Runs a Summary PDF through the chain-test path (pdftotext ->
ElmhurstSiteNotesExtractor -> from_elmhurst_site_notes) into Sap10Calculator
and dumps SAP + per-end-use kWh + the intermediate worksheet trail, for
diffing our calc against the accompanying Elmhurst U985 worksheet PDF.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add Phase 6 (self-improve) to audit-ara-portfolio: when a run confirms a
novel systematic problem, codify it as a check — gated on systematic (>=5
props, root-caused), not-already-covered, and /grill-me-pressure-tested.
Each check records provenance (motivating cause + example properties) so the
registry stays sharp and compounds every run.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Script takes an optional --scenario to restrict to one scenario's plans. New
skill drives the full loop: run the deterministic scan, review groups,
deep-dive samples via run_modelling_e2e, characterise sub-classes, and
cross-reference open PRs/ADRs — then proposes new checks to codify.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A check-registry that reads property/baseline/plan/scenario and flags odd
results (plan-below-baseline, already-meets-goal-with-works, band/score
mismatch, zero-works-post-differs, effective-lodged divergence, negative
bill savings). Writes modelling_audit.md/.csv. Adding a check = one
decorated function.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When refetch_epc=False and no stored lodged EPC exists, the handler no longer
falls back to a live EPC API call — it treats the property as EPC-less and
hands it to the prediction path. This keeps REFETCH_EPC (lodged path) and
REPREDICT_EPC (prediction path) cleanly independent.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Captures the calc-facing ventilation-type gap (older RdSAP mappers drop
mechanical_ventilation_kind), the FE-side Main Fuel display, the sweep survivor
clusters, and the predicted-property display path — all separate from ADR-0037.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Group pending property IDs by postcode and pack them into ~BATCH_SIZE
messages, never splitting a postcode, so each SQS message drives one
batched modelling_e2e Lambda invocation. Adds a completed-since skip
filter and a properties cap.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
In-container Playwright runs were flaky: the chromium renderer crashed
mid-build ("Target crashed") on the 64M /dev/shm, and login intermittently
hung. Added `--disable-dev-shm-usage` + `--no-sandbox` launch args, a
4-attempt login retry loop (domcontentloaded + explicit selector wait),
and an `ELM_GUID` env override so a per-UPRN assessment can be targeted
without editing the module. Tooling only — no calculator impact.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
When a property failed, the handler recorded only its bare property_id and
raised RuntimeError("failed property_ids: [...]"). That string is what
SubTask.fail persists into the subtask outputs.error column, so a failed run
told you which property failed but never why — forcing a CloudWatch lookup.
The per-property catch now captures property_id, uprn, error_type, and the
error message, and the raised RuntimeError embeds those as JSON so the subtask
outputs column is parseable directly. query_failed_modelling_e2e.py reads that
outputs.error into a new Error column in its report.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Closes out the cohort-broadening work with its decision record and consolidates
the retry plumbing.
ADR-0034 documents broadening the EPC-Prediction cohort to the real unit
postcodes nearest the target (via postcodes.io) when its own postcode holds no
same-type comparable — extending ADR-0031 decision 5. Records why postcodes.io
was chosen over council[] (whole-LA, no property_type in rows), a bulk Code-Point
Open / ONSPD dataset, and the OS Places radius API, and the lazy / nearest-first
early-stop / soft-fail policy. Broadening-specific docstrings now cite 0034.
Retry consolidation: extract the EPC client's call_with_retry into a shared
infrastructure/http_retry.py keyed off a generic TransientHttpError marker, so
the mechanism (exponential backoff, Retry-After) is shared while each client
keeps its own transient policy. EpcRateLimitError now subclasses TransientHttpError
(still an EpcApiError); PostcodesIoClient routes through the same helper, raising
TransientHttpError on 429/5xx and soft-failing to the seed once exhausted (the EPC
client propagates instead). Direct tests for the shared helper; EPC + postcodes.io
suites repointed at the shared sleep.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two reconciliations to make the modelling_e2e Lambda handler production-ready.
1. Price through the off-catalogue overlay, drop the workarounds
The handler priced through a plain ProductPostgresRepository and excluded
secondary_heating_removal / system_tune_up / system_tune_up_zoned to dodge
ProductNotFound (and a poisoning pgEnum DataError). Those measures are now
priced by catalogue_with_off_catalogue_overrides (already used by the e2e
runner and PostgresUnitOfWork), so the exclusions are removed and ALL measure
types are considered. This also fixes gas-boiler / single-glazed properties,
which Dan's handler never excluded and so still crashed (the standard
system_tune_up option is built unconditionally — the considered-measures
exclusion never actually gated it).
2. Broaden the EPC-Prediction cohort to nearby real postcodes (ADR-0031)
A property with no lodged EPC and no same-type comparable in its own postcode
(e.g. the only flat among houses) used to gate out and fail the subtask. The
gov EPC API cannot search by radius/outcode, so we resolve the real unit
postcodes physically nearest the target via postcodes.io (keyless; already a
trusted in-repo dependency) and walk them nearest-first until enough same-type
comparables surface. New PostcodesIoClient (transient-failure retry with
exponential backoff, soft-failing to the seed so broadening never breaks
prediction) and EpcComparablePropertiesRepository.candidates_near. Wired into
the handler and e2e runner; broadening is lazy (only on gate-out) and memoised
per (postcode, property_type).
Validated live: property 728476 (gas boiler) prices system_tune_up at GBP295;
property 718580 (lone flat in BR6 6BS) now predicts via nearby BR6 postcodes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Corpus validation of the modelling_e2e photovoltaic_supply-as-list fix. Cert
6102-6227-8000-0083-2292 (RdSAP-20.0.0 semi, gas combi + 2× 1.14 kW PV arrays)
crashed from_rdsap_schema_20_0_0 on the measured-array list; the fix routes it
through the dict-tolerant _map_schema_21_pv. PV correctly credited: engine 61
(no PV) → 66 (+5). Built in Elmhurst (evidence: epc.json + summary + worksheet,
fabric+heating; the PV "New Technologies" Panel-details grid deferred): worksheet
55 = engine-on-Elmhurst-inputs 55 exactly → calculator faithful. The +6 engine-vs-
Elmhurst base-dwelling residual is the documented RdSAP-default gap (band-C cavity-
uninsulated suspended-floor semi). Pinned engine 66.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Corpus validation of the modelling_e2e built_form fix. Cert 8742-6624-9300-2780-4926
(SAP-Schema-16.0, ground-floor electric-storage-heater flat) omits built_form; the
mapper now derives it from dwelling_type. built_form is ML-only so the fix is
SAP-neutral: engine 66 = lodged 66 exactly. Built in Elmhurst (evidence: epc.json +
summary + worksheet): worksheet 54, engine-on-Elmhurst-inputs 53 ≈ 54 → calculator
faithful. The +12 engine-vs-Elmhurst is a build/input gap (cert size-1 small cylinder
unrepresentable in Elmhurst's Normal/110L-minimum entry → higher HW + reduced-field
16.0 defaults). Pinned engine 66.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- run_modelling_e2e --from-db re-models from already-persisted inputs (reads
each Property's Effective EPC + planning protections + solar from the DB) and
skips every live fetcher — zero gov-API calls. With --persist it re-writes the
Plan and, for lodged-EPC Properties, the Baseline. Self-contained loop; the
live-fetch path is untouched. Makes local re-runs instant and avoids tripping
the gov API's per-IP rate limit (6000 req / 5 min) during iteration.
- EpcClientService.REQUEST_TIMEOUT 10s -> 30s: a cold per-UPRN search can exceed
10s and the old timeout turned it into a timeout-then-retry; 30s rides it out.
Note: an open perf question remains — modelling is fast in isolation (<0.5s/
property) but a long-lived --persist run shows ~1 min/property; suspected in the
persist path (plan.save / baseline) or connection handling, NOT the API. Left
mid-diagnosis for handover.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
profile_corpus_error.py: print median signed error alongside the mean in each
feature bucket. The mean is dragged by fat-tail register anomalies (e.g. electric
room heaters mean -1.09 but median -0.01) — median is the outlier-resistant lens
for finding TRUE systematic slices, so hunt by |median|, not |signed|.
heat_transmission.py: document why permanent-shutter R is deliberately NOT applied
(Elmhurst uses R=0.04 curtains on every window incl. insulated shutters, proven on
case 46; API-path trial worsens MAE). Comment-only.
pyright not installed locally — strict type gate not run.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Make run_modelling_e2e the single script that does everything for a portfolio,
so the 291-property run needs one invocation with per-property recovery (no
all-or-nothing chunking):
- On --persist, a lodged-EPC Property now also gets its Baseline Performance
row written via PropertyBaselineOrchestrator (per Property, so one bad cert
does not abort the batch). Predicted (EPC-less) Properties have no lodged
figures, so they get a Plan but no baseline row.
- The run CSV gains api_sap (register) vs baseline_sap (calculator) + sap_delta,
so calculator-vs-API divergence is reviewable per property.
Fill the off-catalogue overlay for the measures the live material catalogue
cannot price, so they stop crashing the run:
- double_glazing (£550/window) and secondary_glazing (£400/window): priced
per window (the generator multiplies by single-glazed window count, matching
the legacy window_glazing). Grounded in 2025/26 UK installed costs; per-window
is the right unit for windows (fixed per-unit install dominates) — per-m2 fits
walls/floors, not glazing.
- gas_boiler_upgrade / system_tune_up / system_tune_up_zoned: these are priced
off the heating rate sheet (Products()), with get() reading the catalogue only
for an id — so the overlay entry exists to satisfy that lookup (material_id
stays None, as with ASHP); the rate sheet remains authoritative.
Validated on a 12-property sample (incl. a secondary-glazing case and a
SAP-Schema-16.2 cert): 12/12 baseline rows + plans, 0 errors.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
