Lands the production code that the just-committed Elmhurst conformance
fixtures (6455d48b) exercise: the SAP10.3 calculator orchestrator
(domain.sap.calculator.Sap10Calculator), the RdSAP-driven cert→inputs
mapper (domain.sap.rdsap.cert_to_inputs), and the EpcPropertyData
strict-type pass that P6.1 starts.
calculator.py is the entry point. Two surfaces depending on the caller's
shape:
- Sap10Calculator().calculate(epc) — full RdSAP mapper + worksheet loop
- calculate_sap_from_inputs(inputs) — pure physics over typed inputs
P6.1 introduces BuildingPartIdentifier as a strictly-typed replacement
for bare-string matching on SapBuildingPart.identifier (motivated by
the pain point at worksheet/dimensions.py:74-82). Two boundary factories
canonicalise raw inputs: from_api_string for the gov-EPC API, and
extension(n) for site-notes / construction id flows.
Also catches up two transitive deps that 6455d48b implicitly required
but I missed:
- ml/rdsap_uvalues.py — party-wall U-value rows that heat_transmission
resolves; the U=0.0 branch the 000516 fixture exercises lands here.
- ml/tests/_fixtures.py — make_minimal_sap10_epc that every Elmhurst
fixture imports. Without this catch-up, checking out 6455d48b in
isolation would ImportError.
Out of scope (will commit separately): ml/transform.py legacy envelope
drift; backend/ FastAPI + documents_parser layer; etl/ scratch.
824 tests pass.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Lands real-cert ground-truth conformance tests for the SAP10.2 worksheet,
asserting our §1 dimensions, §2 ventilation, and §3 heat-transmission
output line-by-line against six Elmhurst-lodged worksheets (000474,
000477, 000480, 000487, 000490, 000516). Each fixture covers a distinct
shape: with/without room-in-roof, single-part vs main+extensions, age
A and B, party-wall U=0.0 vs U=0.25, 1/2/3 sheltered sides, varying
draught-proofing %, and the (12) suspended-timber quirk.
§1/§2/§3 module updates back the new line-refs (LINE_31 external-element
area, LINE_33 fabric loss, LINE_37 total fabric loss; per-fixture (12)
floor / (15) window / (21) shelter-adjusted ach; SapRoomInRoof storey
contribution via the 2.45 m §3.9.1 convention).
The §3 test currently asserts invariants only ((33) = Σ per-element,
(37) = (33) + (36)) because SapRoomInRoof only carries floor_area —
gable/slope/stud/flat-ceiling sub-areas the worksheet itemizes are not
yet modelled. LINE_3* constants capture the worksheet ground truth for
when that gap closes.
Adds a SAP-domain README with a step-by-step guide for adding new
Elmhurst fixtures from the assessor's PDF pair (Summary + worksheet),
including the field-by-field cert → EpcPropertyData mapping table and
the gotchas surfaced across the six fixtures (storey-height +0.25
convention, party-wall U code mapping, has_suspended_timber_floor flag
truth table, (25) effective-ach formula, Energy Rating vs EPC Costs
wind-speed trap).
366 tests pass (was 360 pre-pairs 5-6).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
100-cert probe, seed=7, sap_score window 5..99. MAE 4.29
(vs 8.41 on 2026-05-18 with the older 20..95 window — the
delta blends calculator improvements with sample-window
change, so this is logged as the post-P5 reference, not as
"P5 reduced MAE".)
P5 itself was pure trace exposure; the calculator's SAP
output should be numerically unchanged. The headline finding
from this run is primary-energy over-prediction: PE MAE
44.40 kWh/m², bias +39.66 — now the dominant signal with
SAP residuals halved. Each end-use PE contribution surfaces
on SapResult.intermediate per P5.12, so the next session
can localise the bias without re-instrumenting.
Closes the second half of P5 (HANDOVER_SYSTEMATIC_REVIEW §2.5):
- Adds test_bre_worked_examples.py — one comprehensive test that
locks every published SapResult.intermediate key against its
SAP 10.2 worksheet item number ((4) TFA, (33) fabric heat loss,
(39) HTC, (40) HLP, (73) gains, (93) mean internal temp, (98c)
space heating, (240e/247/250) costs, (252) PV credit, (256)
deflator, (257) ECF, (261-272) per-end-use CO2, (275-287)
primary energy per m²). All formulas derived independently from
the worksheet pages 131-148; passes against the synthetic
100 m² baseline.
- Explicit caveat in module docstring: BRE-published worked
examples don't exist in any of the three SAP-spec PDFs we have
(rdSAP10, SAP10.2, SAP10.3 — all greppped). The test is
spec-formula-derived, not BRE-validated. Structure stays if
BRE numbers surface later; only expected values change.
Also surfaces and documents an RdSAP10 spec drift in
PARITY_FINDINGS.md: Table 32 (page 95 of rdSAP10) gives
Energy Cost Deflator = 0.42, vs the code's 0.36 (SAP10.2 Table 12,
worksheet item (256)). Not changed in P5 — needs ADR-level
resolution on whether the calculator targets SAP10.2 (0.36) or
RdSAP10 (0.42) ratings.
P5 (SapResult.intermediate population + BRE worked-example
fixtures) is now complete on this branch.
Closes the second §11-sketch gap noted in HANDOVER_SYSTEMATIC_REVIEW
("primary energy AND CO2 per end-use"). Lifts the single co2 = total
× factor expression into five named locals (main_heating, secondary,
hot_water, pumps_fans, lighting) and exposes them on `intermediate`.
The five components sum exactly to the top-level co2_kg_per_yr — no
PV deduction in the current implementation.
P5.9 exposed the four primary-energy components as absolute kWh/yr
keys (space_heating_primary_kwh_per_yr, …). HANDOVER_SYSTEMATIC_REVIEW
§11 specifies these as `_pe_kwh_per_m2` because primary energy enters
the rating equation per floor area. Renamed to match the sketch:
- space_heating_pe_kwh_per_m2
- hot_water_pe_kwh_per_m2
- other_pe_kwh_per_m2
- pv_pe_offset_kwh_per_m2
Chain check now verifies max(0, sum − pv_offset) ≈
result.primary_energy_kwh_per_m2 (the top-level per-m² field).
Absolute kWh/yr values remain recoverable via tfa_m2 on `intermediate`.
Final P5 slice. PV credit was the missing term linking the per-end-use
fuel costs (P5.6) to the top-level total_fuel_cost_gbp: total =
max(0, sum(per-end-use) − pv_credit). With this key, every step of
the §13 cost chain — per-fuel cost → PV credit → total → ECF →
rating — is auditable from `intermediate`. P5 trace exposure is
complete.
Promotes _FLOOR_AREA_OFFSET_M2 → FLOOR_AREA_OFFSET_M2 (§13 ECF
denominator, Table 12) and _ECF_LOG_THRESHOLD → ECF_LOG_THRESHOLD
(SAP rating linear/log regime boundary at ECF = 3.5). Together with
the deflator (P5.7) they fully document the §13 rating curve in
trace mode.
Lifts the inlined primary-energy sum into four named components:
space-heating (main + secondary × space_heating PEF), hot water,
other (pumps_fans + lighting × other PEF), and the PV offset at
other PEF (Appendix M). Together with the top-level
primary_energy_kwh_per_yr they make whether the floor-at-zero
clipped visible.
Adds delivered_fuel_kwh_per_yr (sum of all five end-use kWh) and
co2_factor_kg_per_kwh (mirrors the SAP10 input). Together with the
top-level co2_kg_per_yr they make the §15 equation traceable:
co2 = delivered_fuel × factor.
Promotes `_ENERGY_COST_DEFLATOR` to `ENERGY_COST_DEFLATOR` so the
§13 Table 12 constant can be referenced in trace mode alongside the
ECF it scales. ECF mirrors the top-level field; the deflator is the
only fixed worksheet constant the SAP rating depends on.
Per-end-use £/yr costs (main heating, secondary heating, hot water,
pumps_fans, lighting) lifted from the inlined total_cost sum into named
locals and populated on `intermediate`. §12 sweep slices can now diff
each line against the spec (Table 12 unit prices, future Table 12a
fractional blending, Table 12c heat-network DLF) without re-deriving
the cost decomposition.
Behaviour-preserving — `total_fuel_cost_gbp` reconciles bit-for-bit.
136 SAP tests pass.
§9 / Table 9c step 10 output keyed by worksheet name on `intermediate`.
Mirrors the top-level `space_heating_kwh_per_yr` field so spec sweep
slices refer to the worksheet name regardless of field renames.
135 SAP tests pass.
heat_transfer_coefficient_w_per_k (HLC), heat_loss_parameter_w_per_m2k
(HLP), time_constant_h, and the two annual averages
(internal_gains_annual_avg_w, mean_internal_temp_annual_avg_c) populated
on `intermediate`. The averages let sweep slices verify monthly-loop
outputs without re-summing 12 months.
134 SAP tests pass.
infiltration_ach (the cert-derived input) and infiltration_w_per_k
(the derived HLC_V = ACH × volume × 0.33 from SAP 10.2 §4.1) populated
on `intermediate`. Diagnostic surface for the §4 / Table 4g sweep.
133 SAP tests pass.
First slice of P5 trace mode mechanical half (ADR-0010 / handover §11).
SapResult.intermediate: dict[str, float] now exposes worksheet-named
variables for per-section diffing against BRE worked examples and hand
calcs. Dimensions group lands first: tfa_m2, volume_m3, storey_count.
Subsequent slices (P5.2 heat transmission → P5.8 primary energy)
extend the same dict; field defined here so the structural change
lands once and later slices are pure additions.
131 SAP tests pass; 310 packages/domain tests pass.
ADR-0010 §2: the cert-calibration price table was bug-masking
pre-March-2025 SAP values fit against a mixture-distribution of two
spec-version regimes. P2.1 swapped the probe to SAP_10_2_SPEC_PRICES,
P2.2 migrated the golden fixtures, leaving no external consumers.
File deletion is mechanical at this point.
Also updates the cert_to_inputs() docstring at L741-L751: removes the
stale reference to CERT_CALIBRATION_PRICES, points at ADR-0010 and
the Validation Cohort filter as the parity-validation mechanism.
All 152 SAP + ml_training_data tests pass with the file gone.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
ADR-0010 §10: the cert-based fixtures contained compensating errors
under cert-cal prices and are scheduled for replacement by BRE
worked-example fixtures (P5). Until P5 lands they stay as a loose
smoke test catching catastrophic regressions only.
Changes:
- Swap prices=cert_calibration_prices() → prices=SAP_10_2_SPEC_PRICES.
Last external consumer of cert_calibration_prices — P2.3 can now
delete table_12_cert_calibration.py cleanly.
- Loosen tolerance: SAP ±1 → ±5, PE ±10 → ±25. The cert-cal prices
had been numerically tuned around these specific certs, so spec
prices alone produce a -3 to +3 SAP drift across the set.
- Retire 9390-2722-3520-2105-8715 early (heat-network mid-floor
flat). It drifted to SAP residual -7 because cert-cal had absorbed
heat-network DLF + Table 12c interactions. Cert JSON remains in
fixtures/golden/ per ADR-0010 §10; a BRE worked-example covering
the heat-network path will subsume it during P5.
Remaining 6 fixtures pass at ±5 SAP under spec prices. The whole
suite retires when P5 lands.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds ADR-0010 superseding ADR-0009's spec-version target, PCDB
sequencing, and cert-calibration layer. Captures the conclusions
of a grill-with-docs session:
1. Active spec target is SAP 10.2 (14-03-2025), not SAP 10.3 — no
SAP-10.3-lodged certs exist in the corpus to validate against.
2. table_12_cert_calibration is deleted (not "re-derived at the
end"). It was pre-March-2025 spec prices fit against a mixture
distribution of two spec-version regimes, with downstream-
component bugs absorbed into the fit — not Elmhurst deviation.
3. Validation Cohort: filter the corpus to inspection_date ≥
2025-07-01 so every cert in the probe was lodged on SAP 10.2
(14-03-2025) prices. One spec, one signal.
4. PCDB integration is promoted from "Session C deferred" to
prerequisite P4 — dominates residual variance on heat pumps and
the 78% of gas-boiler certs lodging main_heating_data_source=1.
5. Trace mode (SapResult.intermediate) and BRE worked-example
fixtures replace the 7 cert-based golden fixtures, which
contained compensating errors.
6. Strict-type EpcPropertyData via codes.csv-derived canonical
enums (P6) — the in-source motivation lives at
dimensions.py:74-82 (Khalim's comment, included in this commit).
7. Worksheet-faithful structure is a sweep-time principle: each
worksheet module mirrors SAP 10.2 worksheet line numbering.
CONTEXT.md additions:
- Refined "Calculated SAP10 Performance" and "SAP10 Calculation"
to reference SAP 10.2 + ADR-0010.
- New term "SAP Spec Version" — domain-meaningful because the
same EpcPropertyData yields different sap_score under different
spec revisions.
- New term "Validation Cohort" — the version-locked sub-corpus.
HANDOVER_SYSTEMATIC_REVIEW.md is rewritten section-by-section to
reflect ADR-0010: §1 framing, §2 status pointer, new §2.5 with the
six prerequisites P1–P6 in dependency order, §3 diagnosis (cert-cal
was stale prices, not Elmhurst deviation), §4 scope (PCDB IN,
SAP 10.3 stays OUT), §5 approach (worksheet-faithful principle as
§5.5), §7 tension dissolved, §7b findings re-framed, §8 dead-ends
re-classified as conditional, §9 cohort filter, §10 fixture
strategy, §11 trace mode as prerequisite, §12 prereqs-first,
§13 Phase 0/Phase 1 workflow, §14 ADR-0010 reference, §15 final
note.
P2.1 (commit ac1aa56a) already lands the first ADR-0010 slice
(probe swap to spec prices).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
ADR-0010 P2: cert-calibration layer is deleted, the probe uses
SAP_10_2_SPEC_PRICES (already defined in cert_to_inputs.py). Extracts
a pure predict_sap_for_cert(cert_document, *, prices) -> int helper
out of main()'s inline pipeline so the spec-prices path is unit-
testable in isolation; the helper is also reusable for P3's cohort-
filtered probe variant.
The pinned regression value (SAP=67 for cert 6035-7729 under spec
prices, vs the cert's lodged SAP of 73 under cert-cal prices) lives
in services/ml_training_data/tests/unit/test_sap_parity_probe.py.
It will drift as P4 (PCDB) and the section sweep land their fixes;
that's expected.
cert_calibration_prices is still imported by test_golden_fixtures.py
and the table_12_cert_calibration module is intact. P2.2/P2.3 retire
those.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
§7b "Outstanding findings to pick up during the systematic pass"
collects spec-correct fixes that were reverted because they regressed
SAP MAE against the corpus — but the spec basis is unambiguous and
they WILL be the right answer once cert-calibration is re-derived.
Treat as TODOs, not dead-ends. Documents:
Finding 1 — HW cylinder zero-loss for combi (PE MAE -6.64 measured)
Finding 2 — Standing charges Table 12 note (a)
Finding 3 — Cat=10 room-heater Table 12a fractional blending
Finding 4 — Lighting Appendix L proper (L1-L12 cascade)
Finding 5 — Internal-gains Table 5 water-heating + losses rows
Finding 6 — Storage-loss-factor table values 3× off spec
Finding 7 — Heat-pump fallback (needs PCDB)
Finding 8 — Smaller gaps carried forward
Each documents the spec section/page reference, the current code
bug, empirical impact where measured, and when to pick up during the
section-by-section sweep.
PCDB section strengthened from "deferred to Session C" to an explicit
roadmap: data source URL, lookup key (main_heating_index_number),
fields needed, recommended sequencing (after spec sweep so cert-cal
is re-derivable), and why-not-now (cert-cal currently masks PCDB gaps).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The slice-by-slice "fix the biggest residual" approach has hit a
ceiling at SAP MAE ~4.6 because the cert-calibration prices absorb
multiple structural deviations from spec. Any spec-correct fix in one
component breaks the calibration for others. Three failed slices this
session (standing charges, cat=10 routing, combi zero-loss) made the
pattern unambiguous.
Pivot: systematic section-by-section spec verification. Read the
RdSAP 10 + SAP 10.2 spec in order, check each table / formula /
footnote against the corresponding code, fix gaps one at a time.
Build the spec-correct engine first; re-derive cert-cal calibration
once at the end as a thin Elmhurst-compatibility layer.
Handover doc covers:
- Critical framing (deterministic, not assessor judgement)
- Current state (SAP MAE 4.61, PE MAE 43.32 at f4a8d2a0)
- Why the slice-by-slice approach won't converge
- Scope decisions (RdSAP 10 + SAP 10.2 only; park full-SAP + PCDB)
- Section-to-code mapping
- Known dead-ends to skip
- Cert-calibration vs spec-correctness tension and how to resolve it
- The 7 golden fixtures and their compensating-error caveats
- Trace mode recommendation (ADR-0009's `intermediate` field)
- Specific §1-3 starting tasks
- Workflow recap
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Pins 7 certs from a 1000-cert random sample that satisfy:
|SAP rounded-int residual| ≤ 1
|PE residual| ≤ 10 kWh/m²
main_heating_category != 4 OR main_heating_data_source != 1
(non-PCDB-heat-pump — PCDB lookup is deferred)
Cert mix: 6 cat=2 gas/oil boilers (3 PCDB, 3 Table 4b) + 1 cat=6 heat
network. Age bands A, C, D (×3), F, J, L. TFAs 75-526. Mix of
detached / semi-detached / mid-terrace / mid-floor flat. The cleanest
PE match in the set (cert 7536-3827) has PE residual -0.29 kWh/m².
Purpose: regression anchor. Future slices that improve aggregate MAE
silently break individual certs unless caught here. Each cert's
expected residual is recorded in `_EXPECTATIONS` so the diff is
human-inspectable when a regression fires.
The set is acknowledged to contain compensating-errors cases: some
certs match SAP within ±1 because the cert-calibration prices absorb
multiple structural deviations from spec. Hand-trace of 7536-3827
showed PE matched (-0.29) but cost was £143 (12%) under cert's implied
cost — a multi-factor gap (price calibration + missing gas standing
charge + lighting over-prediction) that cancels back into SAP ±1. We
accept this with the tolerance choice: tightening to PE ±5 in our
sample would have yielded zero fixtures.
Tolerance can tighten over the session as we close the PE bias
(currently +38 kWh/m² systematic).
All 301 domain tests pass; no behaviour changed.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Heat-network certs (cat=6) were under-predicted in cost — SAP bias
+6.31 across 13 sample certs, PE bias -15.6 (we under-predicted PE).
Root cause: missing distribution-loss-factor application.
SAP 10.2 spec references:
- Table 12 note (k): "Cost is per unit of heat generated (i.e.
before distribution losses); emission and primary factors are per
unit of fuel used by the heat generator."
- §C3.1: "Where a heat network is listed in the PCDB, the DLF is
already factored into the cost, CO2 and PE factors recorded
therein, so a DLF of 1 should be entered in worksheet (306) to
avoid double counting." (Implication: non-PCDB networks MUST
apply DLF.)
- Table 12c (p. 193): DLF by age band, 1.20 (A pre-1900) →
1.50 (K+ 2007+).
- RdSAP 10 §10.11 Table 29 cross-references Table 12c.
Mechanism: setting main_heating_efficiency = 1/DLF (and water_eff
when HW inherits from main via codes 901/902/914) makes the
calculator's main_fuel_kwh = q_useful × DLF = q_generated, which
multiplied by the per-kWh-generated unit price gives the cost the
spec mandates.
Affects:
- Heat-network main heating (sap_main_heating_code in 301-304 OR
main_heating_category == 6)
- HW from main on such certs (water_heating_code in 901/902/914)
Trade-off: CO2/PE for heat-network certs will under-predict ~20%
versus the spec's "fuel-burned × per-fuel-factor" formula, because
our architecture uses one main_fuel_kwh value for cost AND CO2/PE.
For SAP-rating purposes (the priority) this is acceptable; the PE
bias actually moves in the right direction here (cat=6 PE bias
-15.6 → -5.6) because the under-counting partially cancels a
pre-existing larger under-count.
Parity probe at 300 certs, seed=7:
SAP MAE 4.69 → 4.61 (-0.08)
SAP bias 0.98 → 0.87 (-0.11)
PE MAE 43.32 → 43.11 (-0.21)
cat=6 PE bias -15.6 → -5.6 (+10.0, correct direction)
cat=6 PE MAE 40.3 → 35.8 (-4.5)
cat=6 our_pe 158.5 → 225.0 (cert 230.6 — converged)
Cumulative across S-B23 → S-B31:
SAP MAE 5.34 → 4.61 (-0.73)
PE MAE 57.28 → 43.11 (-14.17)
PE bias 51.56 → 38.64 (-12.92)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
An attempted slice (S-B30, not committed) hypothesised that
`main_heating_fraction=1` on the cert meant "no secondary heating" and
overrode Table 11's 10% default. Probe at 300 certs penalised it:
SAP MAE 4.69 → 4.85, SAP bias 0.98 → 1.61. The hypothesis was wrong
and I should have read the spec before coding.
SAP 10.2 Appendix A1 (p. 43) defines `main_heating_fraction` as the
allocation between TWO main heating systems when both exist; not as
the main-vs-secondary fraction. 99% of corpus certs have =1, meaning
"single main, 100% allocation".
SAP 10.2 Appendix A4(d) (p. 45) is explicit: "If any fixed secondary
heater has been identified, the calculation proceeds with the
identified secondary heater" and "Table 11 gives the fraction of the
heating that is assumed to be supplied by the secondary system" —
no override based on main_heating_fraction.
Adds:
- Regression test pinning the spec behaviour
(test_main_heating_fraction_does_not_override_table11_secondary_default)
- Regression test for the already-spec-aligned fallback path
- _secondary_fraction docstring explaining why main_heating_fraction
is NOT consulted (with reference to the failed attempt)
- secondary_heating_type kwarg on make_sap_heating (test-only, was
missing — needed to construct the regression fixture)
Probe at 300 certs unchanged from prior baseline:
SAP MAE 4.69, bias 0.98
PE MAE 43.32, bias 37.69
The hand-trace finding that cert 9036-0827 over-predicts cost remains
real, but the secondary-heating fraction is per-spec. The residual
~£33 gap on that cert is most likely missing PCDB efficiency lookup
(cert has main_heating_data_source=1 and index_number=10241 — PCDB
data — and we fall back to category-default 0.80 vs typical PCDB-
listed condensing-boiler 0.90+). Deferred to Session C per ADR-0009.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Parallel of S-B24 (walls) for the other envelope elements. Full-SAP
assessments lodge a measured/calculated U-value directly in the
description ("Average thermal transmittance X W/m²K") for floors
(~1 391 corpus certs) and roofs (~1 140 certs). Per spec:
- §5.11 (roofs) opening clause defers to assessor's value when
present
- §5.12 (floors): "Unless provided by the assessor the floor
U-value is calculated according to BS EN ISO 13370"
Both u_floor and u_roof now invoke `_measured_u_from_description`
first; if it parses a value, they return it directly and skip the
cascade. No range cap (consistent with S-B24 design — calculator
mirrors what the assessor lodged).
Parity probe at 300 certs, seed=7: headlines unchanged (same parquet
sampling gap as S-B24 — full-SAP certs filtered out upstream). Slice
correctness proved by:
- 1 unit test for u_floor measured-U parse
- 1 unit test for u_roof measured-U parse
- existing 287 tests passing, no regressions
A bulk-zip-based probe to measure the corpus-wide impact remains the
needed tooling investment (see S-B24 commit message).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
346 corpus certs lodge roof_insulation_thickness="NI" (Not Indicated,
parsed to 0 by _parse_thickness_mm). When the description also signals
retrofit insulation ("Pitched, insulated (assumed)" / "Flat,
insulated" / "Roof room(s), insulated (assumed)"), our cascade
returned the uninsulated Table 16 row-0 value (U=2.30).
RdSAP 10 §5.11.4 (page 44, end of section): "If retrofit insulation
present of unknown thickness use 50 mm". That maps to Table 16 row
"Insulation at joists at ceiling level, 50 mm" = 0.68 W/m²K. The fix
is the analog of S-B27 for roofs: when insulation_thickness_mm==0
(the "NI" sentinel) and _described_as_insulated(description), return
0.68 instead of the row-0 lookup.
Per-cert delta: ΔU = 1.62 W/m²K on the affected slice; for typical
80 m² roof = 130 W/K HLC reduction ≈ 12 kWh/m² PEUI per cert.
Parity probe at 300 certs, seed=7:
SAP MAE 4.72 → 4.69 (-0.03) ← first SAP MAE drop in 3 slices
PE MAE 44.19 → 43.32 (-0.87)
PE bias 38.56 → 37.69 (-0.87)
Cumulative across S-B23 → S-B28:
PE MAE 57.28 → 43.32 (-13.96)
PE bias 51.56 → 37.69 (-13.87)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The cert's `floor_insulation_thickness` field carries "NI" (Not
Indicated) on 58% of corpus certs — by far the most common value. For
~2 413 of those (12% of corpus) the description also says "Solid,
insulated (assumed)" or "Suspended, insulated (assumed)" — the
assessor saw insulation but didn't measure the thickness. Our
`_parse_thickness_mm("NI")` returns 0, which feeds `u_floor` as an
explicit "0 mm" → r_f=0 → uninsulated-floor U-value. Wrong.
RdSAP 10 §5.12 Table 19 footnote (2) (page 46): "For floors which
have retrofitted insulation, use the greater of 50 mm and the
thickness according to the age band". `u_floor` now accepts a
`description` kwarg; when `_described_as_insulated(description)` is
true and the lodged thickness is missing/zero, ins_mm =
max(50, age-band default).
Geometry sanity-check, 100 m² × 40 m perimeter, w=0.3 (B=5):
- Uninsulated solid floor: d_t = 0.615, U = 0.60 W/m²K
- 50 mm assumption: d_t = 2.758, U = 0.31 W/m²K
Parity probe at 300 certs, seed=7:
PE MAE 45.37 → 44.19 (-1.18)
PE bias 39.75 → 38.56 (-1.19)
Band J bias +41.2 → +29.7 (-11.5)
Band K bias +34.1 → +22.4 (-11.7)
Band L bias +19.6 → +11.3 (-8.3)
Band M bias +86.3 → +55.1 (-31.2)
Bands A-H mostly unchanged (max(50, 0) = 50 either way; description
overrides on older stock are rarer in this sample)
The K-L-M dwellings improved most because for them the age-band
default insulation (100-140 mm) is now applied instead of 0 mm.
Cumulative across S-B23 → S-B27:
PE MAE 57.28 → 44.19 (-13.09)
PE bias 51.56 → 38.56 (-13.00)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Two related bugs both produced U=1.7 for retrofit-insulated solid-brick
walls when the spec says U=0.55 (Table 6 footnote: "If a wall is known
to have additional insulation but the insulation thickness is unknown,
use the row in the table for 50 mm insulation"):
1. _insulation_bucket(0, True) returned 0 instead of 50. The "NI"
sentinel parses to 0 via _parse_thickness_mm, then the bucket
function's "< 25 -> 0" branch ignored the insulation_present signal.
Affects 56 corpus certs lodging solid-brick with type=1 or type=3
plus thickness="NI".
2. wall_ins_present was set False whenever wall_insulation_type == 4
("as-built / assumed"), even if the description said
"...insulated (assumed)" or "...partial insulation (assumed)".
Affects 128+51 = 179 corpus certs.
The same root pattern as S-B25 (cavity-wall description disambiguation),
extended to non-cavity constructions. `_cavity_described_as_filled`
generalised to `_described_as_insulated`; now used by:
- u_wall (cavity-wall dispatcher to the Filled-cavity row, S-B23/B25)
- heat_transmission_from_cert (override wall_ins_present for non-cavity
walls so the 50 mm bucket routes per Table 6 footnote)
Parity probe at 300 certs, seed=7:
PE MAE 45.74 → 45.37 (-0.37)
PE bias 40.19 → 39.75 (-0.44)
Band D bias +42.7 → +41.6 (-1.1)
Band F bias +12.6 → +10.7 (-1.9)
Modest aggregate movement — the affected population is small (~0.6% of
corpus, ~2 certs in the 300 sample). The slice's correctness is proved
by 4 unit tests in test_rdsap_uvalues.py + 2 end-to-end tests in
test_heat_transmission.py.
Cumulative across S-B23 → S-B26:
PE MAE 57.28 → 45.37 (-11.91)
PE bias 51.56 → 39.75 (-11.81)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The RdSAP schema's `wall_insulation_type = 4` ("as-built / assumed")
covers two distinct cert populations that previously both routed to
the Cavity-as-built row (U=1.5 at band E):
686 certs: "Cavity wall, as built, no insulation (assumed)" — U=1.5 ✓
1171 certs: "Cavity wall, as built, insulated (assumed)" — should be 0.7
147 certs: "Cavity wall, as built, partial insulation (assumed)" — 0.7
The description string disambiguates. The legacy production map at
recommendations/rdsap_tables.py:753 routes the latter two to "Filled
cavity" — we match that interpretation here for parity with the cert
assessor and the production recommendation engine.
`_cavity_described_as_filled` adds the description check; the existing
filled-cavity dispatcher in u_wall now fires on either signal:
- wall_insulation_type == 2 (S-B23 — explicit filled-cavity code)
- description contains "insulated" or "partial insulation" without
the "no insulation" negation marker (S-B25 — assumed cavity-fill)
Parity probe at 300 certs, seed=7:
PE MAE 46.78 → 45.74 (-1.04)
PE bias 41.78 → 40.19 (-1.59)
Band F bias +23.2 → +12.6 (-10.6)
Band G bias +31.8 → +25.1 (-6.7)
Band H bias +30.7 → +15.5 (-15.2)
Improvements localise to bands F-H (1976-1995), the era when Building
Regs mandated cavity insulation for new-builds — making "as built,
insulated (assumed)" the modal description. SAP MAE drifted up
+0.12 (cost-side residuals surfacing now that envelope is closer to
spec; tracked for follow-up).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Full SAP assessments (~15% of corpus, 4 403 of 30 000 scanned bulk-zip
certs) lodge a measured/calculated wall U-value per BS EN ISO 6946 in
walls[i].description, e.g. "Average thermal transmittance 0.18 W/m²K".
These certs typically have wall_construction, wall_insulation_type and
construction_age_band all None, which the cascade defaults previously
resolved to U = 1.5 (uninsulated cavity at band E). RdSAP 10 §5.3:
"U values are obtained from … the construction type, date of
construction and, where applicable, thickness of additional insulation"
— but a measured value supersedes the cascade.
Corpus U-value distribution among parsed:
median 0.21, mean 0.225, range 0.06-1.84
80% at U ≈ 0.2 (Part L-compliant new-builds)
10% at U ≈ 0.1 (passivhaus / very low)
7% at U ≈ 0.3 (older retrofitted full-SAP)
3% in the tail (conversions, edge cases)
Per affected cert (100 m² new-build at U 1.5 → 0.21):
walls_w_per_k drops 129 → 21 W/K
PEUI drops ≈ 120 kWh/m²
Implementation:
- _measured_u_from_description() regex-parses the phrase from the wall
description; returns None on no-match or non-numeric so the cascade
fall-through is preserved.
- u_wall checks the measured value FIRST, before any cascade logic.
- No range cap — calculator mirrors what the assessor lodged, per the
"deterministic except for input errors" principle. Parse failure
falls through cleanly.
Parity probe at 300 certs, seed=7: headlines unchanged. Direct check
on the sample: 0/300 certs carry an "Average thermal transmittance"
description. The v18a parquet filters full-SAP certs out somewhere
upstream, so this slice is invisible in the parquet-based probe. The
slice's correctness is proved by:
- 4 unit tests in test_rdsap_uvalues.py (tracer + regression on
ordinary descriptions + parse-failure fallback + filled-cavity
description still routes correctly)
- 1 end-to-end test in test_heat_transmission.py exercising a
synthetic full-SAP cert through heat_transmission_from_cert
- All 274 domain tests passing, no regressions
Follow-up tooling: a bulk-zip-based parity probe that doesn't filter
to the parquet's subset is needed to measure this slice's corpus
impact. Separate dig.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Previous bound (20, 95) excluded full-SAP new-builds (sap_score 90+,
which carry the dramatic wall U-value gap) and deepest-tail heritage
certs (sap_score ≤ 20). Widening so the sample reflects the
populations where the calculator's biggest spec gaps live.
New baseline at 300 certs, seed=7:
SAP MAE 5.34 → 4.59 (-0.75)
PE MAE 48.99 → 46.78 (-2.21)
PE bias 42.07 → 41.78 (-0.29)
Note: the v18a parquet only contains ~0.7% certs with age_band=None,
while the raw bulk zip has 15% full-SAP "Average thermal transmittance"
certs. The parquet is filtering them somewhere upstream — to be chased
in separate work. Until then, parity-probe MAE will under-show the true
corpus impact of slices that target full-SAP certs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The cert encodes filled-cavity walls as
(wall_construction=4 cavity, wall_insulation_type=2 filled,
wall_insulation_thickness="NI"). The previous cascade parsed "NI"→0
and ran the thickness-bucketed table, returning U=1.5 (the
"Cavity as built" row) — treating retrofit-filled cavities as if they
were uninsulated. Spec (RdSAP 10 Table 6, page 33) has a dedicated
"Filled cavity" row at U=0.7 for bands A-E, 0.40 at F, 0.35 at G-H,
and "as built" from band I onward.
Adds:
- WALL_INSULATION_FILLED_CAVITY constant (code 2 per RdSAP schema,
confirmed empirically on 8 000 corpus certs against walls.description)
- _CAVITY_FILLED_ENG row in domain.ml.rdsap_uvalues
- dispatcher in u_wall when (construction=cavity, insulation_type=2)
- wall_insulation_type plumbing through heat_transmission_from_cert
Parity probe (300 certs, seed=7) before → after:
- PE MAE 57.28 → 48.99 (-8.3)
- PE bias 51.56 → 42.07 (-9.5)
- Band C bias +65.3 → +47.8 (-17.5)
- Band D bias +67.9 → +45.7 (-22.2)
- Band E bias +77.0 → +58.8 (-18.2)
- Band F bias +43.8 → +25.4 (-18.4)
- Band K-L bias unchanged (filled-cavity row falls back to as-built
from band I onward per spec footnote; correct no-op)
Future slices already lit up by the same enumeration:
- type=1 external / type=3 internal insulation rows (~440 certs)
- type=6 filled + external / type=7 filled + internal (~22 certs)
- type=None "Average thermal transmittance X W/m²K" string parse
(1 358 certs — biggest follow-up)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds primary-energy breakdown (space heating, hot water, lighting,
pumps, PV) per cert plus stratified bias reports by main_heating_
category, construction_age_band, and dwelling_type. Used to localise
the +51 kWh/m² PEUI bias to envelope-side over-prediction on pre-1996
fabric, which the bare SAP-residual ranking didn't surface.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Per user suggestion: the iteration history in this chat has likely
accreted blind spots that a long context window can't shed (e.g. I
spent slices comparing our delivered kWh to the cert's primary kWh
without noticing the apples-to-oranges error). A fresh agent reading
the SAP 10.2 + RdSAP 10 PDFs cold against the current calculator may
spot gaps faster.
HANDOVER_FRESH_REVIEW.md gives the fresh agent:
- Current state (MAE 5.34, primary-energy bias +51 kWh/m²)
- Repo layout pointer
- Priority-ordered dig list (PEUI mystery first)
- Validated truths
- Dead-end list (don't repeat S-B5 NI thickness switch etc.)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Wires SAP 10.2 Table 12 "Primary energy factor" column into Table 12
helpers and onto CalculatorInputs as three per-end-use factors (space
heating, hot water, other). calculate_sap_from_inputs now emits
primary_energy_kwh_per_yr and primary_energy_kwh_per_m2 on SapResult,
matching the cert's `energy_consumption_current` field (PEUI).
Triggered by a decomposition that revealed I'd been comparing our
delivered energy to the cert's primary energy — apples to oranges.
With proper primary-energy comparison the actual finding is:
300-cert primary-energy diff (cert calibration prices):
energy MAE: 57.3 kWh/m²
energy bias: +51.6 (we over-predict by ~50%)
energy P50: +49.5
This is a much bigger systemic bug than the SAP MAE 5.34 suggested.
Closing it requires investigating either (a) demand model
over-prediction, (b) HW losses, (c) PEF values per fuel, or (d) cert
reporting convention differences. Targeted for the next context.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Per SAP 10.3 §2 worksheet line 22 / RdSAP10 §4.1: effective infiltration =
raw_ACH × (1 - 0.075 × sheltered_sides). Default 2 sheltered sides for
typical UK terraced/semi-detached layout (the cert doesn't lodge a
sheltered-sides count, so we apply the spec's typical default).
infiltration_ach() gains a `sheltered_sides` kwarg defaulting to 0
(spec-pure intermediate result; existing unit tests keep that contract).
cert_to_inputs passes sheltered_sides=2.
Found via energy decomposition: our predicted total energy was running
+15.7 kWh/m² over cert (10% over) — wind shelter knocks ~15% off
infiltration, contributing to closing that gap.
300-cert parity probe:
MAE 5.43 → 5.34 (-0.09)
bias -0.52 → +0.29 (back near zero)
within ±10: 86.3% → 86.7%
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
SAP 10.2 Table 11 allocates a fraction (10-20%) of space heating to a
secondary system based on main heating category. Per Appendix A §A.2.2,
this is applied:
- Always for electric storage heater main systems (codes 401-407, 409,
421); a portable electric heater (code 693) is defaulted when no
secondary is recorded.
- Otherwise only when the cert lodges a secondary_heating_type.
Calculator gains secondary_heating_fraction, secondary_heating_efficiency,
secondary_heating_fuel_cost_gbp_per_kwh on CalculatorInputs and a
secondary_heating_fuel_kwh_per_yr on SapResult. Monthly loop splits
demand: q_main = q_heat × (1 - frac), q_secondary = q_heat × frac, each
converted to fuel via its own efficiency. Cost = main_kwh × main_price
+ secondary_kwh × secondary_price + ... .
Initial implementation applied 10% unconditionally and regressed 300-
cert MAE 5.45 → 6.58 (bias -2.65). Restricted to the conditional rule
above and aggregate returns to flat:
300-cert: MAE 5.45 → 5.43 (flat)
bias +0.22 → -0.52
within ±5: 62.7% → 64.3%
The slice is spec-correct and architecturally enables the secondary-
heating channel; aggregate MAE moves are small because most certs
don't lodge a secondary and most non-storage mains don't force one.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Per user suggestion (switch from probe-driven to worksheet-driven
iteration), enumerates the §§1-15 worksheet + Appendices A-U state in
the calculator with a status grade and a prioritised gap list. Becomes
the roadmap for Session B remaining slices.
Next slice from this list: Table 11 secondary heating allocation —
10% fraction on most boiler-main certs that we currently model as 0.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Wires photovoltaic_arrays into the calculator as a per-kWh cost credit
against the ECF numerator. Total annual PV kWh = sum(peak_power_kw)
× 850 (UK-average yield per Appendix M, single national figure since
ratings use UK-average weather per S-B18). Credit rate is Table 12
code 60 (PV export tariff) — 5.59 p/kWh under SAP spec prices, 13.19
p/kWh under cert-calibration prices.
This is the first slice from the worksheet-driven phase (per user
suggestion). PV was identified as a clear systemic gap that probe-
driven iteration hadn't surfaced because only ~5-10% of certs have
PV and the corpus probe is biased toward the most-frequent shapes.
100-cert: MAE 4.39 → 4.49 (small regression; bias -0.17 → -0.07)
300-cert: MAE 5.44 → 5.45 (essentially flat; bias 0.11 → 0.22)
Net spec-correct, aggregate MAE neutral. The certs that DO have PV
should see the right cost story now; ML residual will pick up the
fidelity gap (no orientation/overshading/pitch on our yield).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
SAP 10.2 Appendix U explicit rule: "Calculations for fabric energy
efficiency (FEE), regulation compliance (TER and DER, TPER and DPER)
and for ratings (SAP rating and environmental impact rating) are done
with UK average weather. Other calculations (such as for energy use and
costs on EPCs) are done using local weather."
Our calculator was using the cert's region_code for everything. Spec
mandates region 0 (UK average) for rating outputs. Net MAE neutral on
the 100-cert sample (most certs sit close to UK average) and on the
300-cert sample but it's spec-correct, and aligns with what the cert
assessor's SAP rating actually computes.
Found by switching from probe-driven to worksheet-driven iteration —
per user suggestion this is the more efficient mode once the easy
wins from probe-driven have been extracted.
100-cert: MAE 4.39 (unchanged)
300-cert: MAE 5.44
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Refines S-B16 with a fuel-conditional rule for the Unknown tariff code
(RdSAP energy_tariff=3): all-electric dwellings whose meter_type the
assessor couldn't pin down are almost always E7-eligible (gas dwellings
default to Single). For non-electric end-uses (gas main heating), the
meter_type doesn't affect cost, so Unknown stays standard for them.
Hand-trace confirmation: 3 of the 4 worst residuals (0800-1364,
0036-1125, 0340-2394) all have meter_type=3 AND electric main fuel —
applying off-peak to these recovers the parity loss S-B16 introduced.
100-cert parity probe:
MAE 5.04 → 4.39 (recovered to S-B15 best state)
bias -1.20 → -0.17
within ±10: 93% → 96%
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Confirmed against the official RdSAP enum in
datatypes/epc/domain/epc_codes.csv:
1 = dual (off-peak / Economy-7)
2 = Single (standard tariff)
3 = Unknown (verified against Elmhurst assessor software:
treated as Single)
4 = dual (24 hour) (off-peak)
5 = off-peak 18 hour (off-peak)
Different from the SAP-Schema enum (1=standard / 2=off-peak) — the
transform.py docstring referenced the SAP enum, not RdSAP. Our corpus
is RdSAP so we use the RdSAP codes.
This locks in the meter_type-based tariff selection from S-B15 with
the authoritative enum, replacing the earlier heating-code heuristic.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Per user guidance: trust the cert's lodged meter_type as the source of
truth for tariff selection, rather than inferring tariff from heating
code lists. SAP10 meter_type enum (verified empirically on the 250k
corpus: 75% type 2, 14% type 1, 11% type 3):
1 = Off-peak (Economy-7 / dual rate)
2 = Single (Standard)
3 = Off-peak (24-hour heating)
The transform.py docstring describes 1=Standard / 2=Off-peak but that
contradicts the 75% type-2 distribution (UK demographics don't put 75%
of dwellings on off-peak). The inverted reading parity-tests correctly.
Tariff routing rules:
- Space heating: off-peak rate when main fuel is electric AND meter is
off-peak; else standard main-fuel rate.
- Hot water: off-peak rate when water fuel is electric AND meter is
off-peak; else water-fuel rate.
- Lighting + pumps + fans: always standard electricity (Table 12a
notwithstanding — cert software empirically uses standard here).
100-cert parity probe:
MAE 4.40 → 4.39 (flat in aggregate; structurally cleaner code)
RMSE 5.63 → 5.56
bias +0.16 → -0.17
within ±10: 96% (unchanged)
The meter_type seam replaces the e7_eligible_main_codes set on
PriceTable. Conceptually cleaner: tariff is a property of the meter,
not the heating system.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Hand-trace cert 0340-2394 (73m² mid-floor flat, code 691 electric room
heater, actual SAP 75, predicted 56) confirmed the cert software
applies off-peak rates to electric room heaters when the dwelling has
the E7-tariff hallmarks (electric immersion HW cylinder). Extending
the cert-calibration E7-eligible set from {191-196, 401-409, 421-425}
to add {691-696}.
100-cert parity probe:
MAE 4.48 → 4.40 (-0.08)
RMSE 5.81 → 5.63
bias -0.52 → +0.16 (essentially centered)
within ±10: 95% → 96%
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Water heating codes 907 (single-point gas) and 909 (electric instantaneous)
describe no-cylinder, point-of-use systems with no primary circuit.
The predicted_hot_water_kwh model was adding 366 kWh cylinder-storage
loss + 245 kWh primary-pipework loss on top of useful demand for these
certs — over-counting HW by 600+ kWh.
Discovered hand-tracing cert 2903-8339 (11m² Top-floor flat studio,
water_heating_code=909, actual SAP 75, predicted 55).
100-cert parity probe:
MAE 4.53 → 4.48 (-0.05)
RMSE 5.96 → 5.81
bias -0.57 → -0.52
Smaller MAE delta than S-B12 because instantaneous-HW certs are a
smaller subset, but the affected dwellings are exactly the worst-
residual tail.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The legacy water_heating_efficiency(901, main_code) returns 0.80 (gas
boiler default) when sap_main_heating_code is None — even if the main
system is a heat pump (category=4, efficiency 2.30). For "from main
system" water codes (901/902/914), we must inherit through the FULL
main-heating cascade including the category fallback.
Discovered by hand-tracing cert 0320-2850 (Semi-detached bungalow,
heat-pump main with no SAP code lodged, actual SAP 70, predicted 49).
HW was being charged at 0.80 eff for a 2.30-eff dwelling — 2.9× too
much HW fuel.
100-cert parity probe:
MAE 4.66 → 4.53 (-0.13)
RMSE 6.27 → 5.96
bias -0.70 → -0.57
within ±10: 94% → 95%
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Hand-tracing cert 0800-1364 (Detached bungalow, code 191/direct-electric,
actual SAP 71, predicted 37) showed the cert assessor applies off-peak
rates to direct-electric main heating despite SAP 10.2 Table 12a
specifying 90% high-rate. Adds e7_eligible_main_codes to PriceTable so
each price source carries its own rule:
- SAP_10_2_SPEC_PRICES: {401-409, 421-425} (storage only, per Table 12a)
- CERT_CALIBRATION: {191-196, 401-409, 421-425} (empirically what
the cert software does)
100-cert parity probe:
MAE 4.99 → 4.66 (recovered to pre-S-B9 best state)
bias -1.03 → -0.70
within ±1: 23% → 24%
within ±3: 47% → 48%
within ±10: 93% → 94%
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
