Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Angle
The end of a descriptive job is proved from the object described and never from the work accomplished: "I have read all my sources", "all my records cite one another", "every area has been read everywhere" establish that you have finished what you started, and a product can be half covered under any of those sentences. What finishes a description of a piece of software is the enumeration of that software along several independent angles — its modules, its tables, its events, its screens, its API, its other applications, its customers' words — and the finding that nothing there is still in the "not yet looked at" state. The price of that shift is that the denominator has to be built before you start, when nothing has been done yet and the list costs a lot for nothing, rather than at the finish line, where you would pick the criterion from among those you are already sure of satisfying.
Synthesis
The ideas gathered here come from a single undertaking — describing the objects, the rules and the vocabulary of a software product — and they answer one another on a point that none of them shows alone: measuring a job and measuring its object are two different things, and every convenient mechanism measures the first.
The cohesion check closes the network of records back on itself and knows only the records; the semantic search engine ranks and never counts; exhaustively reading the four places in a product measures depth and ignores breadth; and the agent, finally, confuses reporting with delivering. Four different instruments, a single blind spot, and it is structural: what was never produced shows up nowhere in what was produced.
The shared remedy always takes the same form — build a list external to the work, set it against the work, and treat the disagreement between two lists as the useful information. That holds for the coverage counter, whose score per angle proves nothing but whose gap between angles points to where to look; it holds for the set of framing questions, external by construction since it comes from support; it holds for the three-state counter, where only "not yet looked at" counts. And it holds for the moment of choice: a criterion established before anything is done is the only one that cannot be chosen for its convenience.
Tensions / contradictions
One tension runs through these notes without being resolved: enumeration presupposes an enumerable object. Describing software lends itself to it — a repository has modules, a database has tables, an interface has screens. A body of thought, an open corpus, a field of inquiry have no denominator, and the thesis loses its footing there: all that is left is the cohesion check, the very one these notes show to be insufficient.
Second tension, internal to the mechanism: the coverage counter rests on a matching of names that fails half the time, supplemented by a hand-written correspondence list. The external list that is supposed to judge the work therefore contains judgment, and an indulgent judgment inflates its own score.
Questions
- On material where no inventory is possible, what mechanism plays the part of the external list without falling back on an impression of completeness?
- Who, in an organization, has the mandate to reject a stopping criterion and extend a job everyone thought was finished?
- How do you establish that one angle of enumeration is genuinely independent of another, before treating their agreement as a confirmation?