# Existing Cruise evidence and closest related work ## Current local implementation The following observations describe files inspected on 11 November 2026. They do not establish the exact implementation used in the historical export. | Local evidence | Observed behavior | Research implication | |---|---|---| | [CodexPromptPlan.swift](../Mac/leanring-buddy/CodexPromptPlan.swift), `input`, approximately lines 74–95 | Planner context includes the saved goal, latest user message, completed-work reports, recent conversation/progress, latest result, and previous follow-ups | Current Cruise is report-aware, not a strict delivered-artifact-only observer | | Same file, instruction around lines 98–152 | Planner chooses a specific remaining increment; may inspect up to three local files; cannot run tests/builds/browsers/web searches or implement work | Existing source-file access exceeds the proposed strict research boundary | | Same file, optional `edit_prompt` | A planner can propose a revised saved goal, subject to validation and integration checks | The proposed immutable-goal/separate-memory protocol is a research variant, not an existing guarantee | | Same file, `validateNovelty` | Normalized exact-repeat detection | Useful engineering guard; does not prove semantic novelty or useful progress | | [Closed-loop fixture README](../Mac/checks/closed-loop/README.md) and [runner](../Mac/checks/closed-loop/run.py) | Ten worker passes, separate read-only planning calls, optional goal edits, saved snapshots, replayed conversation | A starting point for an isolated experiment; not evidence of live desktop queue reliability | The production planner's specific operational rules are not being edited by this dossier. A research wrapper should enforce the strict observation interface independently and freeze it before measurements. ## Historical Launchboard run Sources: [run context](../launchboard/results/cruiseharness/README.md), [submitted-prompt index](../launchboard/prompts/cruiseharness/README.md), [final notes](../launchboard/results/cruiseharness/prompt-notes.md), and [challenge description](../launchboard/README.md). What is documented: - A sequence of ten completed work passes with planner-generated follow-ups after the seed. - A human-supplied initial work prompt and a separate owner correction before pass 5. - Separate planning calls, including a proposed final follow-up that was saved but never executed. - A local prototype and intermediate snapshots. - A planner-generated “no persistence” restriction that became durable guidance despite not being an original owner requirement. - Missing exact underlying model identifier, historical harness commit, usage, and cost in the export. What this supports: a feasibility example of a planner continuing work and revising textual guidance, plus a concrete memory-drift failure to study. What it does not support: autonomous zero-intervention performance, external-evaluator-free artifact-only operation, statistically reliable quality gains, lower total cost, revenue generation, or superiority to another optimizer. A local prototype with a revenue goal did not demonstrate actual revenue. Do not retroactively repair the historical artifacts for the paper. Quote and label the human steering, distinguish seed prompts from generated follow-ups, and preserve the unexecuted final proposal as unexecuted. ## Related-work map GEPA received a complete paper reading. For the other papers below, this dossier uses primary abstracts and targeted method descriptions for screening; a full implementation-level reading is still required before experiments or final novelty claims. This is a focused map, not an exhaustive literature review. | Work and primary source | Established idea relevant to us | What our comparison must establish | |---|---|---| | [GEPA](https://arxiv.org/abs/2507.19457), ICLR 2026 | Reflective prompt evolution with rich feedback and Pareto-based selection | Whether restricted artifact access and no scored selection give a useful tradeoff; fixed weights alone are not new | | [Self-Refine](https://arxiv.org/abs/2303.17651), 2023 | Iterative feedback and refinement of a model's own outputs | Whether explicit next-action principles and memory add beyond ordinary critique/revision | | [Reflexion](https://arxiv.org/abs/2303.11366), 2023 | Linguistic reflection retained in episodic memory from varied feedback | Whether our evidence restriction and provenance treatment change outcomes | | [TextGrad](https://arxiv.org/abs/2406.07496), 2024 preprint | Textual feedback used to optimize components of compound systems | Whether artifact-grounded control adds beyond generic textual optimization | | [MIPROv2](https://aclanthology.org/2024.emnlp-main.525/), 2024 | Joint optimization of instructions and demonstrations in multi-stage programs | Useful native-supervision context; not necessarily the closest strict-loop baseline | | [SPO](https://aclanthology.org/2025.findings-emnlp.479/), 2025 | Reference-free prompt optimization using pairwise output judgments | Removing gold answers is insufficient novelty; test a faithful reference-free competitor | | [ProRefine](https://arxiv.org/abs/2506.05305), 2025 | Inference-time prompt refinement using textual feedback without gold labels | A particularly close baseline for the primary text track; read the actual prompts and selection procedure | | [ACE](https://arxiv.org/abs/2510.04618), ICLR 2026 | Context evolves through generation, reflection, and curation; supports unlabeled adaptation with execution feedback | Compare memory update and evidence access, rather than claiming evolving memory is new | | [Feedback Descent](https://arxiv.org/abs/2511.07919), 2025 | Text-artifact improvement through comparative preferences and rationales, with no weight updates | Artifact critique itself is prior art; isolate the effect of removing comparative acceptance and using next-action planning | | [Prompt Duel Optimizer](https://aclanthology.org/2026.findings-acl.490/), 2026 | Label-free prompt selection under a limited LLM comparison budget | Report comparator cost and judge dependence; include if its task setup is a good fit | | [Constitutional AI](https://arxiv.org/abs/2212.08073), 2022 | Principles guide critique and revision within a broader training procedure | Principles are not inherently novel; show their contribution to fixed-agent prompting | | [Topic-disjoint reflective optimization study](https://arxiv.org/abs/2610.12601), October 2026 preprint | Studies transfer, repeated runs, feedback detail, and prompt-level overfitting for a deep-research agent | Use disjoint topics, independent judging, and verifier ablations; treat its claims as reported preprint findings | Do not infer that any listed work prohibits a behavior merely because its abstract does not mention it. The novelty audit must inspect full methods, released prompts, and available code. ## Priorities before the pilot 1. Read and diagram the full methods of Self-Refine, ProRefine, SPO, ACE, and Feedback Descent. For each, record exact input channels, mutable state, candidate generator, selection rule, memory, and stopping. 2. Pin official implementations or document faithful reimplementations. Verify which mechanisms are essential and which adaptations are required for persistent tasks. 3. Use equal development budgets to choose configurations. Do not hand-tune C extensively while keeping competitors at an arbitrary default. 4. Repeat a small sanity task for each baseline. Confirm that its implementation can actually improve a known defective output before using it as evidence against the method. 5. Decide whether C is a distinct algorithm or an especially constrained instance of existing self-refinement. Update the paper framing before preregistration. ## Claim ledger | Candidate claim | Status now | Evidence needed | |---|---|---| | Cruise can generate follow-up prompts from reports and saved guidance | Supported by inspected implementation and historical records | Existing source/export references | | Current production Cruise is artifact-only | Not supported; current planner receives broader context | A separately enforced restricted-input variant | | The proposed method changes no code in its primary text study | Design requirement, not experimentally verified | Configuration and write-access audit during runs | | Artifact access causes better next prompts | Hypothesis | Controlled artifact interventions and downstream quality | | The principles preserve user intent | Hypothesis | Principle/provenance ablations and drift measurements | | The method beats self-refinement or GEPA | Unknown | Faithful paired comparisons with uncertainty and full cost | | The method transfers to unseen workers/tasks | Unknown | Separate frozen-protocol transfer studies | | The study requires no evaluation | Incorrect framing | Independent assessment is essential; external scores are excluded only from adaptation |