diff --git a/references/trend-scoring-modes.md b/references/trend-scoring-modes.md
index bbb6937..635be09 100644
--- a/references/trend-scoring-modes.md
+++ b/references/trend-scoring-modes.md
@@ -92,6 +92,36 @@ The same priority bands apply to both modes (the composite is on the same 0–10
 | 2.0–3.9 | **Low** | Note, skip for now | Park unless the angle sharpens |
 | 0–1.9 | **Skip** | Off positioning | Off positioning |
 
+## Calibration — the expected score distribution (MR-F6)
+
+The five dimensions are **model judgment**, and unanchored judgment inflates: an observed sweep
+put 13 of 20 candidates in **Immediate** (≥8.0), which makes the band meaningless — if most
+things are "draft now", nothing is. The composite is only useful if the bands are **scarce at the
+top**. So a scoring pass must calibrate against this distribution expectation, not score each
+candidate generously in isolation:
+
+- In a **typical sweep**, most candidates belong in **Medium–High (4.0–7.9)**. That is the honest
+  home of "worth writing about, eventually".
+- **Immediate (≥8.0) is the exception, not the rule** — reserve it for candidates that are
+  genuinely exceptional *relative to the rest of this sweep*. As a working target, expect on the
+  order of **≤3 of 20** in Immediate; a sweep that floods the top band is uncalibrated, not lucky.
+- **Skip/Low (<4.0)** is a real outcome — an off-pillar or already-resolved topic should land
+  there, not get floated to Medium to be polite.
+
+Mechanics for the pass:
+
+- **Score relatively across the batch.** A 9–10 on a dimension means "exceptional versus the other
+  candidates in *this* sweep", not "good in the abstract". Rank the batch, then assign — a flat
+  batch of 8s is the failure mode this rule exists to catch.
+- **The distribution shape is the mechanism; the threshold of what counts as exceptional is the
+  user's calibration** (the dimension rubrics above + the profile), never a hard-coded fact about
+  any one operator's topics. Domain-general: the anti-inflation rule travels; the specific bar does
+  not live here.
+- **This is now measurable.** The composite + band are **persisted** on each record (RE-R3a,
+  `TrendRecord.score`), so the distribution can be checked across runs — e.g. count the Immediate
+  share of a stored sweep. The MR-F7 band-cap gate (N7) builds on the same persisted signal: a
+  candidate with no formulable reader-grip is capped below Immediate regardless of composite.
+
 ## Consumers
 
 - `agents/trend-spotter.md` — reads the requested mode and applies the matching rubric