Request-to-evidence coverage

Requested outcome	Retained evidence	Final disposition
Pull latest changes	Exact source revision in configuration.json and run-plan.json	Completed before testing in an isolated campaign worktree; unrelated working changes preserved.
Better than current Qwen3.8 27B on this RTX 5090	summary.json, native quality and capacity artifacts	No better replacement established. Signal is the strongest new efficiency research lead; incumbent retained.
Test four variations, including fine-tunes	Four original preflight and reasoning artifacts; recipes.toml	Signal, Swift, Qwopus Flash, and 20B Minitron loaded and passed six functional checks; two additional no-spec controls tested.
Explore shorter reasoning	Signal/Swift/incumbent reasoning artifacts	Signal/Swift used 16.5%/16.1% fewer total completion tokens on ten questions, with 9/10 versus incumbent 8/10 strict passes. Separate reasoning-token counts unavailable. Not broad quality proof.
Hardware-specific recipes	feasibility.json, configuration.json, both recipe registries	Four exact single-5090 recipes loaded. Candidates configured at 64K versus incumbent 262K. Actual preflight input was 47,349 tokens, not a validated 57K input plus 8K output envelope.
Stop at 50% Codex usage	campaign-state.json	Checks observed 19% initially and 28% at final verification; threshold not reached.
Unattended selection/promotion	run-plan.json, summary.json, restoration.json	Authority present; no challenger cleared replacement gates. Selected incumbent restored healthy, all candidate containers unloaded, protected process preserved.
Fix forward with bounded attempts	friction-log.md, independent review records, tracked lifecycle/evidence-gap ticket	Lifecycle and Windows regressions repaired with independent tests. Model failures investigated with no-spec, revised-output, syntax, and budget diagnostics; original failures preserved. General lifecycle release held on shared locking and immutable baseline process attestation.

Completion means bounded research, four-profile measurement, selection and restoration are complete. It does not mean a new model qualified, a production SWE/client workflow passed, 262K challenger context was tested, or the lifecycle patch is approved for general release. No test result hides these gaps. No new promotion occurred.