diff --git a/jouleclaw/jouleclaw-rs/crates/jouleclaw-cascade/src/router.rs b/jouleclaw/jouleclaw-rs/crates/jouleclaw-cascade/src/router.rs
index 0e6fb96eb..cf24ce879 100644
--- a/jouleclaw/jouleclaw-rs/crates/jouleclaw-cascade/src/router.rs
+++ b/jouleclaw/jouleclaw-rs/crates/jouleclaw-cascade/src/router.rs
@@ -55,6 +55,16 @@ pub struct RoutingPlan {
 
     /// Joules spent producing this plan. Counted against the query
     /// budget.
+    ///
+    /// **Ungraded.** This is a bare `f64`, so a figure a router measured and a
+    /// declared constant are indistinguishable here — and because it is a
+    /// debit, a figure below the work actually done under-charges the budget,
+    /// which is the unsafe direction at a ceiling (`jouleclaw/SPEC.md` §1.3).
+    /// A router whose routing cost scales with its own state, or that runs a
+    /// model to choose a route, must report what it spent rather than a
+    /// constant; see `jouleclaw_routing::ROUTING_JOULES` for why the flat
+    /// constant is a floor. Carrying a provenance beside this field is what
+    /// would let a verifier tell the two apart.
     pub router_joules: f64,
 
     /// Human-readable explanation of WHY this ordering. Surfaces in
diff --git a/jouleclaw/jouleclaw-rs/crates/jouleclaw-routing/src/lib.rs b/jouleclaw/jouleclaw-rs/crates/jouleclaw-routing/src/lib.rs
index acb1a2a30..8f213ade9 100644
--- a/jouleclaw/jouleclaw-rs/crates/jouleclaw-routing/src/lib.rs
+++ b/jouleclaw/jouleclaw-rs/crates/jouleclaw-routing/src/lib.rs
@@ -43,7 +43,31 @@ pub use privacy::{
     SensitivityLevel, TierSensitivityPolicy, PRIVACY_FILTER_JOULES,
 };
 
-/// The (negligible) joule cost charged for producing a routing plan.
-/// L5 lives in the meta-cognitive control plane; its energy is client
-/// CPU only and is dwarfed by every execution tier it orders.
+/// The joule cost charged for producing a routing plan: an **`Estimator`-class
+/// declared floor**, not a measurement.
+///
+/// L5 lives in the meta-cognitive control plane, and for a table-driven or
+/// heuristic router its energy really is client CPU only and dwarfed by every
+/// execution tier it orders. Two things keep that from being the whole story,
+/// and both matter because [`RoutingPlan::router_joules`] is **counted against
+/// the query budget**, so a charge below the work done under-debits it — the
+/// unsafe direction at a ceiling (`jouleclaw/SPEC.md` §1.3).
+///
+/// 1. **It does not scale with the work.** [`LearnedRouter`]'s `route` tokenizes
+///    the query and scans its episode memory for neighbours, which is work
+///    proportional to the number of episodes; this constant is flat. Nothing
+///    bounds that memory, so the gap grows silently with it.
+/// 2. **A router that runs a model must not use it.** A classifier or an LLM
+///    choosing a route costs orders more than this, and charging the constant
+///    would report a model inference as picojoules.
+///
+/// A router whose routing cost scales with its own state, or that runs a
+/// model, **MUST** report what it actually spent rather than this constant.
+/// Nothing enforces that today: `router_joules` is a bare `f64` and carries no
+/// provenance, so a measured figure and this floor are indistinguishable in
+/// the type. Giving [`RoutingPlan`] a provenance field is what would make the
+/// distinction checkable instead of documented.
+///
+/// [`RoutingPlan`]: jouleclaw_cascade::router::RoutingPlan
+/// [`RoutingPlan::router_joules`]: jouleclaw_cascade::router::RoutingPlan::router_joules
 pub const ROUTING_JOULES: f64 = 8e-9;