diff --git a/.claude/skills/find-weird-recommendations/SKILL.md b/.claude/skills/find-weird-recommendations/SKILL.md
index 04e0ee36c..b51723957 100644
--- a/.claude/skills/find-weird-recommendations/SKILL.md
+++ b/.claude/skills/find-weird-recommendations/SKILL.md
@@ -61,6 +61,17 @@ Writes `modelling_audit.md` / `.csv`. The check that carries this skill:
   no budget cap, the only reasons to fall short are (a) a measure was wrongly
   withheld — a bug — or (b) the dwelling genuinely can't reach the band. Phase 4
   decides which. Budget-capped scenarios are excluded (falling short is expected).
+  **Caveat: this check is noisy on flat-heavy portfolios** — it fired 322/338 on
+  portfolio 814 (every flat a band short of B). Use `plan-stuck-in-low-band` to
+  cut straight to the real hits.
+- **`plan-stuck-in-low-band`** (HIGH) — the plan leaves the dwelling **still
+  rated E/F/G with the band unchanged** from baseline (unlimited budget). This is
+  the sharp "E→E, no big movement" worklist — the genuinely-trapped tail, not
+  every dwelling a band short of goal. On portfolio 814 it flagged just **4**
+  (pids 742121, 742210, 742265, 742347) vs `plan-stops-short-of-goal`'s 322. A
+  wrongly-withheld measure or a physically-trapped dwelling both surface here;
+  Phase 4 tells them apart. Band-keyed (no SAP-points threshold), and the
+  strictly-worse case is left to `plan-below-baseline-band` so the two partition.
 
 Also read `plan-below-baseline-band`, `already-meets-goal-with-works`, and — if
 you pass `--with-recommendations` (EXPLAIN-gated; confirm with the user first) —
@@ -136,11 +147,39 @@ Match each deep-dive against these before hypothesising a novel cause:
   main fuel / heat-network flag, then fix by restoring the gate in
   `_hhr_storage_eligible` (community heating ⇒ not HHR-eligible). A community
   network is a shared asset a single dwelling can't rip out for storage heaters.
-- **Measure withheld → plan stops short of goal** (pid 742121, "why no HHRSH").
+- **Measure withheld → plan stops short of goal** (the "why no HHRSH" class).
   The dwelling can't reach the goal band, and the expected bundle is *absent from
   candidates*. Read the generator's eligibility predicate against the live EPC —
   an over-tight gate (a fuel/description/category condition that shouldn't
   exclude this dwelling) is the usual cause.
+- **Measure OFFERED but not selected because it *lowers* SAP** (pid 742121, "why
+  no HHRSH" — re-modelled 2026-07, NOT the stale mains-gas story). This dwelling
+  is an **electric** maisonette (`main_fuel_type=29`), so HHRSH **is** in the
+  candidate set — eligibility passes. But scoring the HHRSH overlay in isolation
+  gives **−6.3 SAP** (54.3 → 48.0): swapping its efficient existing electric
+  system for off-peak storage heaters *worsens* the cost-based SAP, so the
+  Optimiser correctly drops it. Its fabric is already done (double-glazed, walls
+  insulated, no cylinder, no own roof), so the plan is **empty** and it sits at E,
+  unreachable to B — a legitimate £0 result, not a bug. **Lesson: a measure
+  *absent from candidates* is a generator-gate bug; a measure *present but not
+  chosen* is the Optimiser judging it not worth it — score the overlay
+  (`PackageScorer.score(epc, [EpcSimulation(...)])`) before blaming a gate.** The
+  old note here claimed 742121 was a mains-gas maisonette that "correctly gets
+  only lighting" — that was stale; verify fuel against the live EPC every time.
+- **Community-heated flat stuck low, correctly** (pid 742265, "might need loft
+  insulation?"). A **community-heated** flat (SAP main-heating code 301, category
+  6) on portfolio 814. HHRSH and ASHP are both gated out by `is_heat_network_main`
+  (the topology guard) — right, per the community bug above. Its roof is
+  `"Pitched (vaulted ceiling)"` with insulation lodged as **`ND` (no data)**, so
+  the loft generator withholds insulation: it only fires on a definite `0 mm`
+  (`roof_insulation_thickness != _ROOF_UNINSULATED_MM` → None), and a vaulted
+  ceiling isn't the loft-fallback's target anyway. Its only real lever is IWI
+  (+6.7), not enough to clear E — so it stays E→E, legitimately. **Watch item:**
+  a vaulted-ceiling roof with `ND` insulation gets *no* roof measure at all
+  (neither the sloping-ceiling branch, which matches the literal `"sloping
+  ceiling"`, nor the loft branch, which needs `0 mm`). If several such roofs turn
+  out to be genuinely uninsulated upstream, that's a mapper/gate gap — but `ND`
+  is unknown, not zero, so withholding is defensible for now.
 - **Missing `secondary_heating_removal`** (pid 742264, "gov site shows secondary
   heating"). Check whether the live EPC carries a secondary-heating system and
   whether
diff --git a/scripts/audit/anomalies.py b/scripts/audit/anomalies.py
index 7e9f627a8..532108f7c 100644
--- a/scripts/audit/anomalies.py
+++ b/scripts/audit/anomalies.py
@@ -175,6 +175,50 @@ def _plan_stops_short_of_goal(a: PropertyAudit) -> Optional[str]:
     )
 
 
+@check("plan-stuck-in-low-band", Severity.HIGH)
+def _plan_stuck_in_low_band(a: PropertyAudit) -> Optional[str]:
+    """The plan leaves the dwelling STUCK in a low band (E or worse) with the band
+    UNCHANGED from baseline — the engine spent (or spent nothing) yet moved the
+    dwelling nowhere off the floor.
+
+    Complements ``plan-stops-short-of-goal``. That check fires on anything below
+    the goal band — 322/338 on portfolio 814, mostly flats a single band short of
+    B, which is expected and drowns the real hits. This one isolates the
+    genuinely-stuck tail: the handful still rated E/F/G *after* works, where a
+    wrongly-withheld measure or a physically-trapped dwelling actually shows up.
+    It is the "E→E, no big movement" worklist the tech team asked for.
+
+    Motivated by portfolio 814 (scenario 1271, goal B, unlimited budget): only 4
+    properties land E-or-worse and all 4 stay at E — pid 742121 (electric
+    maisonette, £0 works: HHRSH *is* offered as a candidate but scores −6.3 SAP,
+    so the Optimiser correctly drops it, and the fabric is already done — no lever
+    left); pid 742265 (community-heated flat, SAP code 301: HHRSH and ASHP are
+    gated out by the heat-network topology, leaving IWI its only real lever); pids
+    742210 / 742347 (flats that move ~5-8 SAP but never clear E). Legitimate-but-
+    trapped and real-bug cases both surface here — the deep-dive
+    (``run_modelling_e2e`` + the candidates CSV) tells them apart. Budget-capped
+    scenarios are excluded: staying low is expected once the money runs out.
+
+    Band-keyed, not a SAP-points threshold, so it needs no tuned magic number —
+    "post band E/F/G and no band improvement" is the whole rule. The strictly-
+    worse case (post below baseline) is left to ``plan-below-baseline-band`` so
+    the two partition cleanly."""
+    post, base = _band_rank(a.post_band), _band_rank(a.effective_band)
+    if post is None or base is None:
+        return None
+    if post < _BANDS.index("E"):  # A–D: out of the low tail, not "stuck"
+        return None
+    if post != base:  # band moved (a lift, or the worse case above) — not stuck
+        return None
+    if a.scenario_budget is not None:  # budget-capped: staying low is expected
+        return None
+    return (
+        f"still {a.post_band} ({a.post_sap}) after works — no band lift from "
+        f"baseline {a.effective_band} ({a.effective_sap}) on an unlimited-budget "
+        f"scenario (cost_of_works £{a.cost_of_works or 0:.0f})"
+    )
+
+
 @check("post-band-score-mismatch", Severity.MEDIUM)
 def _post_band_score_mismatch(a: PropertyAudit) -> Optional[str]:
     """The persisted post band disagrees with the band the post SAP implies — a