+# Falsifiable Code Review Findings
+
+Use these instructions for Copilot code review when you want fewer speculative comments and more findings a reviewer can confirm or reject with a test. They suit pull requests that may contain AI-generated code, where a plausible-looking patch can pass the obvious case and fail at a boundary.
+
+## Intent
+
+- Treat the pull request description as the intent of the change: what it must do and what it must not change.
+- If the description does not state the intent, say so in the review and ask for it before calling behavior a bug.
+
+## What to report
+
+- Report only problems that change behavior, security, data, or what a reader must understand.
+- Skip style issues that a formatter or linter would catch.
+
+## Make every finding falsifiable
+
+For each finding, give:
+
+- the file and line;
+- the concrete input or state that triggers the problem;
+- the expected result;
+- the result you predict the code produces;
+- the smallest test that would fail today.
+
+If you cannot name a triggering input, label the item `question`, not `bug`.
+
+## Review order
+
+Check in this order:
+
+1. The stated intent.
+2. Boundary inputs: empty, one, many, duplicates, mixed types, ordering.
+3. State and side effects: mutation, retries, partial failure.
+4. Error paths.
+5. Security-relevant input: authorization, injection, secrets.
+6. Tests that pass without proving the intent.
+
+## Honesty about coverage
+
+- Write `not traced` for any code path you did not follow.
+- Do not claim that you ran code or tests.
+
+## End of the review
+
+End with:
+
+- the three highest-risk findings;
+- what you did not review;
+- one decision: approve, request changes, or need more information.
+
+## Example finding
+
+Code under review: `return [...scores].sort().reverse().slice(0, 3);`
+
+```text
+bug: src/leaderboard.js:6
+Trigger: topScores([9, 100, 25, 3])
+Expected: [100, 25, 9]
+Predicted: [9, 3, 25], because sort() without a comparator compares values as strings
+Failing test: assert.deepEqual(topScores([9, 100, 25, 3]), [100, 25, 9])
+```
+
+A test that uses only two-digit scores passes on the same code, which is why the trigger matters.