diff --git a/.github/PULL_REQUEST_TEMPLATE.md b/.github/PULL_REQUEST_TEMPLATE.md
new file mode 100644
index 000000000..f0ed5a737
--- /dev/null
+++ b/.github/PULL_REQUEST_TEMPLATE.md
@@ -0,0 +1,42 @@
+## What & why
+
+<!-- One paragraph. Link the FEATURE / issue this PR satisfies. -->
+
+**Linked:** <!-- #issue / TAF-XXX -->
+
+## Definition of Done
+
+> **Red-then-green:** a fix lands with a test that **failed before** it and **passes after**. The expected value comes from the spec, never from what the code currently returns.
+
+- [ ] Requirement linked & scope matches the FEATURE / issue
+- [ ] Added/updated a test that is **red without this change, green with it**
+- [ ] Tests are at the right layer (unit / contract / e2e) and assert **intended** behavior, not current output
+- [ ] Local checks pass **before push** — Rust: `cargo fmt --check`, `cargo clippy -- -D warnings`, `cargo test`; .NET: `dotnet test`; Docs: `cpt validate`
+- [ ] Coverage held or raised vs the base branch — the **ratchet** (CI fails if coverage drops below the base)
+- [ ] Negative paths covered (null / empty / error / unauthorized) — no silent failure
+- [ ] Docs / spec (DESIGN / FEATURE) updated
+
+### Data (dbt / connectors / SQL) — or check N/A: [ ]
+
+- [ ] New connector: its data actually reaches the unified `silver:class_*` set
+- [ ] Changed silver schema: dependent **gold views still resolve** (no missing column)
+- [ ] New/changed model: strict **`contract: {enforced: true}`** with declared `data_type` (prevents type drift)
+- [ ] Changed a metric: added/updated a **golden-value** regression test (parity with prior values)
+- [ ] ReplacingMergeTree `order_by` dedup keys **cannot be NULL** (NULL ≠ NULL → double counting)
+- [ ] Source freshness covered where applicable
+
+### API (serving layer) — or check N/A: [ ]
+
+<!-- The React frontend lives in the insight-front repo; its render-contract /
+     Vitest / Playwright checklist belongs in that repo's PR template. -->
+
+- [ ] **Honest NULLs:** when there's no data the API returns `NULL` / a `ComingSoon` flag, never `0`
+- [ ] **Error vs empty:** a real failure is returned as an error status (e.g. HTTP 500), distinct from an empty-but-OK result, so the UI can tell them apart (not a silent empty body)
+
+## Evidence
+
+<!-- Paste the test summary + coverage % (before → after) and any gate output. -->
+
+```text
+
+```