Discussion 2026-10-15-jdu-ifee-mes

Back to the meeting notes

Answer what you can and skip the rest. Everyone's answers show under each question and update every 15 seconds.


Shared files



Docs, PDFs, images, spreadsheets, up to 25 MB. Anyone with this link can open what you share.

County data and scale

1. Which data plan are we following?
    2. Can someone get a real sample from a county archive (size, page count, file type)?
      3. Add storage to the Dell quote?

      Only needed if we keep whole county archives ourselves.

        Confirm the tool for each arm

        4. Arm 1 (one agent with tools)
          5. Arm 2 (fixed workflow)
            6. Arm 3 (team with a checker)
              7. Arm 4 (general agent with a workspace)
                8. Anything to change?

                  Which tool for each arm?

                  9. Arm 1 (one agent with tools, the control)
                    10. Arm 2 (fixed workflow)
                      11. Arm 3 (small team with a checker)
                        12. Arm 4 (general agent with a workspace)
                          13. Run all arms on the same Claude model first, then try our own model?

                            Test emails

                            14. During the test, should the "unsure" emails be sent, or written and saved for us to check?

                              Experiment 1, revised: how do we build the co-worker?

                              15. Which ways of building the co-worker should we try?
                                16. Is "one agent with tools" the right control?
                                  17. Which tools should it have from day one?
                                    18. What can it do without asking, in the test?
                                      19. Which model first?
                                        20. Anything missing or wrong?

                                          Experiment 1: which arms do we run?

                                          21. Which arms should experiment 1 include?
                                            22. Is the control right: today's flow with no settings changes and no AI?
                                              23. For the AI arms, which model size first?

                                              Large shows the best case; small shows the cost floor.

                                                24. Should the AI arms see the page image, or only the text?
                                                  25. Any lever missing?

                                                    Before experiment 1 can run

                                                    26. Who builds the single-document path?
                                                      27. By when?

                                                      e.g. Friday Oct 17

                                                        28. Should Claude propose the exact 25–30 documents now, for Reid to approve?
                                                          29. Tax statements: only 3 deals have them. What do we do?
                                                            30. Who should receive the "unsure" emails?
                                                              31. Where does the experiment run?
                                                                32. When do we meet to look at the results?