junte/Model
1
0
Fork 0
mirror of https://github.com/Hestia-Homes/Model.git synced 2026-07-12 13:29:04 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

docs(corpus): find the second driver of 100061275133's residual gap

Browse source 
Walked the worksheet section-by-section (dimensions, ventilation, fabric,
internal/solar gains, MIT, space-heating requirement, fuel cost) against
the calculator's matching intermediates. Internal gains and heat loss
matched closely, but solar gains (worksheet line 83) was 2.2x off in
January (56.9 W vs our 123.4 W) — traced to the Elmhurst BUILD entering
all 12 real windows (3 orientations: NW 7.05 m2, NE 1.21 m2, SE 7.06 m2)
as a single synthetic 15.32 m2 "North" group, understating solar gain.

Verified by reproducing both build artifacts in the calculator: real
windows + real Table-11 gating = 76.33 (our answer, matches lodged 76
exactly); collapsing to a synthetic single-North window + forcing Table
11 on = 72.62, reproducing Elmhurst's 72 almost exactly. Confirms neither
residual is a mapping or calculator bug — one is a deliberate,
regression-tested spec choice (Table 11) and the other is an artifact of
how this specific Elmhurst reconciliation build was entered.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Jun-te Kim 2026-07-03 09:52:22 +00:00
parent f69d345493
commit d15ea8b7d4
1 changed files with 15 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
tests/domain/sap10_calculator/test_real_cert_sap_accuracy.py
View file

@ -898,9 +898,21 @@ _EXPECTATIONS: Final[tuple[RealCertExpectation, ...]] = (
# note a few lines below: applying Table 11 unconditionally regressed
# the full 1000-cert corpus MAE by +0.16 (lodged software doesn't
# apply Elmhurst's blanket convention). Verified: forcing Table 11 on
# for just this cert moves 76.33 -> 73.62, ~62% of the 76.33-vs-72 gap
# — confirms this is the dominant driver, not a mapping/entry error.
# PINNED to the observed exact match (vs lodged, the correct target).
# for just this cert moves 76.33 -> 73.62, ~62% of the 76.33-vs-72 gap.
# The remaining ~1.6 pts is a second, separate cause: (6) Solar gains —
# the BUILD entered all 12 windows as one 15.32 m2 "North" group
# (E.set_single_window), but the cert's real sap_windows lodge 3
# orientations (NW 7.05 m2, NE 1.21 m2, SE 7.06 m2 incl. the largest
# single window) — collapsing to due-north understates solar gain
# (Elmhurst Jan (83)=56.9 W vs our real-orientation 123.4 W). Verified:
# forcing our engine to ALSO use a single synthetic North window (i.e.
# replicating the build's simplification, not a cert or engine fact)
# plus the forced Table 11 above reproduces 72.62 -- matches Elmhurst's
# 72 almost exactly. Conclusion: the calculator is faithful (it already
# matches lodged exactly using the cert's REAL window orientations);
# both residuals are Elmhurst-worksheet-reconciliation artifacts (a
# deliberate spec choice + a build-entry simplification), not mapping
# or calculator bugs. PINNED to the observed exact match vs lodged.
RealCertExpectation(
schema="RdSAP-Schema-21.0.1",
sample="uprn_100061275133",