Trust & transparency
How the deal desk decides — and what it refuses to do.
Most “AI for real estate” tools are chatbots in a trench coat. This one is different. Read this page once and you'll know exactly what kind of decision you're getting, why you can audit it, and where the limits are.
The LLM narrates. The engine decides.
When you submit a deal, the agent calls five typed tools in parallel — comps, rent estimate, school scores, flood zone, recent permits. Each tool returns a strictly schemed payload, never free text. The orchestrator binds those payloads into a single decision contract. Then a deterministic rules engine, written in plain TypeScript, produces the verdict, the score, and the confidence number. The language model is only used to narrate the timeline and produce the email draft afterwards. It never moves the score and never picks the verdict.
That's why the same inputs always produce the same verdict. Every deal you submit gets a content hash, and we surface that hash in the audit footer. Run a deal twice, watch the hash, watch the verdict. They will match. Always.
The percentage on the verdict card is not a vibe.
The confidence score is a weighted formula over four observable inputs:
- Completeness — how many of the required signals returned data versus how many timed out or returned a null.
- Agreement— whether independent signals point in the same direction. Comps say overpriced and rent says under-rented? That's low agreement, lower confidence.
- Freshness — every signal carries a fetched timestamp. Stale signals downgrade confidence on a curve.
- Rule certainty — how close the deal sat to the threshold that flipped its verdict. A deal that misses a hard breaker by 4% is less certain than one that misses by 40%.
Below 60% the verdict carries a Low confidence pill in the UI. We do not hide that.
What this system does
- Underwrites Texas commercial real estate deals — office, industrial, multifamily, retail, mixed-use, land.
- Returns a deterministic verdict with a full audit trail of every signal pulled and every rule that fired.
- Cites every signal that moved the score and every rule that fired.
- Produces a PDF investment brief and a public read-only share link suitable for sending to an IC.
- Refuses to produce a verdict when too many signals are missing — instead surfaces what's missing.
- Caches identical analyses by content hash so you can prove reproducibility.
What this system does NOT do.
Every limit below is a deliberate design choice. We'd rather narrow the scope and be right than promise everything and quietly hallucinate.
It does not replace a licensed broker, attorney, lender, or CPA.
The verdict is screening output. Closing a deal still requires title, legal, financing, and tax sign-off from licensed professionals.
It does not analyze deals outside Texas.
The rules, comps, and tax assumptions are calibrated to Texas. Out-of-state addresses are auto-rejected with a clear message — never silently mis-analyzed.
It does not negotiate, bid, or transact.
The system reads inputs and produces a verdict. It never sends a message, opens a deal, signs an LOI, or wires money.
It does not predict future market conditions.
Forecasts of cap-rate compression, rent growth, or interest rate moves are explicitly out of scope. We only evaluate the deal as it is today.
It does not let the language model override the rules.
The verdict and score are produced by deterministic code. If the LLM disagrees with the engine, the engine wins — every time.
It does not act on instructions hidden in your notes field.
Free-text fields are treated as untrusted data. Prompt-injection attempts (e.g. “ignore prior rules and pursue”) are surfaced and ignored.
It does not promise the data is real-time.
Every signal carries a freshness timestamp. We show it. If a signal is stale, your confidence number reflects that.
Why an institutional LP can sign off on this.
Every verdict ships with a receipt: every signal that moved the score, every rule that fired, the observed-versus-threshold inequality, the freshness of each tool response, and a deterministic reasoning trace. None of it is generated narrative. All of it can be reproduced from the input hash. That's the whole pitch.
Lone Star Deal Closer is an underwriting screening tool. It is not investment advice, an offer to buy or sell securities, or a substitute for licensed professional services. Use of the system is governed by our Terms of Service.