Show the First Refusal Before You Retry
Picture this review with me, as a teaching scenario. The design lead's cursor hovered over the Approve button. The second model pass agreed with our preferred label. The first pass had refused that same recovery modal. I asked who owned the click if a user got stuck. The scenario says the design lead owned the decision. I asked where we could still undo the approval. The room went quiet in that…
In the review process, the design lead hesitated before approving a recovery modal, as it had initially been refused in the first model pass. The team discussed who was responsible for the decision, and it was agreed that the design lead owned the decision. The group recognized that a cheap retry could make agreement appear as proof, but agreement and evidence were not the same thing.
They considered how a second pass could erase dissent and make the card look clean, hiding the refusal. The owner of the decision should not be the model, but the design lead. The consequence of this setup is a recovery path that hides the refusal. The stage of reversibility should occur before the card is marked as approved. After the approval click, engineering may consider the copy as settled, and it may be too late to walk back the approval.
The debate over retrying has been a recurring theme in public developer discussions. The urge to retry should be seen as the design problem to address. A nod is a sign of mood, but a stored refusal is a record. The decision should be named before the model speaks. Begin with a clean slate, even if the file already exists. Write the decision in a single sentence that any outsider could understand.
Do not start the process by asking the model for its opinion. In the example scenario, the question is posed: "May we replace the refused recovery label with the team's preferred label?" The owner, the user, and the potential harm must all be identified. The card is presented for the team to save first. This card includes details such as the decision ID, the owner (design lead), the user moment (account-lockout-recovery), the proposed action (to replace the refused label with the team's label), the potential harm if the action is incorrect (the user cannot differentiate between refusal and success), its reversibility until the review card status is approved, and information about the first refusal.
The refusal is marked as kept verbatim and contains a summary of the missing elements: the verbatim label, the absence of an undo path, and the missing reason. The source handle for this refusal is recorded as 'pass-1-output.txt', and the refusal is marked as non-overwriteable. The stops, or conditions that would prevent a retry, are listed, including the absence of a verbatim first refusal, the absence of the undo gap, the absence of the reason for the refusal, and the lack of a named decision owner.
Noise factors that do not contribute to the decision-making process include model confidence scores, alternative wording variants, and praise for the team's label. The only acceptable outcome is a clean exit, where all the necessary fields are present and correctly filled. If any of the required fields are missing, the command should fail.
The flow of the process is outlined, starting with a user in lockout mode, requesting a recovery path, storing the first refusal before any retry is offered, and then checking for stops before an optional retry challenge. The owner ultimately approves or rejects the changes. The design ensures that the model's nod is not considered a part of this flow.
The refusal is locked in place before any retry is attempted. When checking the card, verify that the first refusal is present and marked as non-overwriteable. If these conditions are not met, the retry is not considered a valid review but an erasure of the critique.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
