# Bare agent versus the shipped fix command

Frozen on November 8, 2026 before collecting either arm or any comparative ballot.

Use all ten original Next.js projects from the previous repo study. Each arm starts at the same committed seed in a separate disposable checkout. One trajectory per arm per domain. Reuse the Codex model, reasoning setting, restricted editor permissions and installed application dependencies. No other project, kit or evaluator evidence is readable by the bare editor.

The bare arm receives exactly `make this look like a real product` as its single user request. Supply no screenshot, task hint, target route, critique, intervention, handoff or second turn. Measurement occurs after the CLI exits and is never returned to that editor. The paid arm invokes the actual public `unvibed fix --route=<route> --agent=codex` from the same initial repo using the installed paid kit. The command controls its own render, critique, edits and termination. This compares two product workflows with different amounts of work; it is not a same-budget comparison.

Only after edits are finished, the common observer separately captures the specified route on desktop (1440 × 900) and phone (390 × 844) and runs its protected interaction on each viewport. Keep the first completed admissible attempt. Preserve infrastructure failures, allow one identical-seed retry for an infrastructure failure, and never rerun an arm because of its design grade. Report render failures, incomplete trajectories, failed interactions, unchanged sources and CLI stop reasons. No mixing of results from different agents.

Present the two final screens side by side to a blind vision judge, first bare left and then paid left in independent calls. Both receive the same neutral reference observation and browser-verified outcomes. Include the other nine final screens from each condition as two separate anonymous galleries. Use the previous nine visual dimensions plus cross-domain distinctiveness. Normalize the 1 to 7 preference scale so higher always favors the paid loop. A stable win or loss requires directional agreement across both orders; otherwise mark tied or unresolved.

## Clear-win decision rule

A clear win requires all ten completed pairs, at least seven stable paid wins, no more than two stable bare wins, a mean normalized preference of at least 4.75/7 and all 20 final paid interaction checks passing. Report all results even if this rule fails. Evaluate Codex independently. If Claude Code is authenticated, run its own complete independent study under the same rule and never pool the agents.

The previous equal-budget repo-loop study and earlier one-shot HTML comparison remain secondary. Preserve the preregistration, raw tool receipts, accepted source, checks, final PNGs and judge ballots.