# Japa two: Cruise research dossier Prepared 11 November 2026. This is a research proposal, not a report of completed experiments. The paper you meant is **GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning**. GEPA already improves prompts while keeping model weights fixed. The opportunity for our paper is more specific: > Can a fixed agent make sustained progress toward a user's goal when a separate planner sees only delivered artifacts and explicit requirements, and can change only subsequent prompts and bounded textual memory, without receiving benchmark scores, gold answers, execution traces, or source code? This is a hypothesis to test. It does not yet establish that Cruise outperforms GEPA. Artifact inspection also involves judgment; the precise claim is **adaptation without an external evaluator in the loop**, rather than improvement without any evaluation whatsoever. Independent measurement after a run is necessary to publish a credible performance claim. Read these documents in order: 1. [GEPA reading notes](01-GEPA-reading-notes.md): the abstract, method, experiments, appendices, limitations, and implications for our study. 2. [Research plan](02-research-plan.md): the contribution, formal method, principles, hypotheses, novelty risks, and staged work plan. 3. [Experimental protocol](03-experimental-protocol.md): tasks, fair baselines, information restrictions, budgets, statistics, ablations, and release requirements. 4. [Paper blueprint](04-paper-blueprint.md): titles, proposal abstract, section outline, planned figures and tables, and rules for writing the eventual results. 5. [Planner prompts and examples](05-planner-prompts.md): a concrete starting protocol and illustrative follow-up prompts. 6. [Cruise evidence and related work](06-evidence-and-related-work.md): what the current implementation demonstrates, what is proposed, and the closest competing literature. 7. [Initial manuscript](07-initial-manuscript.md): connected Introduction and Method sections for the untested text-artifact proposal, with verified references and an explicit separation between adaptation and independent assessment. The [sources folder](sources/README.md) contains the downloaded 87-page conference paper, page-marked extracted text, and a source manifest. [references.bib](references.bib) contains starter citations. The [development task set](development-tasks/README.md) supplies twelve hand-authored fictional text-only assignments, four per proposed task family, with complete participant packets and separate assessor-only guides. Twelve remains an unfrozen development default, not a registered sample-size decision; these are not held-out tests or experimental results. The [baseline prompt assets](baseline-prompts/README.md) provide the shared initial worker instruction, B1 generic continuation, and separate B2 critique/revision templates, with B0 retained as a single-call compute-allocation control. All templates are untested; B2 is a Self-Refine-style study adaptation rather than a verified reproduction. The [treatment ablation assets](treatment-ablations/README.md) provide a matched treatment reference and complete no-memory and no-principles planner templates. They document retained validity rules, input/output differences, prompt-length reporting, and the limits of each untested comparison. ## How to interpret “without improving the code” The primary study uses text deliverables and makes no code edits during adaptation. In the optional software study, the worker can edit the requested application, but the planner cannot inspect its source or modify the agent, harness, tools, or model. These are separate claims and separate tracks. The software track is unnecessary if the intended restriction forbids even application-code changes. ## Recommended first experiment Run a small, preregistered pilot comparing artifact-guided prompts against generic continuation, ordinary self-refinement, and a reference-free prompt optimizer. Use the same worker, equal total budgets, fresh tasks, and scoring that stays hidden until all runs finish. Launchboard is a useful worked example; use newly authored tasks for the decisive tests. The practical contribution should survive these questions: does the planner use the actual artifact, preserve the user's requirements, outperform simply spending the same compute on the worker, and avoid turning its own assumptions into permanent rules? ## Status - Read the complete conference-paper text and appendices, including the image-only NPU prompt. - Inspected the local Cruise planner and historical Launchboard records. - Screened relevant primary literature; full reproduction of competing methods remains future work. - Created this plan and templates. No experimental runs, performance estimates, superiority claims, publication submissions, or product-code changes were made.