Reality First

This is the brief of the Reality First manifesto.

Reality First is a way of working with AI on anything you will have to answer for: a case, a contract, a project, a study. What actually happened is recorded, and that record outranks every plan and summary built on it. A machine does the recording and the summarising. The judgment and the signature stay with you.

The problem

Three failures motivate the practice. Two are old; the third is new.

Faithful records of work die from friction. The lab notebook decays into "we tried some stuff, it didn't work." The journal stops the day it becomes a chore. Writing down what happened competes with the work, and the work wins.

Every plan begins as a faithful account and then silently stops being one. Reality moves, and the plan is tidied where it should be redrawn. A plan that has drifted is worse than no plan, because it still carries authority.

Language models can absorb the first cost almost entirely. They can also, for the first time, absorb the judgment, fluently enough that the difference is invisible on the page. Left to itself, an agent takes the decisions without saying so, and stops to ask only when it lacks the information or the permission to continue. The human gets the snags of execution, and the machine keeps the judgment.

Two principles

1. Representations are answerable to traces. The explorer's rule: the chart is always redrawn from the coastline, never the coastline "corrected" to match the chart. A signed contract, a test result, a measurement are traces of what happened. Plans, summaries and to-do lists are charts: built on the traces, rebuilt from them, and overruled by them where the two disagree.
2. One node is non-delegable. Ask of any task: if someone else did it and the result were identical, would it be worth the same? Clean dishes, yes. A marathon run in your name, no. Drafting, searching, formatting and summarising are dishes. Choosing what to do, and signing that it was done right, are the marathon.

One session, six steps

1. Orient. Look at what is actually there before you trust any document.
2. Surface. Bring up only what this moment needs.
3. Form intent. Write down what you think and what you expect before you ask the machine: your registered prior. Then ask it to argue against you.
4. Act. Do the work, and give the machine every part whose value does not depend on who did it.
5. Chronicle. The machine writes up what happened. You sign it.
6. Reflect. Compare what you intended with what came out, and keep what will matter later.

The machine carries the middle. The two ends, choosing and witnessing, are yours.

Where to go

• Read the full manifesto: the argument, where the practice can break, where it does not apply, and what would prove it wrong.
• Try praxis, a command-line tool and an assistant skill that run this loop on one bounded piece of work with one signer.

Join the discussion

The practice is young. It improves where people try it on their own work and say what happened. Three things are the most useful to hear:

• where it cost you effort and gave nothing back;
• when looking at the record surprised you;
• where you were tempted to skip your own judgment and sign.

Write in the discussions on GitHub or to work@dimitri.ai.

By Dimitri Diomaiuta. This text is licensed under Creative Commons Attribution 4.0 (CC BY 4.0).