Executing the card

This is the loop closing. The card names a metric, a test and a sample size; the run below computes them against a real index and reports whatever comes back. If the result is null it is shown as null — a demonstrated null is better teaching than a manufactured win.

Demo 1 harness: LIVE — Live: executing against Demo 1's harness over 12 resolved questions.

HYP-2026-014 — Cross-encoder reranking over a fixed candidate pool improves context recall@5

Control

The identical hybrid first stage, k=5, no reranking. Same index, same chunks, same query set.

Intervention

Hybrid first stage, then ms-marco-MiniLM-L-6-v2 cross-encoder reranking, k=5.

Metric

context_recall@5 from atlas_shared.evals.context_recall

Tests

stats.mcnemar on per-question hit@5, and stats.paired_bootstrap on per-question recall@5

Decision threshold

+10 points, and the lower bound of the interval above zero

Items needed per arm

146

Integration status

Demo 1 present

True

retrieval.search importable

True

harness.run_eval importable

True

index built

True

ONNX models present

True

questions resolved

12 of 12

Demo 5 imports Demo1.pipeline.harness.run_eval. Neither demonstration forks the metric code: both score through atlas_shared.evals.