Non-scores
A cell that is decided and is not a verdict, in three reasons that are not interchangeable. Added
for the first observational board, sirens-echo's eval/self-report-2026-09-17.
Why a board needs them
A derived board asks a written challenge, so every case is answerable by construction. An observational board grades replies nobody wrote a challenge for, and a dimension can simply fail to apply: a reply that claimed no tools has no self-report to check. Collapsing that into a pass inflates the board with cells the subject never earned, and collapsing it into a fail blames the subject for the instrument. Both read as coverage, so the state has to exist rather than be inferred.
not-applicable- nothing was claimed, so there is nothing to check.unreachable- the answer exists and this instrument cannot get it.no-record- the record exists and is empty. A turn that died early produces no records either, so absence of a failure record is not evidence of success.
They license different repairs, which is why they stay apart: an unreachable cell is a telemetry
gap somebody can close, and a not-applicable cell is the corpus saying what it is.
What they change about a rate
Nothing may divide by them. counts reports scored beside annotated, and the page's headline,
progress bar and map tally all count scored over scorable. A board that seeds its non-scores
mechanically opens with a third of its cases decided, and a bar reading thirty percent before a
human has graded anything is a ceiling artifact drawn as progress. They are not deductions either,
so none of them asks for a critique, and DEDUCTIONS and NON_SCORES are separate sets for that.
Seeding them is mechanical
Which cells are non-scores is derived rather than judged, so the deriver writes them into every grader's file before anyone opens the board. Disagreement between two graders is what an inter-rater study measures; a coverage hole is not, and seeding it per grader keeps the two apart.
The exclusion needs its own check
Dropping cells an instrument cannot reach is safe only where unreachability is independent of what
is being measured. On the first board it was not: see that board's README.md and
independence.txt.