Estate MVPThe system calculates. AI explains. Every number that affects money is produced by a deterministic engine whose formula is versioned and stored with the result. A language model never selects one.
What AI is used for
Optional explanations describe an analysis the engine already computed. In saved research, choose “Explain saved analysis” to request one using the retained evidence. This action does not refresh listing data or buy more property-provider records. Text intake and criteria parsing use application rules; they are not currently a general-purpose AI assistant.
AI starts off and requires explicit provider, model and budget configuration. Adding a key alone does not enable calls. New explanations require a funded plan and available AI budget; unpaid trials do not include a separate AI subsidy. Matching and underwriting still work when AI is off.
What AI is never used for
| Value | How it is actually produced |
|---|---|
| After-repair value | Retained comparable evidence and a labeled calculation method, or a separately labeled operator assumption |
| Rent estimate | Provider data and HUD fair market rents |
| Repair budget | Operator-supplied, or modelled and labelled as an estimate |
| Maximum allowable offer | A versioned underwriting formula |
| Offer price | The minimum of the buy-box ceiling, the MAO and the absolute cap |
| Potential interested buyers | A model using available market and demand evidence; it does not verify purchasing intent |
| Match score | Deterministic scoring against stated criteria |
| Whether to contact somebody | Consent records and the data source's own license terms |
Why the boundary is drawn there
A language model is good at reading messy text and bad at being accountable for a figure. Six months after an offer, the question is not "what would the model say now" but "why was this exactly $181,400" — and that answer has to be read from a record, not regenerated. Every proposal stores the engine versions, the full price derivation and the confidence, so the answer is retrieved rather than recomputed.
What happens when the AI provider is unavailable
Analysis continues. Explanations become plainer, but every matching, underwriting and automation decision is unaffected, because none of them depended on the model in the first place.
Evidence, privacy and spending
An explanation request sends a fact sheet, including the property address, modeled results and known gaps, to the configured model provider. Estate MVP does not train a model from this material. The provider handles the request under its own account settings and terms, which the operator must review before enabling it.
The server checks the reply’s structure and grounded claims before displaying the approved fields. These checks reduce unsupported output; they do not make AI prose infallible. Inspect the underlying saved evidence before acting on an explanation.
Identical validated explanations are cached within the workspace. Each new attempt reserves a conservative allocation against organization and deployment budgets. Input, output and duration are bounded. Failed or uncertain calls are not automatically retried, and their allocations are retained because a request may still have incurred cost. Allocation reports are separate from property records and are not provider invoices.