UNIFIED STATE SYSTEM
UNIFIED STATE
LOVE & FREEDOM — ALWAYS
UNIFIED STATE · Tools

Carrier Retrieval Lab 2 — questions first, answers on trial

October 5, 2026
Unified State
Carrier / Retrieval Lab R02

Carrier Retrieval Lab · Study R02

Questions first.
Answers on trial.

Explore what happens when information needs are written before the concepts are revealed—and when a better-looking result must still help someone answer the question.

48tasks across six domains
146qualified assertions in a separate research overlay
36questions reserved for final evaluation

The lab compares a development-selected text or dense baseline with a small contribution from typed relations. Every task has explicit answer obligations and conditions. The public explorer keeps the selected methods’ top five definitions beside the question, then lets you inspect source provenance, full relation qualifications, recorded reader answers, and their assessments.

The retrieval result did not establish an advantage for the typed method. Its final pooled nDCG@5 was 0.670678, compared with 0.674080 for the selected dense baseline. The paired difference was -0.003402, with a 95% interval of [-0.010205, 0.000000]. This is a result about one frozen experiment, not a verdict on every use of a graph.

Answers were tested separately. The separate answer endpoint also does not establish an advantage for the typed-context assignment. The recorded mean score difference was +0.013889 on the 0–1 rubric (95% interval [-0.003472, 0.031279]). Both endpoints, their jointly adjusted tests, and all recorded answers are available in the report.

Open the Retrieval Lab →Read the technical report

Four ways to inspect the evidence

Start with a needRead the task, conditions, and prewritten answer obligations.
Compare contextInspect both top-five lists and the supplied definitions.
Follow a qualified pathKeep the predicate, direction, literal evidence, and limits together.
Read the answerCompare the recorded response with its evidence and assessment.

Twenty-four tasks are AI adaptations of attributed public human questions; another 24 are original AI-authored scenarios. These origins remain labeled separately. The browser filters the frozen collection locally. It does not upload your wording or call an inference service.

A typed path is a retrieval hint, not an entailed answer. The sparse overlay reaches 3.77% of the corpus and changes the ordered top five on only two of the 36 final questions. The questions, relevance judgments, readers, and graders all carry the study’s stated model-family and sampling limitations.

Technical reportFrozen protocolTasks and attributionRetrieval resultsAnswer resultsFull experiment kitReader challenge

The full experiment kit includes the pinned local model and its assets. Carrier’s 6,257 checked records remain unchanged; their existing author credits, source attribution, and terms are preserved. Public question sources retain their individual attribution and API-reported CC BY-SA licenses.