An AI analyst you can audit.

The stations, in the order a job runs them

1. Station 00 · Doors

   Your files go in

   Where you hand over your files. A terminal command, an HTTP API, an MCP server for AI assistants and a Telegram bot: whichever you use, the same job starts.

   No model: plain code.

   Input folder

   examples/coffee-roaster

   Files handed over

   4

   Learn: Ingestion & profiling

2. Station 01 · Ingest

   Read the files

   Opens every file and finds the real tables in it. Titles, totals and several tables on one sheet are told apart; documents and emails are cut into sections. Every row remembers its file and row number.

   No model: plain code.

   Files read

   4

   Tables found

   3

   Document sections

   9

   Learn: Ingestion & profiling

3. Station 02 · Profile

   Understand the columns

   Works out what each column means. Dates, money and its currency, customers; it flags problems such as duplicates or refunds, and screens cell text for instructions aimed at a model.

   No model: plain code.

   Data-quality issues

   3

   Learn: Ingestion & profiling · Security

4. Station 03 · Metrics · METRICS_DRAFT

   Agree what to measure

   Decides what to measure, and lets you approve it. A model proposes each definition; code checks every column it names and compiles it to SQL, and you confirm or edit it before anything is measured.

   Calls a model: the planner.

   revenue

   sum of amount, per month

   Status

   assumed

   Learn: Metric layer

5. Station 04 · Analysis · MEASURE · DETECT · DECOMPOSE

   Calculate the numbers

   Every number is made here, by the database, not by a model. SQL computes each figure, tests a change against its own noise and splits it by who caused it. Each number is stored as a fact with an id, its query and the rows it read.

   No model: plain code.

   Revenue, 2026-03

   $44,100 N12

   Revenue, 2026-04

   $36,200 N11

   Change

   −$7,900 N29 −17.9% N30

   By channel: wholesale

   −$8,500 N144 108% N147

   Learn: Analysis

6. Station 05 · Index docs · INDEX_DOCS

   Make documents searchable

   Makes your documents and emails searchable. Passages keep their exact source text, page and position. Text addressed to a model is screened and locked away first.

   No model: plain code.

   Sections indexed

   9

   D1

   email-0412.eml

   D7

   price-list.pdf

   Learn: Retrieval · Security

7. Station 06 · Web lane · WEB_CONTEXT

   Look things up online

   Optional: outside context, such as an industry benchmark. Before any search leaves your computer, customer names and amounts are removed from it; fetched pages are read only by the quarantined extractor.

   No model: plain code.

   W1

   coffee-benchmarks.example

   Gap to the benchmark

   1.4 pp N270

   Source

   offline web fixture (coffeeWeb)

   Learn: Web lane · Security

8. Station 07 · Explain

   Find out why

   A model looks for why things changed. It sees the list of facts with their ids, never the raw data, and works through a few fixed tools. It writes {{N…}} placeholders, never a number.

   Calls a model: the planner and extractor.

   Cause found in

   D1 · email-0412.eml

   Graded

   document supported

   Learn: The planner (agent) · Security

9. Station 08 · Decide

   Suggest what to do

   A model suggests what to do next. What each action could be worth is a range computed by code, with the metric and the date that will show whether it worked.

   Calls a model: the planner.

   Top action

   Offer wholesale accounts a volume-tier discount to win back paused standing orders

   Impact, next quarter

   $3,995 N350 $7,191 N351

   Checked later by

   revenue >= $44,100 by 2026-09-30

   Learn: The planner (agent)

10. Station 09 · Verify

    Check everything

    Nothing reaches you until it passes. Its checks recompute every number from its SQL, find every quote in its file and drop or reword any claim they can't prove. A small judge model only confirms that a quote supports its claim.

    Calls a model: the judge.

    Numbers recomputed

    101 · 0 mismatched

    Citations checked

    6

    Claims dropped · reworded

    2 · 2

    Learn: The verifier (V1–V8)

11. Station 10 · Render

    Write the report

    Builds the final report. Each placeholder is filled with its checked fact; in the report, any number opens how it was computed.

    No model: plain code.

    Headline

    Revenue fell after March, driven by wholesale (a change of −$7,900).

    Learn: Report & fact IDs

Beside the line

• Beside the line · Workspace · job.duckdb

  The job's database file

  One DuckDB file per job, where every station saves its work. Tables, facts, document passages, claims and checks live in it, and each stage is checkpointed there, so a job that stops resumes where it was.

  No model: plain code.

  Events recorded

  31

  Learn: Report & fact IDs

• Beside the line · Model room · planner · extractor · judge

  The AI models

  Where the models work, each with one narrow job, behind glass. The planner drafts definitions, causes and actions; the extractor alone reads untrusted text; the judge checks quotes. They run on free endpoints or your own machine.

  The models themselves.

  In this run

  scripted planner (coffeePlanner), no LLM

  Learn: Model router & free providers · Security