junte/Model
1
0
Fork 0
mirror of https://github.com/Hestia-Homes/Model.git synced 2026-07-27 23:35:01 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Re-baseline the pashub MAE ceiling to 1.415 for the water-only DHW fix 🟪

Browse source 
Second sanctioned exception to "ratchets never loosen" — needs @KhalimCK's
call, since the prior note expected to re-tighten here. Exactly one cohort
property is WHC 950 (16 Bingley Close, +5.946 -> +6.818 = the whole +0.0043);
within-0.5 unchanged at 23.9%. Its pre_sap is verified correct but its ~+6
residual is a known extraction bug, and this hybrid gauge cannot adjudicate a
calculator fix (#1592 Problem 2). The gov-API corpus, which gauges the
calculator alone, improves.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Daniel Roth 2026-07-15 10:57:16 +00:00
parent a400005353
commit 53ba1f1694
1 changed files with 30 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

31
backend/documents_parser/tests/test_pashub_sap_accuracy.py
View file

@ -142,7 +142,36 @@ _MIN_WITHIN_HALF: float = 0.23
# RdSAP corpus). 16 Bingley Close reverts +3.2 -> +4.3, so observed MAE 1.405 ->
# 1.4103; re-tighten when the water-only branch is fixed correctly under #1592
# (the cylinder-insulation MAPPER fix is retained, so the revert is partial).
_MAX_SAP_MAE: float = 1.411
# DELIBERATELY LOOSENED 1.411 -> 1.415 by the #1592 sub-task B water-only DHW
# fix — the SECOND sanctioned exception to "ratchets never loosen", and the
# note above expected to RE-TIGHTEN here, so this needs @KhalimCK's explicit
# call before merge.
#
# The fix is spec-verified against the SAP 10.2 PDF (not ticket prose): §4.3
# p.24 "The efficiency for water heating is incorporated in the price of heat
# from a heat network in Table 12"; C4 p.53 scopes the plant efficiency to
# "the calculation of CO2 emissions and Primary Energy"; C6 p.54 + worksheet
# (310) = (64) × (305a) × (306) carry no efficiency term. The dwelling-side
# `plant_eff / DLF` divisor was applying it a SECOND time.
#
# Exactly ONE cohort property is WHC 950 — 507644414148 (16 Bingley Close),
# +5.946 -> +6.818 = 0.872/205 = +0.0043 MAE, the whole movement. within-0.5
# UNCHANGED at 23.9% (49/205); mean signed +1.050 -> +1.054. Observed 1.41458.
#
# Why this is not a regression signal: Bingley's `pre_sap` E52 is Khalim-
# verified correct (2026-07-14), but its ~+6 residual is a KNOWN extraction bug
# still being hunted (see HANDOVER_PASHUB_EXTRACTOR_ACCURACY.md: "community-
# heating residual — ... community standing charge/tariff detail, or DLF
# 'Unknown' distribution type"). This hybrid gauge conflates that extraction
# error with the SAP-10.2-vs-lodged-2012-RdSAP engine offset, so per #1592
# Problem 2 it CANNOT adjudicate a calculator fix — the same reasoning the
# orientation exception above rests on. The gov-API RdSAP corpus, which gauges
# the calculator alone, IMPROVES on this change (within-0.5 78.7% -> 78.8%,
# MAE 0.6273 -> 0.6272; exactly 3 certs move, all WHC 950).
#
# Re-tighten under #1592 sub-task C, once an accredited Elmhurst 10.2 worksheet
# pins WHC-950 — Bingley is the natural capture candidate.
_MAX_SAP_MAE: float = 1.415
_KNOWN_GAP_REASON = (
"pashub `from_site_notes` mapper does not int-code a main-heating "