Business Case

10 September 2026. Why this exists, what we set out to build, what we have built, what it costs, and where it goes next.

The engineering companion is technical-overview.md.

1. The problem we are solving

A title search answers one question: what is recorded against this property?

Today it is done by hand, and the work is the same every time. An analyst:

1. opens a county recorder's name index,
2. searches the current owner,
3. finds the deed that gave that owner title, takes its grantor, searches them, and repeats backwards through the chain of ownership,
4. reads every instrument the index returns — deeds, mortgages, liens, judgments, releases,
5. decides which of them concern this parcel rather than a different parcel belonging to someone with the same name,
6. and writes the survivors into a search note, which is what the customer buys.

Steps 3 and 5 are the expensive ones. Step 3 is mechanical but tedious. Step 5 is judgement, and it is where a search goes wrong: a name index returns everything filed under a name, and most of it belongs to somebody else.

The economics follow from that:

• Cost per order is analyst time. Nothing else in the process costs anything material.
• Throughput is headcount. Twice the orders needs twice the analysts.
• Quality varies by analyst, and a wrong note is invisible — it looks exactly like a right one.
• Turnaround is the product. Customers are waiting on a closing.

2. What we set out to build

An order goes in, and a search note comes out that an analyst would recognise as their own work — without anyone opening a terminal, and without a person having to redo the judgement the system already made.

We set the bar deliberately high and deliberately narrow:

Success is not "does it run." Success is "does its note match the one an analyst wrote."

That is why fourteen delivered analyst notes sit in the repository as ground truth, and why every accuracy claim is scored against them rather than against an internal metric. A system that runs cleanly and produces a plausible wrong note is worse than one that produces nothing, because nothing is obviously nothing.

Three principles came out of that and have shaped everything:

Fetching never judges. The component that retrieves records is forbidden from deciding which ones matter. It returns everything it found, each item carrying a record of how it was found; a separate stage decides relevance. This exists because the predecessor system stamped records as useful inside the fetching code — invisible downstream, with no record of why, and untestable.

Producing less is fine. Producing something untrustworthy is not. A missing tax bill makes a thinner note and everything else on it is still true. A missing legal description makes the instruments that follow not fewer but unfiltered, and no reader can tell. The first degrades; the second must stop and ask a person.

Every failure must be legible. A missing model key, a missing parser key, a missing cache — each of these used to look like a normal, quiet result. Each now says what it is.

3. What we have built

3.1 The working loop

   email intake  ──┐
                   ├──▶  order record  ──▶  analyst assigned
   manual entry  ──┘          │
                              │  Run
                              ▼
        ┌──────────────────────────────────────────────┐
        │  plan      what this county can be searched  │
        │  collect   recorder index, card, tax bill    │
        │  extract   read the parcel's legal identity  │
        │  classify  decide which instruments are ours │
        │  review    ask a person, only when needed    │
        │  generate  search note · report · package    │
        └──────────────────────────────────────────────┘
                              │
                              ▼
              note attached to the order, delivered

An order goes in over HTTP and a note comes back out. That loop runs today.

3.2 The milestones that mattered

We got live county access working. Two very different kinds:

• Virginia — the state's Secure Remote Access land records, reached as a JSON API from inside a browser session, behind a per-county paid subscription.
• Brevard County, Florida — a public search with a CSV export for the recorder, and an assessor site behind Cloudflare. We measured both halves of that block rather than guessing: from India the site refuses the country outright; through a US exit node it serves a solvable challenge; and a headless browser never clears it while a headed one clears it instantly. That county now works, and we know exactly why it works.

We built the search the way an analyst actually does it. The grantor-grantee walk — finding the vesting deed, taking its grantor, searching them, repeating — is the primary method of a real title search, and it was the largest missing piece. It runs bounded: limited depth, a back-to date, institutional grantors never followed, and it abandons a hop rather than flooding the result set.

We made the note match. On the largest order in our corpus — 69 instruments, the hardest case we have — the system finds 54 of them, with zero false positives. Precision holding at zero as volume grows is the single strongest signal we have that the judgement layer is sound.

We made the system refuse to guess. When a run has no way to tell this parcel's instruments from anyone else's, it stops and says so instead of returning a name-index dump wearing the name of a search note. It now asks that question before opening a browser, so a hopeless order does not spend a shared county login discovering it.

We built a measurement discipline. Fourteen analyst notes as ground truth, scoring the rendered document rather than internal state, and a standing rule that no accuracy claim is made from memory. Several promising ideas have been built, measured, found not to pay, and reverted with the measurement written down — which is cheaper than building them twice.

We made it durable enough to trust. 1,111 automated tests. Work is persisted as each stage completes, so a crash loses the stage in flight and nothing before it.

3.3 Where that leaves us, honestly

The system assists today; it does not yet replace.

Across our thirteen scored orders, seven stop and ask a person. The judgement those seven contain is good — 72 correct instruments are sitting in them — but a human has to release it. The orders that finish unattended are mostly the ones where the correct answer was "nothing new was recorded."

That is a real and expected position for where we are. It also tells us exactly what the next phase of work is, and §6 is built around it.

4. What it costs

The costs here were measured against this system, and the external prices are public and dated. Where a number is an assumption it is marked.

4.1 Compute is negligible

Running an order through the pipeline costs:

At 500 orders a month, that is under $1 of inference and, for parsing, within the free monthly allowance. Compute is not a cost centre and will not become one.

4.2 Coverage costs are fixed, per county

This is the number that governs expansion, and it is per county per month, not per order. Virginia sells land-records access per circuit court, and published rates cluster around $50/month or $300–600/year per county. Terms are set locally and must be confirmed with each Clerk.

Florida differs — Brevard's recorder needs no subscription, but its assessor needs the US network path and headed browser described above, which is an infrastructure cost rather than a licence fee.

On top of the subscription, each county needs a connector: a step list watched working against the live site before it is trusted. That is engineering days, not a fee.

4.3 People are the entire cost

Using $25/hour and 20 minutes on an order that stops — both assumptions you should replace with your own figures — the human step costs roughly 220× what the compute costs. The conclusion is not sensitive to those inputs: at any plausible rate, review dominates.

This is the most important commercial fact about the product. Cheaper models, faster inference and prompt tuning are close to irrelevant. The economics are governed by how often a person is needed, and by nothing else.

5. The goal

Stated concretely enough to be checked:

An order for a covered county goes in and a delivered search note comes out, unattended, matching what an analyst would have written — and a person is involved only on the genuinely ambiguous cases, not on the routine ones.

Three things follow from that phrasing:

• "Unattended" is the target, not "accurate." We are already accurate on the work we complete. We are not yet completing enough of it without help.
• "Genuinely ambiguous" is an accepted, permanent category. The goal is not zero human involvement. Title work has real ambiguity and a person is the right answer to it. The goal is that the person sees only those.
• "Covered county" is a deliberate constraint. Breadth without depth is worse than no coverage, because it produces thin notes that look thorough.

6. Roadmap

Phased by what each phase proves, not by what it builds.

Phase 1 — Prove the loop on one county

Goal: one county where an ordinary order goes in and a note comes out with nobody watching.

• Run real orders end to end through the production entry point, not the test harness, and fix what that surfaces.
• Time an analyst through a real review queue. Every economic number in §4.3 rests on an estimate nobody has measured; this is a day of work and it prices the product.
• Decide and document what "delivered" means operationally — who checks, who signs off, what happens on a failure.

Done when: a week of real orders in one county complete without intervention at a rate we are willing to quote.

Phase 2 — Drive down how often a person is needed

The economic phase. This is where the money is, and it is the whole of §4.3.

• Attack the review rate directly. Today more than half of orders stop; the target is a minority, and each percentage point is a direct margin gain.
• Give the deterministic layer the ability to rule out, not only to affirm. Today anything the rules cannot confirm becomes a person's problem by default.
• Improve the signals that decide relevance, measured against the fourteen reference notes every time, with results recorded whether or not they pay.

Done when: review rate is low enough that one analyst supports several times today's order volume.

Phase 3 — Make adding a county routine

Goal: a new county is a predictable cost, not a project.

• Turn county onboarding into a repeatable procedure with a fixed shape: subscribe, build the connector, watch it work, promote it, measure it.
• Retire or verify the connector configurations that were written from reading a site rather than watching it answer. They are not coverage and must never be counted as coverage.
• Sequence counties by order volume, priced against §4.2.

Done when: a new county goes from decision to delivering notes on a predictable timeline and a known cost.

Phase 4 — The second deliverable

Goal: sell more than the search note.

• The title report — defects, gaps, a marketability verdict — is produced today but has never been checked against a real one, because no reference report exists. One analyst-written report unblocks this, and it is the cheapest item on this roadmap.
• Once validated, it is a second product from the same run at no additional collection cost.

Done when: the report is scored against real reports the way the note is.

Phase 5 — Scale the operation

Goal: the system survives growth and staff turnover.

• Remove the operational single points: work that is lost on restart, the single-instance constraint, the hosts that must exist in a particular place.
• Harden access control and finish test coverage on the order-management half, which is the part that touches customers and money.
• Make the infrastructure requirements — including the US network path and headed browser one county needs — explicit and reproducible, so a rebuilt host does not quietly break a county.

Done when: growth is a capacity question, not an engineering one.

7. What could stop us

The review rate does not fall. If judgement cannot be pushed further without losing precision, the product is an analyst accelerator rather than a replacement. That is still a real business, but a different one, and Phase 2 is where we will find out.

County access changes underneath us. Subscriptions, portal redesigns and bot-protection are outside our control, and one county has already required a specific network path and a real display to reach. Expect this to recur, and budget for it as ongoing maintenance rather than one-time build.

Coverage outruns evidence. The temptation is to claim counties we have configured but never watched work. A thin note that looks thorough is the single most damaging thing this system can produce, and the guard against it is a discipline rather than a feature.

Ground truth stays scarce. Every improvement is measured against fourteen analyst notes. More reference notes — and the first reference report — directly increase how confidently we can change anything.

8. Summary

Intent. Automate the search a title analyst performs by hand, to the standard of the note they would have written.

Accomplished. A working end-to-end pipeline; live access to two very different county systems including one behind a geographic block we diagnosed rather than guessed; the grantor-grantee chain walk that is the real method of a title search; 54 of 69 correct with zero false positives on our hardest order; a measurement discipline with real analyst notes as ground truth; and a system that refuses to produce a note it cannot support.

Goal. Unattended delivery on covered counties, with people reserved for genuine ambiguity.

The lever. Not compute — that is two cents an order and will stay there. The lever is how often a person is needed, which today is more than half the time and is the entire cost of the product.