Approval Attaches to Premises, Not Just Outcomes

The Yes You Actually Got

You present a finding and a fix. "This field is empty, which is wrong, so I will fill it." You get a yes. What was approved is not "fill the field." What was approved is "fill the field, given that an empty one is wrong." The premise rode along inside the sentence, unspoken, and it is carrying the whole mandate.

So when the premise dies, the approval dies with it — even though nobody withdrew anything, nobody was careless, and the work would have been faithfully executed. This is the failure mode that specifically catches people who are doing everything else right.

Two Cases From This Workshop's Own Log

The anomaly that was the norm. A run reported an empty metadata field as a defect and was approved to fix it. Measuring the rest of the corpus afterward showed that fifty-one of seventy-eight items carried the same empty field. The item was in the majority — filling it would have moved it into the minority, which is the opposite of the conformity the fix claimed to restore. The approval had been given for a repair; there was no repair.

The citation that was a hedge. A run's central story rested on a public issue number, sourced from an internal document that had said the issue "is the same surface as" the problem under discussion. That is a hedge, not a citation. Fetching the actual issue showed it was about something else entirely — and the number, plus arithmetic derived from it, had already propagated into ten places across six files. A decision the approver had made at the plan gate, on that premise, had to be reversed.

What to Do When a Premise Dies Mid-Run

The tempting move is to keep going, because the work is already underway and the fix is still defensible on some other grounds you can construct. Resist that, because constructing new grounds for an approved conclusion is the exact motion that produces confidently wrong work.

The correct move is small: stop, say which premise died and how you know, and re-present. Usually the re-presentation is one sentence and the answer arrives immediately. Occasionally the whole branch is void and you have saved yourself hours. Either way the cost of raising it is bounded and the cost of not raising it is not.

Before acting on an approval, write down the premises it depended on, then go and verify the one you have not measured. The premise you have not measured is almost always the one that reads as obvious — which is exactly why it went into the sentence unspoken, and exactly why nobody in the conversation checked it.

Code

Extract the premise before you act on the yes·text

WHAT YOU SAID
  "Field X is empty, which is wrong. I'll fill it."

WHAT WAS APPROVED
  fill X  <- conditional on -> "an empty X is wrong"
                               ^^^^^^^^^^^^^^^^^^^^^
                               never stated, never checked

THE CHECK (one query, before touching anything)
  $ how many siblings also have X empty?
  51 of 78

VERDICT
  The premise is false. The item was conforming, not defective.
  Filling X would have created the anomaly the fix claimed to
  remove. No repair exists, so the approval authorizes nothing.

THE RULE OF THUMB
  Any premise phrased as "which is wrong", "obviously", "as
  expected", or "since we always" is an unmeasured claim wearing
  the clothes of a shared assumption. Measure that one first.

Exercise

Take a decision you are currently acting on and write it as an explicit conditional: "do A, given B." Then rank the B-clauses by how recently each was measured rather than assumed. Go measure the least recently measured one. If it holds, you have lost fifteen minutes; if it does not, you have found out before the work propagated rather than after.

Hint

If you cannot phrase the decision as "do A, given B" at all, that is the finding. A decision with no stated premise is one whose premise nobody can check, which means it can only be revisited by someone who happens to distrust it — and distrust is not a review process.

The Yes You Actually Got

Two Cases From This Workshop's Own Log

What to Do When a Premise Dies Mid-Run

Code

External links

Exercise

Progress

Comments 0