Skip to content
Português Request access

The method

How EnterDone is built

EnterDone is built by agents, within rules written in code, with a receipt for every step — the same way it works inside your company. That way has a name and is written down: the Suleiman Method, an engineering model for real systems built with agents, by someone who does not read code and validates by what shows up on screen. This page is the summary; the full document is at the end, for anyone who wants to use it in their own project.

The central sentence

Observe what is real, find the difference, fix the minimum, prove in proportion to the risk, and turn valuable errors into permanent properties of the system.

Or, operationally:

AI proposes. Engineering constrains. The system executes. Evidence demonstrates. The human perceives and governs the risk. Error teaches. Code learns.

A system is not correct because the agent said it was done, because the code compiled or because the tests went green. It is correct within what has been demonstrated by evidence. Authority comes from proof, not from trust in the agent.

The central loop

  1. MeasureEstablish the real state.
  2. ContrastCompare what was observed with what was believed. The most valuable finding is often the one that contradicts the hypothesis.
  3. ExplainFind the cause, without confusing correlation with causation.
  4. FixMake the smallest change that restores the invariant.
  5. ProveEvidence proportional to the risk — including trying to refute your own work and, where perceptual quality is involved, human observation.
  6. GeneralizeAsk: what class of error is this?
  7. Wall it offWhen it is worth it, turn the lesson into a permanent protection.
  8. RecordPreserve the finding so nobody has to relearn it.

The human is a sensor, not only a governor

The agent reads the code — the intent. The human sees the screen — the behavior. Some errors only appear in time, motion and perception: the skipped frame, the almost-right layout, the technically correct answer that sounds wrong. For those qualities the agent cannot declare anything proven; it says where to look and what to observe, and human verification is legitimate evidence. A trained eye is a measuring instrument, and the method treats it as one.

The 16 invariants

  1. Measure before you explain.
  2. Every important operational claim needs reproducible evidence.
  3. Absence, error and unavailability are different states.
  4. A test only demonstrates what it actually asked.
  5. The greater the impact, the stronger the proof must be.
  6. Prefer mechanical constraints to behavioral instructions.
  7. Autonomy must be proportional to reversibility and risk.
  8. Important information carries provenance.
  9. Operations at scale have limits and recovery mechanisms.
  10. Recurring errors must become permanent protections.
  11. Local evidence does not become universal law without generalization.
  12. The agent is not the authority on its own success; the observed system is.
  13. Whoever did it is not whoever verifies it, in proportion to the risk.
  14. Risk is defined by what the change touches, not by the confidence of whoever makes it.
  15. Before declaring it done, try to prove you are wrong.
  16. What only the human eye perceives, only the human eye proves.

How it shows up in the product

The method is not a text about EnterDone; it is what EnterDone enforces inside your company. Each rule below has a matching wall in the product, described on the governance page.

The rule in the methodThe wall in the product
Guardrail before instruction: if something must not happen, make it impossible.Passwords go to the encrypted vault and the code fills the field — the model has no way to see them.
Whoever did it is not whoever verifies it, in proportion to the risk.Every irreversible action stops and waits for approval from whoever holds the authority, signed (HMAC-SHA256).
Important information carries provenance.Every fact from the agent says who said it, when and where — and every answer accepts «why?».
Operations at scale have limits and recovery mechanisms.One brake freezes all autonomous action; the monthly spend cap blocks on its own; the trail is chained and cannot be rewritten.
The scoreboard is never typed by hand: numbers come from a reproducible source.Spending on each model shows on a real bill, per model, in the Console.
Autonomy proportional to reversibility and risk.Permissions per role; connector by connector, management decides: read only, ask for approval, or act.
The agent is not the authority on its own success.The agent declares itself on the web and signs every navigation; the visited site verifies it with the public key.
Language is not evidence.The «In production» section of the home lists only what the agent actually did, with a receipt for every step inside.

The full document

The whole method — 44 sections, with its origin, its convergences with established engineering and what is its own contribution — and the operational contract agents follow in every project are published in an open repository, as written, in Portuguese (an English version follows):

Author: Suleiman Ribeiro. License Creative Commons Attribution 4.0: use, adapt and redistribute, crediting the source. The method came from practice, not from research; where it converges with established ideas, the reference table at the end of the document says which.