Follow one call
Add a call, clear the gate, watch it being read, check a claim, join a pattern and make the call. Each stop ticks when it has happened.
Drop in a consented call. It is transcribed, read and graded on this machine, and every learning links back to the line someone said. Here is one company's calls, followed all the way to a decision.
Add a call, clear the gate, watch it being read, check a claim, join a pattern and make the call. Each stop ticks when it has happened.
The decision the calls point to, the evidence behind it, and whether that evidence is strong enough to back you.
What every call is held to. Each line is a test, and a miss is shown as a miss, never hidden.
Running it at your company? Nine steps, from install to your first Monday digest. Each one ticks itself when it is done, and nothing is shared with anyone else's install.
Open the checklist. What needs you, what is being read, and what you are walking into next.
Every file that arrived, by upload or in a watched folder, with its consent state. Nothing is read until consent clears it.
| File | Company | Consent | Status |
|---|
A recording lands in your watch folder, or you upload it. It is hashed and stored, then it waits: nothing is read until you record that the people on it were told.
Speech becomes numbered lines on this machine. Each line gets its speaker and its time, so anything written later can point straight back to it.
A language model on this machine reads the numbered lines and writes down what matters: pains, requests, budgets, objections and decisions. Each one must cite the lines it came from, or it is refused.
A frozen grader scores each learning by what it cost someone to say it, from E0, an inference, to E5, a signed commitment. The same learning gets the same grade every time.
One thread per company. Dormant is computed: an open thread with no call for 90 days. Won and lost are yours to mark; nothing moves a thread for you.
Each learning gets a short statement, and the same statement joins the same signal, whichever call it came from. A signal's stage comes from rules over grades, calls and dates, never from the model.
Open decisions are drafted from what was said. You give each option a probability and say what would prove you wrong. The label says whether the evidence backs you; the model never decides.
Position · drafted from two calls
Names said in calls. Two names become one only when you merge them, and a merge can be undone.
| Name | Kind | Calls |
|---|
Ask finds the lines that answer you and cites each one. It does not write answers. For a written answer, ask your own Claude or Cursor through MCP: it reads the same cited lines.
Ask a question, or pick one of the examples. Results are the lines themselves, strongest first.
You can ask from your own Claude or Cursor as well: it reads the same cited lines, and it can only read.
What every call is held to. A miss shows as a miss, beside the passes. The figures come from this sample's made-up calls.
Four of six checks pass. The two below target say so.The control swaps four definitions on purpose. The scorer has to rate it low, and does.
163 hand-graded learnings. The grader rates plans higher than a person does; that gap is reported, not hidden.
Running this at your company. Each step ticks itself when what it describes exists; there is nothing to tick by hand.
The machine's health, kept out of your way: what is running, what is waiting, and anything that failed, with a plain note on what to do about it.
Every company, person and call in this sample is made up. Grades on the newest call are illustrative.