Why we never invent numbers
You can tell Nestraqanything about the home you want — "quiet street", "room to extend", "good primary nearby", "south-facing garden". A language model reads your words. But it is never — architecturally never — allowed to make up a number about a home. Here's the boundary, published verbatim from our design doc.
The problem with AI-graded property tools
There's a growing class of tools that bolt a model onto listings and let it invent a 0–100 grade. The number looks precise. But ask the obvious questions — which data produced it? would it be the same tomorrow? can anyone reproduce it?— and there's nothing underneath. A model asked to "score this home" will always oblige, whether or not it has any evidence. That's not analysis; that's confidence-as-a-service.
We built Nestraq the other way round: the model is only ever allowed to understandyou. Every number comes from a deterministic engine over real, named data — or it doesn't appear at all.
How the job is split
- Intake — open vocabulary.Say anything. A small, temperature-zero, schema-constrained model maps each requirement to a concept in our audited catalog — and honestly flags what can't be assessed yet.
- Chips — you're in the loop. Our interpretation comes back as editable chips. You see exactly what we understood, and correct it before anything counts.
- DSL — deterministic and whitelisted. The chips compile to a versioned spec over a whitelist of real fields. Thresholds, prices and distances are set by a deterministic parser — never by the model.
- Scoring — a pure function.The same spec against the same home always gives the same result. Your headline is "matches N of M priorities" with a coverage band — not a naked 0–100 conjured from nowhere.
The trust boundary — verbatim
This table is copied word-for-word from the internal design document that governs the feature (a build-time test fails if this page ever drifts from it):
| The LLM IS trusted to… | The LLM is NEVER trusted to… |
|---|---|
| Understand any phrasing | Set a numeric threshold/price/distance (deterministic parser does this) |
| Map a phrase → a catalog concept / proxy / tier | Set a weight |
Echo back the user's words (user_phrase) | Compute or influence a score |
| Flag "can't assess" honestly | Emit a field/operator outside the whitelist (rejected in validation) |
What honesty looks like in the product
When you build a brief, every chip wears its assessability on its sleeve: solid chips are scored today against real data; outlined chips say "can't assess yet — coming with photo analysis" and light up automatically when that layer lands; dotted chips are yours to judge — some things (the feel of a street, the neighbours) no dataset can score, and we won't pretend otherwise. Nothing you say is silently dropped, and nothing we can't see gets a number.
The same rule governs our valuations: every deal-score traces to named sources — HM Land Registry sold prices, EPC, and the rest — with its confidence shown, and our methodology page states the failure modes plainly. It's also how we can tell you to pay less and substantiate it.