ESSAY
Two paths, one ledger.
People use the apps. AI uses the tools. Both write the same database, so a deal underwritten in Claude shows up in forVEX Edge with nothing copied by hand.
I underwrote a deal in Claude.
I didn't export a spreadsheet. I didn't paste numbers into a form. I asked for comps, rent, flood, and a score the way I always do. The server ran the math. Then I opened forVEX Edge.
The deal was already in the pipeline.
Same address. Same inputs. Same numbers. Nothing copied by hand.
That's the whole architecture once you strip out the diagram. People use apps. AI uses tools. Both paths write the same Postgres.
Two paths
Call the left lane the AI path.
A model (ChatGPT or Claude today) loads a skill. A skill is a written playbook for one job: underwriting, a rehab estimate, appointment prep, a competitor sweep. The skill doesn't invent the math. It tells the model which tools to call and in what order.
Those tools live on REcontrol, the MCP server. Named functions. Typed inputs. Validated outputs. Comps. Underwrite. Rent. Flood. Save deal. Preview estimate. Log activity. The model brings judgment and sequence. The server returns the numbers.
Call the right lane the app path. That's the forVEX suite.
Edge is the hub: pipeline, notes, tasks, pulse, rental ops, market and competitive views. Deal is underwriting on the phone, standing in the driveway, on the same engine the chatbots call. Build turns room-by-room condition into a line-item rehab budget. That same engine is packaged for MCP, so a chatbot gets the estimate the app would give. Scan files paperwork against the right deal or property. Train mostly stands alone for seller practice and writes little back.
Two paths. One database.
Why one ledger matters
Most stacks end up with two systems that almost agree.
You underwrite in a chat. You re-key it into the CRM. You estimate rehab in another tool. You reconcile later, or you don't. The phone version drifts from the desktop version. The model's last answer drifts from both.
One ledger kills the hand sync.
When the underwrite tool writes the deal, Edge sees it, because Edge reads the same tables. When Deal on the phone scores a property, it isn't a private pocket calculator. When Build prices a kitchen, the MCP estimate tools aren't a second opinion with different math. Comps, underwrite, and rehab numbers come from the same place, whether a person tapped the app or a model called the tool.
That's also why the cut from Journal 10 still holds, and I won't restate it. The ledger is the brain. The rented chat windows are hands. Hands are useful. Hands are not the system of record.
What this piece adds is the wiring. The hands and the apps share a write path on purpose.
The design rule
One rule keeps the two paths honest.
Numbers come from the server engines, not from model memory.
The model decides what to ask and how to explain the answer. It doesn't get to remember last week's ARV and reuse it as truth. ChatGPT and Claude stay interchangeable for this work because they call the same tools and get the same comps and the same deal score.
Portability across surfaces is a related idea, and Issue 13 already owns that story. The point here is narrower. If the number didn't come from the engine, it doesn't belong in the ledger.
What this unlocks, and what it doesn't
What it unlocks is continuity.
You can think in a chat and operate in Edge. You can stand in a driveway with Deal and know the pipeline won't need a cleanup pass. You can let a model run a competitor sweep through the same governed tools and watch findings pile up in the market views you already use. The operator stays in one business, not three tools pretending to be one.
What it doesn't unlock is a staff that works while you sleep.
Analysts and agents are different jobs. ChatGPT and Claude are analysts. You ask, they pull, they explain. Standing roles that act on a schedule are a separate build, and a separate Update when that roster is ready to show. This piece is about the ledger those roles would share, not the org chart.
History is a later cut too. The database keeps more than current state: events, memory, market signal, trust graded by source. One sentence is enough for now. The four-layer story gets its own essay.
For today the proof is simpler.
Two paths. One ledger. Underwrite in chat. Work it in Edge. Nothing copied by hand.
How many places does one of your deals live right now?