Grading surfaces

Two surfaces, one set of rules. annotate grades in a terminal and serve grades in a browser.

One set of rules

The same loop against the same committed YAML, each rule in one place per surface:

a label must belong to the case's label set
a deduction needs a critique
an evidence span must appear verbatim in the output
the file is rewritten after every decision, so an interrupted session keeps its work
a verdict button or its profile key posts {id, label, critique, evidence} and shows the server's refusal verbatim rather than copying the rules here, and only a live session renders the controls at all, because an export has nowhere to post

What the browser changes

One thing that matters. annotate asks the grader to retype the span and loops until it matches, because a retyped quote gets edited by the hand retyping it. A page anchors the span by selection, so that failure cannot occur and the verbatim check stops being a typo guard. It stays in place as the guard against a page sending a span from the wrong case. serve hands the page housecast.grading.v1, carrying the profile's own keystrokes so one-key grading survives the move.

Two graders on one board

--grader kai writes annotations.kai.yaml rather than the shared annotations.yaml, so two testers stop overwriting each other. The name is validated as lowercase letters, digits and hyphens because it becomes a filename, and it is written into the file under a grader key because a filename is the first thing a copy or an export changes. grade disagreement takes --annotations once per grader, reads that key ahead of the filename, and rates only the cases every grader reached: one nobody finished is incomplete rather than folded in, and nothing compared reports no rate rather than agreement. Required --tester declares the study's graders, so a glob catching a calibrated grader refuses rather than counting. Counted files are named.

The order a pass takes

a grading session opens as a deck, one card, arrows or j/k to move, and d switches back to the scroll an export opens in
annotation-queue.csv beside the dataset sets the order on both surfaces, found the way pin.yaml is. The ranking is a view, so dataset.yaml keeps derivation order, a queue naming a case it lacks refuses rather than ordering a subset, and unranked cases trail and are counted. It outranks the entity-major order serve otherwise uses, trading the charter locality annotation_order exists for against covering the least stable cases first, and --no-queue keeps entity-major on either surface

annotate is not deleted

It keeps working, and the page has now carried decisions end to end.