Controlled multi-match research selection

This follow-up isolates candidate selection from query formulation. Each fresh reader must issue exactly one search, q=freshwater, producing five matches. The same task and corpus as RESEARCH_BROAD_DISCOVERY_V1.md cross direct/paraphrased Beacon wording with first/last placement of the applicable research among those matches. Readers may fetch Threads but may not narrow the search.

Success requires the actual five-match response, the configured target position, retrieval of the applicable observations, a computed digest match and the correct profile/energy. This intentionally supplied query is not evidence of spontaneous search formulation. The prior problem-only trial tests that separately.

Setup failure retained

The first attempt failed while seeding the corpus, before any model started. Its underlying Sprite error was not retained by the old checked-call helper, so the cause is unknown. All five resources were deleted, with zero model usage recorded. Private evidence remains at outputs/research-selection-v1/.

The rerun at outputs/research-selection-v2/ explicitly starts the finite hub in a detached remote process and waits for all four native HTTP nodes to respond, with a thirty-second readiness deadline. Readiness and seed stdout/stderr are preserved privately. This addresses a possible startup race without claiming that it was the established cause of the original failure.

Run experiments/sprite-cohort/research_selection.py with the existing dedicated model key and Sprite authentication. The run creates five finite Sprites and attempts cleanup on exit. One same-model run per case cannot quantify reliability.

Rerun result

All four readers issued exactly one search returning five candidates and selected P1 at 169 mJ from the actual observations. Target positions were 1, 5, 1 and 5 for direct-first, direct-last, paraphrased-first and paraphrased-last respectively. All four computed the correct digest. Durations were 170.37, 38.73, 39.54 and 32.66 seconds. Observed model usage was $0.037653830; all five Sprites were deleted. Public fixtures, oracle and sanitized traces are in results/research-selection-v2/.

The direct-first reader also tried an observation ID as an artifact locator and received HTTP 400. Its report then claimed no separate artifact existed, which that response does not establish. The selection and exact Thread-text digest were correct, but this is a separate reference-handling/reporting error. The agent workflow guide now explicitly distinguishes observation IDs, digests, Thread IDs and signed artifact IDs.

These runs exercise fifth-position selection; they do not establish general order invariance, resistance to maliciously attractive distractors or large-corpus search. The query was deliberately supplied by the operator.