Purpose and design
The main locality study places the root declaration at the front of the context. It therefore does not isolate absolute relevant-path position from distance to that declaration. This completed follow-up tests a simple alternative explanation: late paths should benefit when the unchanged root declaration is moved to the end.
The study uses 48 fresh forests with 512 parent records, four roots, and eight-edge queried paths. Twelve views cross compact or dispersed spacing, early/middle/late placement, and front/end root declarations. The compact span contains eight records; the dispersed span covers 256 record positions. Forest blocks and conditions within blocks were randomized. The 576 planned requests all returned usable responses.
Within each forest, the views preserve graph facts, literal questions, candidate order, state length, and billed input-token counts. All 288 front/end pairs change the declaration position without changing the remaining record order. Relevant-path selection still requires the exact graph. Moving the declaration also changes its distance from the query and other records; it is not an isolated intervention on one internal position variable.
Root results
| Path spacing and position | Front declaration | End declaration |
|---|---|---|
| Compact, early | 36 / 48 | 36 / 48 |
| Compact, middle | 17 / 48 | 12 / 48 |
| Compact, late | 23 / 48 | 15 / 48 |
| Dispersed, early | 9 / 48 | 11 / 48 |
| Dispersed, middle | 12 / 48 | 10 / 48 |
| Dispersed, late | 11 / 48 | 7 / 48 |
The primary interaction is the end-minus-front accuracy difference for late compact paths minus that difference for early compact paths. It is −16.67 percentage points, with a 95% forest-bootstrap interval of −37.50 to +4.17 and Holm-adjusted p = 0.389. The locally specified two-test family also includes fourth ancestor, with an interaction of −2.08 points [−20.83, +14.58], adjusted p = 1. Historical analyses use 10,000 bootstrap draws and 10,000 sign flips under the assumptions described in the methods. Intervals are not simultaneous.
Every direct-parent response is correct (576/576), but these are repeated views of 48 forests. They are not 576 independent demonstrations of general correctness.
Interpretation
Moving the declaration to the end did not produce the predicted improvement for late paths in this sample. The interaction interval is inconclusive. It neither establishes equivalence nor proves that a nearer declaration harms performance.
The question already names all four candidate roots, so the declaration is redundant evidence. Its relocation cannot rule out other evidence-distance, parsing, or positional explanations. Fourth-ancestor candidates retain the historical depth shortcut. This control does not support the simple predicted improvement, while leaving other proximity explanations unresolved. It does not identify architecture.
Verification and evidence
The publication review independently reconstructed 3,456 expected question answers from the literal inputs, verified every request against its frozen plan, and checked all representation invariants. Actual dispatch order matches the plan. Local freeze and request records establish recorded sequencing, not external preregistration.
The evidence download contains this study under data/header/, its protocol and manifest, the unchanged historical analyzer dependencies, and the independent literal audit. These 576 requests and 48 forests are supporting evidence outside the main three-study inventory; they are not added to any main estimator.