diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
new file mode 100644
index 0000000..5051057
--- /dev/null
+++ b/CONTRIBUTING.md
@@ -0,0 +1,96 @@
+# Contributing to the Nexus
+
+## The Hard Rule
+
+**Every PR has a net addition limit of 10 lines.**
+
+Not a guideline. Not a target. A hard limit.
+
+If your change adds 40 lines, you find 30 lines to remove. If you
+can't find lines to remove, your change is too big or you're
+homebrewing something that already exists as a package.
+
+## Why
+
+The codebase must shrink or hold steady over time. New lines are
+debt. The health metric is not features added — it's lines removed.
+
+This rule exists to enforce a philosophy:
+
+- **Import over invent.** If it exists in npm, pip, or another
+  module, use it. New code is a last resort.
+- **Plug in the research.** The AI/ML ecosystem ships new
+  capabilities weekly. Wire them in. Don't reimplement them.
+- **No builder trap.** Writing code feels productive. Shipping
+  the smallest possible change *is* productive. They are not
+  the same thing.
+- **Removal is a first-class contribution.** A PR that deletes
+  50 lines and the tests still pass is more valuable than one
+  that adds 50.
+
+## PR Requirements
+
+Every pull request must include:
+
+### 1. Net line diff within limit
+
+```
++12 -8 = net +4  ✅ (under 10)
++40 -35 = net +5  ✅ (under 10)
++200 -0 = net +200  ❌ (rejected)
++200 -195 = net +5  ✅ (under 10)
+```
+
+### 2. Manual test plan
+
+Describe what you tested by hand. Not "it works" — specific
+steps someone else can repeat:
+
+```
+## Test Plan
+1. Started nexus_think.py with --model timmy:v0.1-q4
+2. Sent chat_message via WS: "Hello"
+3. Verified agent_state changed to "thinking"
+4. Verified chat_response returned within 30s
+5. Checked experience.db — new row with perception + thought
+```
+
+### 3. Automated test results
+
+Paste the output of your test run. If no tests exist for the
+code you're changing, write one — that counts toward your 10
+lines.
+
+```
+## Test Results
+$ python -m pytest nexus/tests/ -v
+test_perception_adapter.py::test_chat_message PASSED
+test_perception_adapter.py::test_heartbeat_filtered PASSED
+test_experience_store.py::test_record_and_recall PASSED
+3 passed in 0.4s
+```
+
+## What This Means for Agents
+
+This rule applies to every contributor: human, Timmy, Claude,
+Perplexity, Gemini, Kimi, Grok. No exceptions.
+
+An agent that can work within this constraint demonstrates
+discipline, judgment, and the ability to find the smallest
+change that solves the problem. An agent that can't is in
+the builder trap.
+
+## Baseline
+
+As of 2026-03-25, the codebase is **4,462 lines** across all
+source files. This number should go down, not up.
+
+## Exceptions
+
+The only exception is the initial commit of a new dependency's
+config file (package.json, requirements.txt, etc.) that enables
+importing existing work. Even then, prefer fewer dependencies
+over more.
+
+There are no other exceptions. If the rule feels impossible,
+the change is too big. Break it into smaller PRs.