Evaluation harness
Accuracy
22 synthetic RICS-style reports with known-correct ratings, triage, and citations. We re-run the extractor and triage engine against that gold set whenever the page loads. This is not a claim about live uploads.
Reports in corpus
22
All synthetic. No personal data.
Overall
100%
Mean of the six checks below.
Ratings
100%
Condition rating matches gold.
Triage
100%
Urgent / negotiate / budget / monitor / ignore.
Citations
100%
Gold phrase present in the extract.
Grounding
100%
Every citation appears in the source.
Unreadable pages
100%
Poor scans stay unread — we do not invent a covering.
Property fields
100%
Address, postcode, type, survey level.
Hallucination rate 0% — we fail a report if the decode invents a swimming pool, a valuation, or any other forbidden phrase on that gold card. Corpus mix: Level 1 × 5 · Level 2 × 12 · Level 3 × 5.
Per report
Overall score for each synthetic report. Featured samples are the three you can open from the homepage.
Gold cards
| Report | Level | Ratings | Triage | Cites | Overall | |
|---|---|---|---|---|---|---|
Featured | 2 | 100% | 100% | 100% | 100% | Open |
Featured | 3 | 100% | 100% | 100% | 100% | Open |
Featured | 1 | 100% | 100% | 100% | 100% | Open |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 3 | 100% | 100% | 100% | 100% | Open | |
| 1 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 1 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 3 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 1 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 3 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open | |
| 1 | 100% | 100% | 100% | 100% | Open | |
| 3 | 100% | 100% | 100% | 100% | Open | |
| 2 | 100% | 100% | 100% | 100% | Open |
How we score
- Gold labels are written with the reports. The extractor is not shown the answers.
- Ratings and triage are scored only on the sections we labelled — extra extracted sections do not count against us.
- A citation passes if the gold phrase appears in the surveyor wording or in a grounded excerpt.
- Unreadable pages must stay unreadable. We fail if the engine invents a covering for a poor scan.
- Live uploads still go through the same engine, then an optional model, then the same grounding step. The model is not in this harness — it would spend API quota on every page load.