Upload SKILL.md

.claude/skills/implement-drift/SKILL.md

@@ -0,0 +1,32 @@
+---
+name: implement-drift
+description: Implement the drift detection layer to satisfy tests/test_drift.py
+disable-model-invocation: true
+---
+
+Implement `osint_core/drift.py` to satisfy `tests/test_drift.py`.
+
+1. **Read first:**
+   - Read `osint_core/drift.py` — it is pseudocode (raises `SyntaxError` on import), treat it as a design spec only
+   - Read `tests/test_drift.py` in full — these are the contract tests to satisfy
+   - Required exports: `DriftAssessment`, `DriftSignal`, `DriftType`, `DriftVector`, `TelemetrySnapshot`, `aggregate_signals`, `assess_drift`, `choose_dominant_drift_type`, `estimate_confidence`, `recommend_correction`
+
+2. **Implement from scratch** (do not edit the pseudocode in-place):
+   - Use `from __future__ import annotations`
+   - Use `@dataclass(frozen=True)` for all value objects
+   - Use `Literal[...]` or `enum.Enum` for `DriftType` and correction verbs
+   - `assess_drift` must be **pure** — no mutations to inputs (see `test_assess_drift_is_pure_and_does_not_mutate_inputs`)
+   - Correction priority: `policy > structural > behavioral > adversarial > operational > statistical`
+   - Adversarial signals must CONSTRAIN before the system ADAPTs; statistical drift may ADAPT only when nothing higher-priority fires
+
+3. **Iterate test-by-test:**
+   ```bash
+   pytest tests/test_drift.py -v -x   # stop at first failure
+   ```
+
+4. **When all pass:**
+   ```bash
+   pytest -v
+   ruff check osint_core/drift.py
+   bandit -r osint_core/drift.py
+   ```