Measurement honesty so we optimise SAP-relevant accuracy, not SAP-neutral
misses (ADR-0030 Component Accuracy):
- Add construction_age_band_pm1: an exact-or-adjacent-band hit. Adjacent
RdSAP age bands carry near-identical U-values, so an off-by-one is
~SAP-neutral. Full corpus: exact 78.5% but ±1-band 91.7% (fixture
63.9% -> 83.3%) — most age misses are adjacent.
- Drop window_count from the gate's residual ceilings (cosmetic): the
predicted picture clusters at a mapper-default 4 windows vs actuals 1-21,
but total_window_area (the SAP-relevant signal) stays tight at ~3.4 m2.
Gate: + construction_age_band_pm1 floor 0.8333; window_count no longer gated.
Closes#1222
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Investigated recency-weighting (weight cohort votes by an exponential decay
in cert age). Key finding: it must be SELECTIVE. On the validation corpus it
HURTS permanent categoricals (wall 91.2->89.5, age 78.5->75.7 — discards
still-valid data) but clearly HELPS time-varying ones, where a recent
neighbour reflects the current physical state:
roof_insulation_thickness 56.7 -> 60.7% corpus (+4pp)
29.4 -> 41.2% fixture (+12pp)
So apply a recency-weighted mode only to roof_insulation_thickness (loft
top-ups happen over time); keep the plain mode for permanent categoricals.
tau = 4yr (~2.8yr half-life); falls back to plain mode when no registration
dates are lodged. Gate floor ratcheted 0.2941 -> 0.4118.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
These independent fabric categoricals were template-copied; mode them like
the construction categoricals. Verified mode beats template before applying.
Big fixture win on roof insulation thickness (doubled), floor insulation
neutral-to-positive:
roof_insulation_thickness 14.7% -> 29.4% (gate floor ratcheted up)
floor_insulation 90.6% (unchanged on the fixture)
Glazing type was tried too (+1.6pp on the 40-postcode corpus) but REGRESSED
the 36-target fixture (0.50 -> 0.44) — the gate caught it. Glazing moding is
marginal/noisy, so it's left on the template; revisit with a larger corpus.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
"One scorer, two harnesses" (ADR-0030): the committed gate, the local script,
and the future battle-test must run the *same* scoring. Extract it:
- domain/epc_prediction/validation.py — `iter_predictions` (the single
leave-one-out orchestration: latest-per-address hold-out, SAP-10.2 target
filter, all-vintage source) + `evaluate_component_accuracy` (calculator-free
ComponentAccuracy aggregation, the primary signal). Unit-tested.
- harness/epc_prediction_corpus.py — `load_corpus(dir)` IO: corpus dir ->
Comparable cohorts (maps payloads, carries address + registration_date).
validate_epc_prediction.py now just loads + calls the scorer for the component
section and iterates iter_predictions for the calculator-floored end-to-end.
Identical numbers (181 targets, SAP MAE 6.34) — behaviour-preserving.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Heating is the dominant SAP lever (ablating it to actual cut the SAP error
~7 -> ~4.5) yet was entirely unscored. Add the heating group to
compare_prediction's categorical_hits: main fuel / category / control (off
the primary MainHeatingDetail), water-heating fuel / code, has-cylinder,
cylinder insulation, secondary heating (off SapHeating).
Template-copied baseline on the 40-postcode corpus (no predictor change
yet — this just makes the signal visible):
heating_main_fuel 93.4%
heating_main_category 92.7%
water_heating_fuel/code 91.7% / 92.4%
heating_main_control 62.1% <- weak
has_hot_water_cylinder 78.5%
cylinder_insulation_type 35.8% (n=120) <- weak
secondary_heating_type 16.8% (n=125) <- weak
Fuel/category predict well from the template; controls, cylinder, and
secondary heating are poor and now drive the next predictor slices.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ADR-0030 commits Component Accuracy to ~19 categorical components (5 today
+ 8 heating + glazing/renewables). Flat *_correct dataclass fields don't
scale — each needs manual runner wiring. Collapse them into a single
`categorical_hits: dict[str, Optional[bool]]` keyed by component name, which
also matches the runner's name-keyed aggregation (now generic: it tallies
whatever components the comparison reports). No behaviour change; the
classification rates are identical (wall n 578->575 is the 3 certs whose
actual wall is None, now correctly counted as not-applicable via _classify).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The register lists every historical lodgement, so a postcode cohort
contains the same physical address many times (LS61AA: 15 certs / 11
addresses; NG71AA: 15 / 9 — "FLAT 3" appears 3x in each). Two
consequences:
- Production: a re-lodged neighbour was counting up to 3x towards the
cohort mode. select_comparables now dedupes candidates to the latest
cert per address (one comparable per real neighbour) — Comparable
gains address + registration_date (the register metadata its docstring
already anticipated, read straight off the cached payload).
- Validation: leave-one-out leaked — predicting a flat from a near-
identical re-lodgement of itself. The harness now holds out a whole
address (excludes every sibling cert) and evaluates on the latest cert
per address (the best ground truth).
Removing the leak gives the honest numbers (19 distinct addresses):
wall_construction 93.1% -> 89.5%
construction_age_band 65.5% -> 52.6%
roof_construction 79.3% -> 68.4%
floor_area mean|.| 37.9 -> 52.6 m2
The earlier figures were inflated by self-leakage; these are the real
accuracy to beat.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Only main wall_construction was set to the cohort mode; the other
homogeneous categoricals (wall insulation, construction age band, roof
construction, floor construction) were left as template-copied, so one
median-size template's quirks set them. Apply the same cohort-mode
mechanism to all of them per ADR-0029 decision 4 — the template still
supplies geometry, only the categorical codes move to the mode.
Verified mode beats (or ties) template-copy per categorical before
applying. Smoke corpus (29 leave-one-out) classification rates:
construction_age_band 55.2% -> 65.5%
roof_construction 72.4% -> 79.3%
floor_construction 46.2% -> 84.6%
wall_insulation_type 93.1% (tie — already template-strong)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The comparison only scored main wall_construction; everything else the
predictor produces (by template-copy) went unmeasured. Extend
compare_prediction to the rest of the ADR-0029 homogeneous categoricals —
wall insulation type, construction age band, roof construction, floor
construction — and aggregate per-categorical classification rates in the
runner. A categorical hit is "not applicable" (None, excluded from the
denominator) when the actual lodges no value, so absent-roof flats don't
score free wins.
Smoke corpus (29 leave-one-out, all but wall are template-copied today):
wall_construction 93.1%
wall_insulation_type 93.1%
construction_age_band 55.2% <- loud; candidate for cohort-mode
roof_construction 72.4%
floor_construction 46.2% (n=13)
These numbers drive the next slice (extend cohort-mode coverage).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Template (the comparable whose structure/geometry is copied wholesale)
was members[0] — an arbitrary draw from the API search order. With floor
area varying widely within a property_type cohort (NG71AA houses span
51-340 m2), this made the copied geometry noisy and systematically large.
Pick the member whose floor area is closest to the cohort median instead,
implementing ADR-0029 decision 4's unimplemented "closest on size"
criterion while keeping the structure coherent (it is still one real
property, so floor dims / windows / parts stay internally consistent for
the calculator).
Smoke corpus (29 leave-one-out predictions):
floor_area mean|.| 68.0 -> 37.9 m2 (bias +46.8 -> -3.9)
window_area mean|.| 11.1 -> 7.3 m2
parts mean|.| 1.00 -> 0.38
SAP |pred-calc - calc(actual)| MAE 7.19 -> 4.86
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Pure compare_prediction (TDD): wall-construction classification hit + signed
residuals on floor area, window count, total window area, building-parts count.
Plus validate_epc_prediction.py (IO plumbing): drops each cert from its postcode
cohort, predicts from the rest on guaranteed inputs only, aggregates the metrics,
and reports SAP three ways (pred-calc vs lodged / vs calc-on-actual / vs the
neighbour-mean baseline). Smoke run: wall 90.9%, floor-area mean|·| 42.6 m2 (a
real signal — template-copied floor area is noisy), SAP pred-calc edges baseline.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
predict() copies a representative template comparable's structure (coherent for
the calculator), overrides the homogeneous categorical with the cohort mode
(robust to an atypical template), then applies known Landlord Overrides on top
(a known value wins over the estimate). Proven on wall construction; roof/floor/
insulation/age extend on the same mode+override mechanism, driven next by the
validation harness metrics.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Pure-domain select_comparables: property type is an always-hard filter; built
form and known Landlord Overrides (e.g. solid brick) are conditioning filters on
the filter-then-relax ladder — applied while >= minimum_cohort survive, relaxed
otherwise (the mixed-street border case degrades gracefully). PredictionTarget
(known inputs) + Comparable (epc + register metadata) + ComparableProperties
(selected cohort). Weighting (recency x similarity) follows in the synthesis slice.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
