Rendered from docs/obligations/0016-a-cue-has-no-measured-effect-on-traversal-precision.md in the Headwater corpus. Every document on this half of the site is typed by the taxonomy the descriptor names: corpus.json.

A cue has no measured effect on traversal precision

Context

Q20 rules where scent lives and gives the referring end the cue. A cue should raise traversal precision over the summary fallback.

Obligation

A paired campaign over one corpus is the instrument.

Discharge

When this record was written, no campaign had run, and this corpus authored no cue. If a cue does not raise precision, it is authoring cost with no scent gain. The honest response is then to remove it rather than to make it longer.

A campaign has run, and it did not grade a cue. The batch of 2026-09-30 graded the navigability selection in both arms of the campaign tier (#1384). The present arm satisfied 69 of 90 (76.7%, 66.9% to 84.2%). The absent arm satisfied 68 of 90 (75.6%, 65.8% to 83.3%). The difference is +1.1 points, in a 95% Newcombe interval of -11.3 to +13.5 points (the result). So the governance layer did not raise traversal precision by more than about 13 points.

No cue is removed, because this corpus authors none for a session to follow. The one cue in the corpus is on the edge of this record to HW-DR-0020, and no navigability probe traverses that edge. The other cue: lines are examples in spec 1 and spec 2 and a declaration in the base package. So the batch measured the summary fallback in both arms and not a cue. The removal that the paragraph above names has nothing to act on. For the same reason, no headwater capture trend is cited: no hand-written cue was removed to cost or to save.

The reading has limits. 8 of the 90 present-arm sessions and 0 of the 90 absent-arm sessions reached this repository from outside their workspace. That count uses the rule in the Limits section of the evaluation: a call names a copy of this repository, a headwater binary or GitHub. It does not count every path outside the workspace. The corpus was not frozen between the two arms' sessions. docs/spec/09-open-questions.md changed after the recording, so each verdict is graded over a moved read set. The evaluation of that batch states every limit and the cost. This record stays open by the owner's ruling of 2026-09-30, pending #1472.

The campaign of 2026-10-03 graded the navigability selection in six arms, and it did not grade a cue. The present arm satisfied 95 of 150 sessions (63.3%) and the absent arm 121 of 150 (80.7%). The difference is +17.3 points for absent (+7.2 to +27.0). One probe supplies it, and it is a measured loss from the layer on that probe. The key of the governs line probe holds against HW-DR-0104: 0 of 30 present sessions satisfied it and 26 of 30 absent sessions did. In present, 22 of the 30 sessions answered with zero tool calls, so they answered from what was in context and did not look. Without it, both arms are 95 of 120, a difference of +0.0 points (-10.3 to +10.3). The authors chose that exclusion after they read the data, so it is a sensitivity reading and not a result. This corpus authors no cue for a session to follow, so the batch measured the summary fallback. The evaluation states every figure and the limits. The next run uses the successor of that probe, HW-PROBE-a-session-says-whether-a-command-or-a-hand-writes-a-governs-line, and its result does not compare with this campaign (#1709). This record stays open. The earlier reason, the owner's ruling to wait for #1472, is spent, because the campaign ran.