diff --git a/local/docs/IMPLEMENTATION-MASTER-PLAN.md b/local/docs/IMPLEMENTATION-MASTER-PLAN.md
index 6c9b4171..bc9bbb87 100644
--- a/local/docs/IMPLEMENTATION-MASTER-PLAN.md
+++ b/local/docs/IMPLEMENTATION-MASTER-PLAN.md
@@ -303,7 +303,25 @@ No shared block layer exists. Each storage driver reinvents I/O dispatch. Linux:
 
 ---
 
-## 13. Total Effort
+## 14. Scheduler & Threading Assessment (2026-05-04)
+
+### Architecture
+- **Kernel**: DWRR scheduler (577 lines), 40 priority levels, per-CPU queues, futex (222 lines)
+- **Userspace**: proc manager (2,638 lines), pthread (440 lines), signal delivery via proc scheme
+- **IPC bridge**: 3 round-trips for thread creation vs Linux's single clone() syscall
+
+### Strengths
+- DWRR with geometric weights, CPU affinity masks, soft-blocking with monotonic timeout
+- Full POSIX process model (PID/PGID/SID, job control, orphan detection)
+- Futex with physical-address keys for cross-process synchronization
+
+### Critical Gaps
+1. **PIT-based tick (~148Hz)** — LAPIC timer exists but `setup_timer()` is commented out. Should use Periodic/TscDeadline mode at 1000Hz.
+2. **Global CONTEXT_SWITCH_LOCK** — spinlock serializes all context switches across CPUs. Should be per-CPU.
+3. **No load balancing** — idle CPUs don't steal work from busy CPUs
+4. **No RT scheduling** — missing FIFO/RR/Deadline classes
+5. **No cgroups** — no CPU bandwidth control or resource limits
+6. **Thread creation latency** — 3 IPC hops vs single clone()
 
 | Tier | Duration |
 |------|----------|