D538 - Two thirds of the ask list is forty sentences

decided - 2026-09-04

orbistoun-cli questions prints one line per function per open question, so the list is as long as the number of entries resting on something unverified. It reports 719 open questions, and obSCEne's backlog 022 is generated from it.

Grouping them by the sentence they are written in:

719 open questions
275 premises
 40 premises carry 484 of the 719          (the other 235 rest on one entry each)
  6 premises carry 355                     (49% of the list)

The heaviest is one sentence under fourteen functions across two libraries and 872,904 calls: "Modelled on the POSIX call of the same shape. The correspondence is inferred from the name and the guest's usage, never verified against the target library."

Why the shape matters more than the count

Fourteen identical asks are not fourteen pieces of work. A premise shared across a family is answerable by sampling it - two or three functions chosen for how cheap they are to call - and the result speaks to the shared claim. Read one line at a time, that structure is invisible, and a console operator either enumerates the whole group or skips it.

Being exact about what a sample buys, because the temptation is to overclaim: measuring sceKernelWrite settles sceKernelWrite. For the other thirteen it moves the premise's credibility, which is precisely what "never verified against the target library" is about - and a single counter-example refutes the group outright, which is the cheaper outcome to go looking for.

Deduplication, not classification

orbistoun-turn refuses to decide what a question means from its prose: a rule over words fails silently and reads exactly like a question nobody can act on (D356). That still holds, and this does not weaken it, because nothing here reads a question. Two group only when they are the same sequence of words - case, spacing and punctuation forgiven, and nothing else.

So the grouping is deliberately unable to do the useful-looking thing. Two entries asking a closely related question in different words stay apart, because the differing word may be where the difference is. The 275 is therefore a floor: the number of distinct things unknown here is no larger, and may well be smaller.

The defect it found on the way in

Fourteen entries meant one sentence; ten ended the first clause with a full stop and four with a semicolon. Under word-for-word grouping that is one premise with two wordings - but under any exact-text grouping, including the census that started this, it is two. My own count said ten functions and 872,301 calls; it was fourteen and 872,904.

The fix is the data, not a fuzzier match, so the four now read like the ten - and no_premise_in_the_knowledge_base_is_written_two_ways gates it. That test went red on the data that prompted it, which is the only evidence that a guard works.

The second defect, which the mode wrote itself

--top was already applied to the queue, and the grouping was added after it. So the first version of this mode grouped a shortened list: --premises --top 1 reported that one function rested on the heaviest premise, with 806,669 calls behind it, under a header reading 3 premises behind 3 open questions. The truth is fourteen functions and 872,904 calls.

Nothing about that output looks wrong. Same words, same confidence, a number that is a fact about --top presented as a fact about the knowledge base - which is exactly what principle 3 forbids, in a tool rather than in a stub.

Grouping now happens before truncation, and --top shortens the list of premises rather than the premises. crates/orbistoun-cli/tests/premises.rs holds two guards: the heaviest premise must be identical with and without --top, and the header must not move when the window onto it does. Both were made to fail by reinstating the ordering, and both fired.

The guards assert against the tool's own full-run answer rather than against a count written into them, because a count would drift with the knowledge base and start passing for the wrong reason.

What this does not do

Move the wall, or establish anything about the platform. It is a claim about the ask list being countable: the same failure D537 found one level down, where records said nothing while the code held the reasoning. Here the records say the right thing 719 times and the list never admits that it is saying it 275.