Study AG · Asking & calibration
The discourse gap closes on every tier; the visibility clause bites
The anaphora hatch
Study Overview
The Untested Border
One border of the ask-path map was never tested: Study X's discourse construction, where "undo that" against a carrier-less editor failed 0 of 144 with every failure a valid silent guess. The shipped ask sentence arguably covers it (a dangling referent is "not stated in the request"), so Study AG pointed the sentence at it: the carrier-less control, the control plus the shipped hatch, and the last-edit echo plus the hatch, across three models and 1,296 session steps.
The Hatch Covers Discourse
The control replicated Study X to the digit. With the hatch, 138 of 144 silent guesses became questions, on every tier, and the asks show genuine recognition: "the request refers to 'that same node' from a prior instruction, but no such node is identified." Unlike ambiguity, which only the frontier tier notices, the discourse gap is letter-covered, and every model protects itself.
The Tax Check Fails
Then the tax check failed on all three models, and the ask texts locate the mechanism precisely. With the echo supplying everything (the node id, the key, both values), models still asked on roughly 70% of anaphora cells, quoting the echo's own node id while refusing to proceed because that node "is not visible in the current view." An anchored patch needs only the id; the echo-without-hatch arm solved 48 of 48 under identical views.
The Collision
This is the sentence's visibility clause colliding with skeleton views: the rule's letter says an unseen node warrants a question, so a fully specified, solvable edit becomes an interrogation. Ordinary steps in the same arm, whose views carry their targets, produced zero false asks in 288 cells.
A Dependency, Sharpened
The guidance sharpens into a dependency: the hatch's zero tax is conditional on the focus-ids contract. Views that cover every mentioned node make the hatch free (measured three times now); outline-style protocols need target-carrying views, a self-serve view tool, or a re-scoped sentence that gets its own registered test first.