Repository navigation
REX-11: B1 loop: prompt versions + retrieval few-shot evaluated via §5.4 champion/challenger #4366
Closed as not planned
Description
owner: rex
Epic: #4354 (REX-001). Spec: docs/specifications/review-experiment-protocol.md §7 row REX-11.
Done when (all must hold): contract review-champion-challenger-v1: >=1 challenger evaluated end-to-end with receipt, promoted/rejected by rule; test-version evaluation counter increments
Binding: §6 R-1..R-12, §8 STOPs. Ticket id = PMAT-.
Activity
REX-11: champion/challenger promotion (§5.4) + B1 few-shot guard (Refs …
noahgift commentedon Sep 25, 2026
REX-11 gate landed on rex/001 @ d869665 (to be folded into B3 as the cop directs).
champion.rs: the five §5.4 gates, each failing closed (unmeasured counts as failed); 20 evaluations per sealed test version, and refusals spend nothing (R-6); promotion is a pin bump that records the rollback sha; demotion on two consecutive gold escapes.fewshot.rs: B1b retrieval draws only from non-test gold (R-2) and never returns the query; every challenger prompt goes through the contamination guard.- defect at …line after a fence joined the last example's hunk, so a whole-text scan missed a sealed test diff in a rendered prompt. The guard now scans each fenced block on its own; RXF-003 pins this.rex challengeexits 0 when promoted, 10 when rejected, 11 when refused. Contractreview-champion-challenger-v1(RXC-001..003, RXF-001..003) passes pv validate. 3/3 planted mutants (block scan removed, precision gate weakened, budget off-by-one) turn RED.noahgift commentedon Oct 4, 2026
Folded into #4354 as a checklist line (ISSUE-TREE-001: a family of rows becomes one parent with checklist lines). Work and status are tracked there. aprender-wtix-even