Mon, Oct 19

Last: today's to-dos

Confirmed: seed the inventory now, and fill the gaps later.

Last one, Joe's to-dos. Say "yes" if these five are yours:

1. Make the notes page work well on a phone
2. Send Keegan and Reid the mapping tool
3. Deploy the one app to the cloud
4. Review Reid's checker branch and say merge or what to change
5. Answer Reid's 3 questions about the test documents

Two to-dos have no owner yet: deciding where the files go that get a deal to 100%, and picking how to search the county records. Name an owner, or say "next call".

Next to confirm

Confirmed, with Joe's change: deal intake end to end is one flow round 1 tests. Other flows may be added.

Last decision, say "yes" or "that's wrong": Seed the inventory now with the data we have and build the processes on top of it. Fill the data gaps later, as each process gets solved.

Next to confirm

Confirmed: AI has to beat plain code before we use it on a document type.

Next, say "yes" or "that's wrong": Round 1 of the experiment tests deal intake end to end. A closed deal folder goes in, and a complete inventory record comes out, with proof for every fact and questions instead of guesses.

Next to confirm

Confirmed: deal intake is the first job to solve.

Next, say "yes" or "that's wrong": For each document type, AI has to beat plain code's match rate before we use it there (check stubs: 81%).

Closing out: confirm one at a time

Closing out. One thing to confirm at a time. Say "yes" or "that's wrong".

First: Deal intake (a closed deal becomes a complete inventory record) is the first job to solve, before the other jobs.

Coming next: AI has to beat plain code; round 1 tests deal intake; seed the inventory now; then owners for today's to-dos.

Reid's test work: what we can build on

Reid's work from Friday and Sunday gives the experiment three pieces we were missing. We should build on them, not start over. It's in the team's experiments repo, on a branch that isn't merged yet.

What Reid built

1. The checker. Code that grades any AI version against the right answers, using memo B's scorecard. It already measures being wrong without asking, the right deal, the right type, finishing with no person, asking at the right times, and giving the same answer every time. Not yet: proof for each fact, the record of what it did, cost, and hidden instructions planted in documents.
2. The plain-code version (the "No AI" control). Rules-only deal matching. On 198 checked documents, it filed 128 on its own, all of them right, and handed 70 to a person. Its first guess was right 82% of the time.
3. Right answers for 198 documents, deal and type only. That's a much bigger sample for right deal and same answer every time than the 27 documents.
4. A ruling: where HubSpot and the Excel inventory disagree, the Excel wins.

How it fits what we agreed today

• Scoring: Joe picked safety first (option A). Reid's notes still show B as the recommendation, so that line needs updating.
• Deal intake end to end (today's proposal) adds a whole-deal check: closed deal folder in, complete record out. That one uses the closed-deal answers and their existing grader. Reid's checker and his 198 documents cover the document-matching part of intake.
• Keegan's caution and Reid's answer both apply: use these answers to build now, and check them against the recorded papers going forward.

Waiting on Joe

1. Merge Reid's checker and the Excel ruling? Joe reviews the branch.
2. Three questions from Reid: which of seven similar assignments he meant for one test document; whether to keep a division order that sits in three deal folders; and whether the plain-code version is graded on deal only, since it never names a type.

Also proposed by Reid

Between calls, each of us drops a dated update file in the notes' updates folder. At the start of the next call, Claude reads them, files them under their projects, and reads their decisions back for a yes. This needs one added step in the meeting app.

Near-term open items, and what to confirm

Near-term open items, before we wrap up.

To confirm before we hang up

Say "yes" to confirm all of these, or name the one that's wrong:

1. Deal intake is the first job to solve, before the other jobs.
2. For each document type, AI has to beat plain code's match rate to be used (check stubs: 81%).
3. Round 1 of the experiment tests deal intake end to end.
4. Seed the inventory now with what we have, build the processes on top, and fill the gaps later.

Also: who owns #6 and #7?

What we agreed, and the to-do list

What we agreed today, and the to-do list in priority order. Say "that's wrong" about anything and I'll fix it.

What I understand we agreed

1. Deal intake comes first. A closed deal becomes a complete inventory record. Every later step (operator emails, county tax transfer, division orders, monthly check comparison, tax statements) trusts that record.
2. Seed the inventory now with what we have. Build the processes on top of it, and close the data gaps later as each process gets solved. Don't wait for every deal to hit 100%. (Reid, Keegan, Joe)
3. The first experiment tests deal intake end to end, on the AI co-workers with their own computers. The right answers already exist for the closed deals. Scoring is safety first, and every version has to beat plain code.
4. The AI workers should lower what needs a person, for everything. They shouldn't hand Kelsey, Tana or Kathryn a new pile of work.
5. The missing data is mostly scattered, not missing. Integra, the Railroad Commission data and our own files already hold most of the fields. Mine those first, then go get the rest.

To-do list, most important first

Later, once intake works

• The steps after intake that Keegan laid out: operator and county emails, division orders, the monthly check comparison, the tax statement comparison.
• A shareable map of what we own, drilling permits, and live views of drilling sites.

Still open

• Is this our one-sentence goal for all of this: "Confirm with confidence that we get paid correctly and pay taxes correctly on every asset we own"?
• Owners for #5, and for the to-dos above that show more than one name.

What the experiment tests: my understanding

My understanding: the experiment tests one job end to end, deal intake. We hand each AI co-worker a closed deal's folder. It has to turn the folder into a complete, correct inventory record, and ask a person whenever it isn't sure. Everything after intake comes in later rounds, because every later step trusts what intake wrote.

What each AI co-worker gets

• One closed deal's folder: the deed, title papers, the allocation table, stubs and division orders.
• The same tools for every version: read a document, look up deals, open a deal's record, save its answer, ask a person.
• The same written instructions, and no way to change real records or send email during the test.

What it has to hand back

1. The right deal. It matches the folder to the right deal in HubSpot.
2. Every interest we bought: tract, type, royalty, acres, legal description, county, operator and net revenue interest.
3. Who pays us: each payor and the owner number, from the allocation table.
4. The money: the total price, and the basis allocation, meaning how the price splits across each property for accounting. Keegan does this by hand today.
5. Proof for every fact: the page and the exact words it came from.
6. Questions instead of guesses when two answers could fit, a number is hard to read, or the papers disagree with our records.

How we know which version wins

• The right answers already exist. The closed Rio Grande Fund deals have hand-checked answers and a grader that compares a result field by field. Reid doesn't have to write new ones.
• Safety first. The number we care about most is how often it hands in a wrong fact without asking. Asking at the right times comes next.
• It has to beat plain code. Plain code with no AI runs the same deals, like Reid's Friday check (81% on check stubs). A version that can't beat it loses.
• Each deal runs 3 times per version, to see if the answers hold.

Later rounds: the steps after intake

These are the loops Keegan laid out. We test them once intake is solid:

1. Email the operators, and the county appraisal district for the tax transfer.
2. Match each scanned division order to its deal.
3. Send the signed division order back in the same email thread.
4. Each month, compare every check's interest with the one on file.
5. Each year, compare the tax notice and the tax bill with the interest on file. Keegan's idea: the AI could also check county websites directly instead of waiting for the yearly paperwork.

To confirm

1. Is deal intake end to end the right job for round 1?
2. Do we test on the closed deals we already have answers for, or on new deals as they close? Reid said we don't need to finish the old ones before taking in new ones.
3. Should the basis allocation be part of what we grade?

What we've covered so far

Notes are up to date. Here is what's been said so far. I joined partway through, so tell me if anything earlier is missing.

Here today

Joe, Reid, Keegan.

Deal intake comes first

• Keegan: the inventory is the brain of the whole operation. Nothing else matters until every closed deal lands in the inventory as a complete set of facts.
• Reid: agreed. The gate into our system is after a deal closes, so intake is the first thing to solve, and we're close. Paperwork also comes in before closing (title packets), so scanning isn't only an after-transfer step.
• Joe: the core tools are in place. We need to pull more facts into each record, which is business rules, and hand the work to the AI.

Reid's Friday check: plain code vs. AI

Reid ran thousands of check stubs and other documents (division orders, title documents, offers) through plain code with no AI. Code matched check stubs to the right deal 81% of the time. So for check stubs, AI has to beat 81% to be worth using. Each document type gets its own bar.

Decisions to confirm at the end

1. Deal intake is the first job to solve, before filing division orders, operator replies and checking payments.
2. For each document type, AI has to beat plain code's match rate to be used (check stubs: 81%).

Both are marked "not sure yet" until we read them back at the close.

Getting started

Cletus, an AI notetaker, is listening and taking notes.

Three quick questions to get started. Answer out loud after the hellos:

1. Who is here?
2. How long do we have?
3. What do we want to walk out with?

One topic is already waiting: which jobs the experiment tests. Joe would like to test the real jobs from the transfer plan: deal intake, filing division orders, operator email replies, and checking payments. This needs the group to confirm.

To ask Claude something, say "Hey Claude, ..." and pause about 3 seconds. Or type it in the Claude panel on the right.

Listening

Claude is listening. Say "Hey Claude, ..." in the call and the reply appears here.