whetstone_DSL/sprint63_plan.md

Sprint 63 Plan: Learned Adapter Hints (Non-Authoritative)

Context

The system can learn from historical review outcomes, but learned behavior must remain bounded by deterministic policy and explicit governance.

---

Goals

1. Introduce learned suggestion layer for adapter decisions
2. Keep learned hints non-authoritative by default
3. Measure whether hints reduce review load and divergence rates
4. Add rollback controls for bad hint generations

---

Steps

Step 859: Hint feature extraction from decision ledgers (12 tests)

Step 860: Pair-specific hint model interface (10 tests)

Step 861: Deterministic fallback contract when hints unavailable (10 tests)

Step 862: Hint confidence + uncertainty packet format (10 tests)

Step 863: A/B harness for hint effectiveness (10 tests)

Step 864: Hint rollback and suppression controls (8 tests)

Step 865: whetstone_get_adapter_hints MCP tool (8 tests)

Step 866: whetstone_set_hint_policy MCP tool (8 tests)

Step 867: Hint safety guardrails and audit fields (8 tests)

Step 868: Sprint 63 integration summary + regression (8 tests)

---

Governance Rule

• Learned hints may suggest; deterministic policy decides.