# Cruise: Artifact-Guided Prompting for Fixed Agents Without External Evaluators in the Loop *Initial manuscript draft: Introduction, Method, Experimental Design, Discussion, and Conclusion. The method and study are proposed and untested. No experiments have been conducted for this proposal, and no improvement, novelty, or superiority claim has been established.* ## Abstract We propose to study whether a planner can improve a fixed agent's text deliverables through artifact-guided follow-up prompts without external evaluators during adaptation. The planner observes delivered artifacts, explicit user requirements, and authorized source material, while execution traces, worker self-reports, hidden answers, and external scoring or acceptance signals remain unavailable. Fixed principles guide next-step selection, and bounded, provenance-linked memory records observations and provisional tactics. The worker, planner configuration, and harness remain unchanged; adaptation modifies only task-specific prompts, memory, and text artifacts, with no code changes. Proposed comparisons include generic continuation, ordinary self-refinement, reference-free prompt optimizers, and a separate GEPA condition with its supervision and development costs disclosed. Matched resource budgets, artifact-access interventions, and memory ablations would test whether the controller contributes beyond additional computation. Independent assessment would occur only after outputs are sealed. The principal limitation is that artifact critique can miss errors and reinforce unsupported assumptions; qualitative agreement cannot certify correctness. The study is untested and reports no empirical results. Its proposed contribution is a testable characterization of restricted-feedback adaptation, without claiming established novelty, superiority, or transfer across tasks. ## 1. Introduction Completing a substantial task with a language-model agent can require a sequence of instructions rather than a single request. A user may be able to specify the desired document while finding it difficult to determine which revision should follow each intermediate draft. The next instruction must account for work already delivered, identify a remaining gap, and preserve requirements that earlier revisions have satisfied. This motivates a question about the control of an otherwise fixed agent: can a planner use the delivered artifact to choose subsequent instructions that advance the original task, without access to the worker's internal execution or an external evaluator during adaptation? Prior work provides several approaches to improving language-model behavior through language. Self-Refine alternates feedback and revision of generated outputs without additional model training ([Madaan et al., 2023](https://arxiv.org/abs/2303.17651)). Reflexion retains linguistic reflections in episodic memory to inform later attempts ([Shinn et al., 2023](https://arxiv.org/abs/2303.11366)). These approaches establish that output revision and textual memory are existing mechanisms, rather than contributions that can be claimed merely by introducing another prompting loop. The relationship between a proposed controller and these mechanisms therefore requires explicit comparison. GEPA combines reflective prompt mutation with evolutionary search and Pareto-based candidate selection. Its principal formulation uses task instances, execution traces, textual feedback, and numeric performance signals while keeping model weights fixed ([Agrawal et al., 2026](https://proceedings.iclr.cc/paper_files/paper/2026/hash/0e9e708b6f48e14fd0ac29e167413f76-Abstract-Conference.html)). GEPA also considers inference-time search, so neither fixed weights nor adaptation during inference is sufficient to distinguish the present proposal. Other work further narrows the relevant question. Self-Supervised Prompt Optimization derives prompt-selection signals from pairwise output comparisons without external references ([Xiang et al., 2025](https://aclanthology.org/2025.findings-emnlp.479/)), and ProRefine studies inference-time prompt refinement with textual feedback without gold labels ([Pandita et al., 2025](https://arxiv.org/abs/2506.05305)). ACE treats contexts as evolving playbooks maintained through generation, reflection, and curation ([Zhang et al., 2025](https://arxiv.org/abs/2510.04618)), while Feedback Descent uses comparative preferences and textual rationales to guide changes to text artifacts ([Lee et al., 2025](https://arxiv.org/abs/2511.07919)). Removing gold answers, maintaining textual state, or inspecting outputs does not by itself establish a new method. We propose to investigate a restricted observation and control setting. A separate planner receives the user's explicit requirements, a fixed packet of authorized source material, and the text artifacts delivered by an unchanged worker. It may generate the next task instruction and revise bounded textual memory, but it cannot inspect execution traces, consult hidden references, query a task scorer, or use a pairwise acceptance oracle. Fixed principles direct the planner to ground proposed work in observable gaps, distinguish evidence from hypotheses, and preserve the user's requirements. The proposed unit of adaptation is a follow-up instruction for the current evolving task, rather than a model-weight update or a scored choice among competing prompt configurations. This setting overlaps with self-refinement because artifact critique informs subsequent revision. The proposed emphasis is on selecting a concrete next increment under an explicit information boundary and maintaining attributable memory across the task. Relative to GEPA's principal formulation, the proposed loop omits execution-level evidence and numeric candidate selection. These are design distinctions to investigate, not evidence of established novelty or an advantage over either approach. In particular, the planner still exercises qualitative judgment. Excluding an external evaluator does not make the process free of evaluation in the ordinary sense, and it provides no guarantee that successive artifacts improve. The primary study is limited to text deliverables, such as source-grounded briefs, operational documents, and structured plans. Neither the worker nor the planner changes code during adaptation. Independent assessment would take place only after run outputs have been sealed, so that measured quality cannot guide the run being assessed. Comparisons with generic continuation, ordinary self-refinement, and relevant prompt optimizers would be needed to determine whether artifact access, principles, and bounded memory contribute beyond additional computation. This draft specifies that untested procedure and its intended limits; it reports no empirical results. Software-artifact experiments remain an optional extension outside the primary formulation. ### 1.1. Related work: observation, feedback, and retained state The relevant comparison separates what a method observes, which signals guide adaptation, what it changes, and how it retains or selects prior attempts. Table 1 summarizes the mechanisms established by the existing source review. GEPA's entry draws on the conference paper's formulation and algorithms (Sections 2–3 and Appendix D); the other entries use the cited primary abstracts and, for Feedback Descent, its targeted method description. This is a bounded comparison rather than an implementation audit of every variant. **Unspecified** means that the consulted source material does not establish the detail, not that the original method lacks it. Observation entries describe reported access and do not imply an exhaustive permission boundary unless one is explicitly defined. **Table 1. Mechanism comparison.** Memory includes retained adaptation state, which need not be a separate memory module. The proposed row describes the untested text-artifact variant in Section 2, not production Cruise. | Method and source | Observation access | Feedback signal | Adaptation target | Memory / retained state | Candidate selection | |---|---|---|---|---|---| | [GEPA](https://proceedings.iclr.cc/paper_files/paper/2026/hash/0e9e708b6f48e14fd0ac29e167413f76-Abstract-Conference.html) | Task inputs, module execution traces, evaluation diagnostics | Numeric metrics and textual feedback | Module prompts; weights fixed | Candidate prompts, ancestry, score matrix | Minibatch improvement gate; Pareto-based sampling; final validation selection | | [Self-Refine](https://arxiv.org/abs/2303.17651) | Task and generated output | Feedback produced by the same LLM | Current output | Retained-history policy unspecified | Iterative revision; acceptance rule unspecified | | [Reflexion](https://arxiv.org/abs/2303.11366) | Task interactions and feedback; exact access limits unspecified | Scalar or linguistic; external or internally simulated | Later decisions through reflective context; weights unchanged | Reflective text in episodic memory | Trial-selection rule unspecified | | [SPO](https://aclanthology.org/2025.findings-emnlp.479/) | Task requirements and compared outputs; external references unnecessary | LLM pairwise output judgments | Prompts | Memory policy unspecified | Prompt selection through judged output comparisons | | [ProRefine](https://arxiv.org/abs/2506.05305) | Reasoning-task prompts; detailed observation access unspecified | LLM textual feedback without gold labels | Prompts at inference time | Memory policy unspecified | Candidate acceptance and ranking unspecified | | [ACE](https://arxiv.org/abs/2510.04618) | Contexts and natural execution feedback; exact access limits unspecified | Reflection informed by execution feedback; unlabeled adaptation supported | Contexts, including prompts and agent memory | Evolving playbook with structured incremental updates | Curation; candidate ranking or acceptance rule unspecified | | [Feedback Descent](https://arxiv.org/html/2511.07919v2#S2) | Candidate and incumbent artifacts; representations depend on domain | Pairwise preference and textual rationale | Text artifacts, including prompts | Feedback history and incumbent artifact | Preferred candidate replaces incumbent | | Proposed method, untested | Delivered text, public requirements and sources, bounded history; no traces or worker reports | Planner's own qualitative gap diagnosis; no supplied grade or acceptance signal | Follow-up instructions and bounded memory for the current task | Attributable observations and provisional tactics; immutable goal separate | One continuing trajectory; no scored or pairwise acceptance gate | Several mechanisms overlap directly. Self-Refine already makes output critique an input to revision, so inserting a planner between worker turns does not by itself establish an additional capability. Reflexion establishes reflective episodic state as a means of influencing later decisions, while ACE develops an evolving context playbook; the proposed bounded memory therefore belongs to an existing family of context-adaptation mechanisms. GEPA also accumulates information through evolved prompts and their ancestry, even though that state differs from an explicit episodic buffer. These distinctions concern the organization and use of retained information, not the presence or absence of learning in language. Where the consulted records leave history retention or selection details unspecified, they cannot support a stronger claim that another method lacks the proposed behavior. Feedback content must also be distinguished from feedback provenance and selection. SPO removes the need for an external reference answer while retaining an LLM evaluator's pairwise judgments. Feedback Descent couples comparative preferences to rationales and uses preference to retain an incumbent ([Lee et al., 2025, Section 2](https://arxiv.org/html/2511.07919v2#S2)). ProRefine's label-free textual prompt refinement likewise makes the absence of gold answers insufficient as a differentiating claim. GEPA's main formulation uses both rich diagnostic feedback and scores, and its inference-time extension prevents a simple offline-versus-online distinction. The table thus supports comparing specific information channels and update rules rather than treating all these procedures as equivalent or ordering them by a single label such as “self-improvement.” In this proposal, unavailable external scores are separately supplied grading or acceptance signals to the controller. The restriction does not exclude the planner's own qualitative evaluation of a visible gap against public requirements. A pairwise judge remains a selection mechanism even if it uses the same model family as the planner or runs inside the same application; a scalar reward is not necessary for such selection. Conversely, a source citation or structural memory check does not constitute independent evidence that an artifact improved. The distinction is therefore functional rather than a consequence of whether models occupy separate processes. The remaining question is whether a controller can use delivered text and bounded attributable state to choose useful follow-ups when diagnostic execution access and separate grading or acceptance signals are withheld. Comparisons and interventions in Section 3 are intended to determine whether that restriction reveals a useful quality–resource tradeoff and whether current artifacts actually influence outcomes beyond ordinary self-refinement. It is possible that the proposed procedure is adequately explained as an instance of existing refinement methods, or that richer feedback is necessary for reliable progress. The restricted-information question remains testable in either case; the present untested proposal establishes neither novelty nor superiority, and unspecified source details cannot be used as evidence for either. ## 2. Method ### 2.1. Task setting and fixed components We consider a task specified by an immutable user goal and explicit requirements, denoted by \(G\), together with authorized source material \(D\). The source packet is fixed before a run and contains information the worker is permitted to use in producing the requested text. It is distinct from assessment-only material, such as reference answers or scoring rubrics. Tasks must be answerable from the public task specification and supplied evidence; an independent assessment must not introduce requirements unavailable to the system. The proposed system contains a worker \(W\), a planner \(Q\), and a harness \(H\) that delivers instructions, exports artifacts, and records state. The worker's model weights, system instructions, decoding configuration, available capabilities, permissions, and context-retention policy remain fixed throughout a run. The planner's model, meta-prompt, and principles also remain fixed. The harness implementation, observation procedure, memory filter, and resource-accounting rules cannot be rewritten by either model. The primary procedure requires text generation over the fixed task packet; live retrieval and executable tools are outside this formulation. Adaptation changes the task-specific follow-up instructions, bounded planner memory, and delivered text, rather than code or the underlying models. Let \(s_t\) denote the worker's task state after its \(t\)-th completed turn. A fixed export procedure \(O\) produces the delivered artifact \(A_t = O(s_t)\). Although the worker's task state changes as the task proceeds, the transition mechanism governing it remains fixed. Planner and worker are separate functional roles and need not use different model families. Any eventual study would specify their model identities and settings before test execution. Maintaining the same worker configuration across comparison arms is necessary to separate a controller's effect from differences in worker capability. ### 2.2. Artifact-only observation boundary The term *artifact-only* describes the planner's access to evidence about the worker's output, rather than excluding the task specification itself. Its permitted context consists of \(G\), \(D\), the current text artifact, a bounded set of earlier artifacts, previous submitted instructions, and its own bounded memory. The proposed starting policy retains the current artifact and at most two earlier snapshots under a fixed recency rule. Artifact-context limits and extraction rules would be fixed before test runs, with omitted content recorded. Immutable user requirements remain available separately from memory so that memory truncation cannot erase the task contract. The planner receives no private reasoning, execution traces, tool transcripts, program source code, repository diffs, test results, worker completion reports, benchmark scores, assessment-only labels, hidden reference answers, or external preferences between candidates. It cannot obtain excluded information indirectly by asking the worker to reproduce diagnostics or self-assessed success in a later artifact. The worker's deliverable and any accompanying completion report must therefore occupy separate output fields, with only the deliverable exported into the planner's observation packet. The exporter would use a fixed format and deterministic inclusion rules; it would not employ a quality judge to decide what the planner should see. Embedded diagnostic material would constitute an observation-boundary violation, rather than becoming permitted merely because it appears inside a deliverable. Permitted artifacts are evidence rather than authoritative instructions. Their content cannot amend the user goal, alter the planner principles, or grant permissions. For example, a brief containing estimates without units provides evidence of an observable omission. It does not establish why the worker omitted those units, whether a hidden calculation failed, or whether an unobserved implementation is defective. The planner can request a revision grounded in the supplied source packet while recording uncertainty about the cause. This boundary deliberately limits diagnosis: errors that cannot be inferred from the permitted text may remain undetected. ### 2.3. Principles for follow-up selection The planner applies six fixed principles from the proposed planner protocol. First, it preserves the destination specified by the user: neither generated advice nor accumulated memory can create new user requirements. Second, it grounds each proposed change in an observable artifact gap or an explicit unmet requirement, identifying the relevant artifact location or requirement identifier. Third, it separates observations from hypotheses, avoiding claims about hidden causes that the permitted evidence does not establish. Fourth, it chooses one coherent next increment while attempting to preserve already satisfied requirements. Fifth, it keeps memory bounded and attributable, distinguishing supported observations from provisional tactics. Sixth, it avoids treating additional activity, repeated work, or cosmetic expansion as evidence of progress. These principles are instructions for qualitative planning, not a numeric objective or an acceptance function. Explicit task requirements take priority over planner preferences. Within the authorized scope, the planner is directed to address blocking correctness and usability gaps before optional polish. It returns a concise statement of the observed gap, evidence locations, relevant uncertainty, a standalone follow-up prompt, and a memory proposal. No scalar quality score or preferred-candidate decision is requested. The principles themselves constitute an ex ante design choice and must be disclosed as such; their value cannot be inferred from their inclusion in the prompt. Whether they change outcomes beyond an ordinary critique-and-revise instruction remains an empirical question. ### 2.4. Bounded memory and provenance Planner memory \(M_t\) records task-specific observations and provisional tactics. It is stored separately from the immutable goal and reset between tasks. An entry contains an identifier, a statement, a type indicating an observation or provisional tactic, evidence references, and a status such as unresolved or superseded. Observations refer to an exported artifact or explicit task requirement. A tactic may be motivated by such evidence, but remains provisional rather than acquiring the status of a user instruction through repetition. Superseded entries can be removed from live context while remaining in the audit record. The protocol proposes initial limits of 1,200 tokens for live memory and 800 tokens for a follow-up prompt. These are unvalidated starting settings to be finalized on development tasks, rather than tested or optimal values. The memory budget excludes the separately retained user requirements. A fixed filter \(F\) checks entry structure, size, permitted types, valid evidence identifiers, and separation from the immutable requirement store. Its role is structural validation, not semantic assessment: the existence of an evidence reference does not certify that the referenced text supports the statement. Such support would be audited independently after a run. Provenance-linked memory is therefore intended to make drift inspectable, not to guarantee that unsupported assumptions cannot arise. This distinction matters when a temporary choice is repeatedly restated. A planner preference for a particular document structure, for example, must not become a permanent requirement unless the user specified it. The filter prevents generated memory from being written into the requirement store, while the planner principles discourage treating its own tactics as obligations. These safeguards leave a residual failure mode in which a model follows an unsupported memory entry despite its provisional label. Measuring that failure would be part of the proposed study. ### 2.5. Follow-up loop and termination The initial worker instruction presents the task and its authorized source packet. After the first completed turn, the harness exports \(A_1\) and begins the follow-up cycle with empty planner memory \(M_1\). Let \(\bar{A}_t\) denote the permitted artifact snapshots and \(U_t\) the previous instructions retained under a fixed context policy. The proposed transition is \[ (u_t, \widetilde{M}_{t+1}, z_t) \sim Q(G, D, P, \bar{A}_t, M_t, U_t), \] \[ M_{t+1}=F(\widetilde{M}_{t+1}), \qquad s_{t+1}\sim W(s_t,u_t), \qquad A_{t+1}=O(s_{t+1}), \] where \(P\) denotes the fixed principles, \(u_t\) is the follow-up prompt, \(\widetilde{M}_{t+1}\) is proposed memory, and \(z_t\) records the observed gap, evidence, and uncertainty. The notation allows stochastic generation without any update to the fixed planner or worker. The harness records these outputs for auditing and delivers the validated follow-up through the unchanged worker interface. Planner memory informs the worker through the follow-up text; it does not modify the worker's system instructions. Each cycle produces one follow-up for a continuing task trajectory. There is no pool of candidates ranked by an external metric and no quality-based gate that accepts or rejects a worker revision. Structural output validation is distinct from such a gate. Invalid outputs and service failures would be handled by a prespecified, non-scoring retry policy, with every attempt and consumed resource retained in the record. Failed requests cannot provide free additional optimization attempts. The starting primary schedule is ten completed worker turns, with at most nine between-turn planning steps, subject to a total resource cap fixed before test execution. Planner and worker costs both count toward that cap. If the planner identifies no supported remaining gap, the protocol requests presentation of the completed artifact and any unresolved uncertainty rather than inventing new work. Such presentation turns remain part of the fixed schedule and resource accounting. The planner's confidence does not determine early stopping. The reported endpoint is the last delivered artifact before the scheduled endpoint or resource limit, not a historical artifact selected retrospectively for its measured quality. Neither this stopping rule nor the absence of candidate selection guarantees monotonic improvement. ### 2.6. Adaptation and independent assessment Excluding external evaluators during adaptation is a restriction on information flow. During a run, no separate grader, reference-based metric, human quality judgment, or pairwise acceptance oracle supplies a signal that can change the next prompt, memory, stopping decision, retry, or retained artifact. The planner nevertheless compares visible work with public requirements and makes qualitative judgments about what to request next. The proposed method is therefore not evaluation-free in an unrestricted sense. It is an artifact-guided adaptation procedure without an external performance or acceptance signal in its loop. Scientific assessment serves a different role. After all outputs in a prespecified evaluation batch have been sealed, an independent assessor may measure the terminal artifact against a rubric or reference material unavailable during adaptation. Writing \(t^*\) for the actual terminal turn and \(R\) for assessment-only material, this produces \(J(A_{t^*};G,D,R)\). The result is an observation about the completed run, not an input to its transition rule. Because neither \(R\) nor the value of \(J\) can influence that run, post-run assessment is compatible with excluding external evaluators during adaptation. Inspecting intermediate artifacts after sealing can characterize regressions, but cannot be used to replace the prespecified endpoint with a higher-scoring snapshot. The same separation must apply across test runs. Once test assessment is revealed, using it to revise principles, memory rules, configurations, or baseline settings would make that assessment part of subsequent development. Such changes would require a new held-out evaluation and could not be described as belonging to the original frozen test procedure. Development data may inform the design before freezing; final test judgments may not silently become adaptation feedback. ### 2.7. Scope and empirical status This formulation concerns within-task adaptation of text artifacts. It does not demonstrate learning a reusable strategy across tasks, improved model capability, lower cost, or greater reliability. Artifact inspection can miss factual errors, unsupported memory entries can persist, and an apparently useful revision can break an earlier requirement. Comparisons under matched resource budgets and controlled changes to artifact access would be needed to establish whether the proposed controller contributes beyond self-refinement or additional worker computation. No such evidence is reported here. The procedure also describes a proposed research variant rather than a verified characterization of the broader production Cruise planner. Software experiments, if pursued, would require a separate observation protocol and separately stated claims; they are not part of the primary text-only, no-code-change method. ## 3. Experimental Design ### 3.1. Questions, task families, and data separation The proposed study asks whether artifact-guided follow-up selection improves the terminal deliverable relative to generic continuation when the worker and total adaptation budget are held constant. A secondary question is whether any advantage survives comparison with ordinary self-refinement and reference-free prompt optimization. These comparisons concern the quality of work completed within a task. They do not establish that the worker has acquired a reusable capability or that the planner has learned a policy that transfers to other tasks. All design choices in this section are prospective defaults grounded in the research protocol; none describes an executed experiment or a completed preregistration. The primary suite would contain three equally represented families of text-artifact tasks. Source-grounded synthesis tasks would ask for a decision brief from a fixed packet of evidence and stakeholder notes, with attention to claim support, attribution, coverage, and uncertainty. Constraint-heavy document tasks would require an operational specification that coordinates roles, exceptions, dependencies, and explicit scope boundaries. Structured planning and analysis tasks would require a feasible plan from supplied calendars, inventories, numerical facts, and preferences, making arithmetic and consistency between prose and tables assessable. Each task would admit multiple substantive revisions while remaining feasible within the planned budget. No task would require producing or modifying code, retrieving live information, or changing the worker or harness. We propose 12 development tasks for feasibility work, 12 separate validation tasks for configuration selection, and 60 held-out test tasks, with four development, four validation, and twenty test tasks in each family. Topics, source packets, and scenario templates would be disjoint across splits. A manual overlap audit would examine substantive similarities rather than treating renamed entities as adequate separation. Tasks and rubrics would be authored before test outputs are available, and the historical Cruise demonstrations would be excluded from the test set. Assessment-only checks would operationalize public requirements rather than introduce undisclosed goals. The proposed test size is a starting allocation, not a claim of sufficient statistical power; development-run variance would inform any revision before the test protocol is frozen. ### 3.2. Comparison conditions and supervision The core controller comparison would include the proposed method, generic continuation, and an ordinary Self-Refine-style critique-and-revision condition. Each would receive the same initial task instruction and authorized source packet, use the same worker configuration, and follow the scheduled revision opportunities described in Section 2.5. Generic continuation would repeatedly ask the worker to review and improve its deliverable toward the original goal. The self-refinement condition would generate feedback on the current artifact and use that feedback for revision, without the proposed six-principle policy or provenance-linked memory. It would be described as an adaptation of Self-Refine to the common interface, rather than as a reproduction of every setting in that paper. The purpose is to distinguish explicit next-action planning from the effects of repeated generation and ordinary critique. SPO and Feedback Descent would provide stronger reference-free comparisons under the same task and artifact access. Their published proposal and selection mechanisms would be retained where compatible with text deliverables, and necessary interface adaptations would be documented. In particular, their own pairwise judgments would remain part of their methods and would consume adaptation resources. These judgments would not come from the independent final assessor. This panel matches available task evidence while allowing selection mechanisms to differ; it is not a comparison in which every method is prohibited from generating comparative judgments. A single-prompt worker with a larger generation allowance would provide an additional compute-allocation control. Its purpose would be to test whether spending resources on an outer planner is preferable to allowing the unchanged worker more computation directly. We propose equal configuration-search allowances across methods: at most four prespecified configurations per condition, selected by the same family-balanced validation criterion, with tuning expenditure reported separately. Baseline prompts, settings, and interface adaptations would be frozen before test execution. A small development sanity check would verify that each implementation can operate on the task format, without excluding a baseline because its validation performance is inconvenient. ACE or ProRefine could be added after implementation-level review, but a change to the registered comparison set would occur before inspecting test outcomes. A GEPA comparison would be reported separately because its native supervision and optimization target differ from the core controller comparison. The proposed default is to optimize task-family instructions on development tasks using GEPA's metric, feedback, and candidate-selection procedure, select a configuration on validation tasks, and freeze the resulting instructions before applying them to test tasks. The underlying worker model and capabilities would be shared, while the learned instruction would be an explicit condition difference. We would report development cost, test cost, and amortized cost at declared deployment volumes rather than treating a trained family instruction as having no acquisition cost. Removing GEPA's score gate would create a labeled ablation, not a native GEPA baseline. A separate inference-time GEPA condition could use a disclosed public feedback function, but would have no access to assessment-only test material. Neither native-supervision condition would be interpreted as isolating algorithm quality under identical feedback. ### 3.3. Resource matching and repeated runs The primary comparison would use a common billed adaptation-cost cap, with ten worker turns as the maximum schedule for the continuing-task conditions. Native search baselines would also be compared at the common cost cap using their documented call structure; their internal candidate generation would not be treated as free work or disguised as a single worker turn. The study would report completed worker calls and revision opportunities alongside quality so that differences in call structure remain visible. The same worker snapshot, decoding settings, and context policy would be used across compatible conditions, and planner or critic roles would use the same declared model snapshot wherever their methods permit it. Rather than inventing a dollar cap without usage measurements, we propose a reproducible development-only rule: set the medium cap to 1.25 times the empirical ninetieth percentile of total adaptation cost for completed ten-turn runs of the proposed method in the development pilot. If no such run completes, the design would remain infeasible and return to development rather than assigning a cap from test data. Low and high caps would be one-half and twice this value. These multipliers are proposed engineering defaults, not empirically justified optima. They would be frozen before the held-out study and applied equally across conditions. Separate runs would be used at each cap so that intermediate assessment cannot determine whether a low-budget run should continue. The medium cap would define the primary contrast; the other caps would characterize quality against expenditure. Accounting would include every worker generation, planner call, critique, comparison, discarded candidate, and retry, including repeated input context and cache charges. Input, output, and exposed reasoning-token counts, request counts, actual billed cost, and elapsed time would be retained separately. Independent final-assessment costs would be disclosed as study costs but would not be charged to a method's adaptation allowance. Any unobserved usage component would be identified rather than estimated as a measured saving. A cost-based primary claim would require usable billing records; absent those records, a switch to a token-budget study would have to be declared before test execution. The last deliverable obtained within the allowed budget would remain the endpoint under Section 2.5. We propose three independent whole-run repetitions per task and condition, with fresh task state and no memory carried between runs. Execution order would be randomized within task and time blocks to reduce confounding with service variation. Seeds would be recorded where supported; lack of deterministic seed control would be disclosed. Repetitions would characterize generation variability, while task-level replication would support inference across the task suite. Three runs of one task would not count as three independent tasks. ### 3.4. Ablations and controlled artifact interventions The mechanism study would remove the fixed principles, remove live memory, and replace structured memory with an unstructured note of the same token allowance. The first contrast would test whether the decision rules contribute beyond the rest of the planner interface; the latter two would separate retaining task state from the particular provenance scheme. Artifact and history access would otherwise be held fixed. If removing an input reduces actual expenditure, both the configured allowance and consumed resources would be reported. Neutral formatting text could control prompt length in the principles ablation, but would not be interpreted as proving that every difference in semantic content had been isolated. To test dependence on current evidence, another condition would withhold the current artifact while retaining only permitted earlier context, and a stale-observation condition would substitute the previous artifact from the same task. Memory in these runs would be generated under that restricted condition, not copied from a planner that had already seen the current artifact. These interventions concern the planner's access; the worker would continue to possess its ordinary task state. Equivalent outcomes would therefore suggest that current-artifact access adds little to the controller in this setting, rather than showing that the worker can operate without its own output. Controlled one-step probes would complement full-run ablations. An independent task author would introduce a single visible omission or inconsistency into a text deliverable before a paired probe. Both worker branches would start from that same edited deliverable, while one planner would receive the accurate version and the other a matched stale or masked observation. All other task context and resource limits would be held constant. After sealing the branches, assessment would determine whether the follow-up identified the manipulated gap and whether the worker repaired it without breaking other requirements. Merely producing different prompts would not establish a useful effect. The introduced defects and their distribution would be fixed before evaluation, and probe outcomes would be reported separately from unmodified-task performance. An expanded-access condition could supply worker self-reports, including a separately labeled misleading-report stress test, to examine reliance on claims of completion. Such inputs would not be silently added to the artifact-only condition. Likewise, adding an external score or comparative acceptance signal would define an expanded-supervision ablation with its cost explicitly counted. These comparisons would help identify what additional information contributes without weakening the strict condition's definition. ### 3.5. Independent outcomes and rating procedure The primary outcome would be the proportion of prespecified atomic requirements satisfied by the terminal text artifact, normalized to the interval from zero to one. As a proposed scoring default, atomic items would receive equal weight within each task, tasks equal weight within a family, and the three families equal weight in the overall estimate. Each item would be scored as satisfied or unsatisfied under a rubric written before outputs are inspected. Criteria involving arithmetic or explicit format constraints could use deterministic checks, while semantic criteria would require independent judgment. This choice makes the score interpretable but does not imply that all errors have equal practical severity; critical violations would therefore be reported separately. Following the separation in Section 2.6, assessment would begin only after every run in the frozen test batch, including registered ablations, had been sealed. For semantic criteria, we propose two independent human raters for all terminal artifacts and a third adjudicator for disagreements. Raters would receive the public task packet and assessment rubric but not method names, planner messages, usage, or execution order. A separate process audit would inspect memory and prompts after artifact-quality judgments were recorded. Agreement and adjudication rates would be reported. Optional LLM judgments would be supplementary, with frozen prompts and human calibration, rather than the default primary semantic assessor. Secondary outcomes would include blinded usefulness preferences, unsupported claims with explicit denominators, critical requirement violations, noncompletion, and final-state regressions. Post-run scoring of the initial and terminal artifacts would describe within-run change, interpreted alongside controller comparisons because additional worker turns alone could improve a draft. A regression would mean that a requirement satisfied in an earlier sealed snapshot is unsatisfied at the terminal endpoint. Memory analysis would count unsupported normative statements, report their fraction among live entries, and identify the proportion of runs affected. Repeated-work and progress coding would also occur after sealing. These outcomes would remain separate rather than being combined retrospectively into a favorable composite. Intermediate quality measurements could describe the trajectory, but would not change the terminal artifact selected for the primary analysis. ### 3.6. Statistical analysis and disconfirming outcomes The primary estimand would be the family-balanced mean task-level difference between the proposed method and generic continuation at the medium resource cap. For each task, scores would first be averaged across its three repetitions. We propose a family-stratified bootstrap over tasks, retaining all repetitions and condition pairings within each sampled task, to obtain a 95% confidence interval for the mean difference. Ten thousand bootstrap resamples is a proposed computational default. Secondary comparisons with self-refinement, SPO, and Feedback Descent would report effect sizes and confidence intervals, with Holm correction across those three comparisons. The proposed significance-test default is a two-sided paired permutation test with 10,000 within-task condition-label swaps, using the family-balanced mean difference and assuming exchangeability of the paired condition labels under the null. Individual rater decisions or repeated generations would not be treated as independent task draws. For planning purposes, we propose five percentage points on the normalized requirement-satisfaction scale as the smallest practically important mean difference. This is a provisional decision threshold rather than an expected effect; it would be reviewed using development variance and task-rubric granularity before the test protocol is frozen. A confidence interval containing zero would leave the direction of the primary effect unresolved. An interval whose upper bound falls below the practical threshold would count against the proposed magnitude of benefit, even if a smaller positive difference remains plausible. Statistical nonsignificance alone would not establish equivalence, and this initial design would make no noninferiority claim without a separately justified margin. Several outcomes would undermine stronger interpretations of the hypothesis. A gain over generic continuation that disappears against self-refinement would fail to establish value beyond ordinary critique and revision. A gain that disappears when planner and comparison costs are included would weaken the resource-efficiency argument. Little change under current-artifact removal or stale observations would weaken the claim that the controller benefits from inspecting the actual delivered work. No reduction in unsupported requirements under provenance-linked memory, or a corresponding loss of task completion, would count against that memory design. Improvements restricted to presentation preferences while factual support or critical requirement compliance worsens would not support a general quality claim. Effects confined to one task family would require a family-specific conclusion. These findings would motivate a narrower account of the procedure rather than being excluded from the analysis. ### 3.7. Failures, incomplete runs, and scope of transfer The failure analysis would distinguish service failures, resource exhaustion, malformed planner output, repeated or irrelevant work, factual errors, regressions, unsupported memory entries, and violations of the observation boundary. As a proposed operational default, transient transport failures could receive one identical-request retry within the existing resource cap; substantive low quality would not justify a rerun. Structurally invalid planner output could receive one schema-repair request that supplies only the formatting error and original permitted context, also charged to the cap. A second failure would terminate the run and preserve its last delivered artifact. Runs ending early would be scored on that artifact and marked incomplete; a run producing no artifact would receive the minimum primary score of zero. Unused budget in a failed run would not be treated as evidence of efficiency. All assigned runs would remain in the main operational analysis, including runs with information leakage, with noncompliance reported explicitly. A compliant-only analysis could be supplementary, but could not replace inconvenient runs or restore a strict-information claim invalidated by systematic leakage. Human interventions would be logged and assisted runs identified separately. Stress cases would include polished but numerically incorrect text, conflicting public constraints, and errors that the allowed observation does not expose sufficiently for reliable diagnosis. Longer twenty- and thirty-turn studies could examine drift, but would remain secondary and could not replace the ten-turn primary endpoint. Audit logs, frozen configurations, allowed input packets, and unsuccessful trajectories would be retained alongside favorable outputs. Finally, the primary analysis concerns within-task progress from an initial deliverable to a terminal artifact, interpreted through the controlled comparisons above. Fresh test tasks with memory reset can show that a frozen procedure works on tasks not used to configure it; they do not show that knowledge learned on one task transfers to another. A separate optional transfer study would derive a reusable instruction from development-task experience, remove task-specific content, freeze that instruction, and compare it with an equally budgeted undeveloped instruction on disjoint tasks without further prompt adaptation. Any content filtering or instruction selection would occur before transfer-test outcomes are available. A second-worker replication could test portability of the frozen procedure, but would likewise be distinguished from learned cross-task transfer. Neither extension would justify a transfer claim until it had been executed and independently assessed. Software experiments remain outside the proposed primary study. ## 4. Discussion The potential contribution of this proposal is to clarify what can be accomplished under a deliberately restricted observation interface. Relative to GEPA's use of execution-level evidence and scored selection, the restriction could help identify when those signals are necessary and when delivered text provides sufficient guidance for useful revision ([Agrawal et al., 2026](https://proceedings.iclr.cc/paper_files/paper/2026/hash/0e9e708b6f48e14fd0ac29e167413f76-Abstract-Conference.html)). Relative to self-refinement, it could examine whether choosing a concrete next increment and preserving attributable task state adds value beyond ordinary feedback and revision ([Madaan et al., 2023](https://arxiv.org/abs/2303.17651)). A useful outcome could therefore be a characterization of the limits of artifact-guided control, including conditions in which richer supervision remains preferable. Such a finding would not require showing that the restricted procedure is universally better. An information restriction is a research constraint, not evidence of an original algorithm. The proposed loop remains closely related to existing output-feedback and prompt-optimization methods. Its separation of planner and worker, fixed principles, and memory conventions may amount to a particular configuration of those ideas. Establishing a distinct contribution would require both a fuller comparison of mechanisms and empirical evidence that the restriction exposes an informative difference. Conversely, a failure to improve on self-refinement would not make the observation setting meaningless; it would limit the case for the additional controller structure. The relevant question is what the constraint allows the study to explain, rather than whether a different interface alone merits a novelty claim. Artifact-only critique also places a limit on what repeated revision can establish. A fluent document can omit an important qualification, present an incorrect number consistently, or make an unsupported claim appear plausible through confident wording. When the permitted evidence does not distinguish between competing explanations, further critique cannot by itself validate the missing evidence. Even a detectable error may be overlooked if planner and worker share assumptions about the task. Agreement between them could then reflect a common blind spot rather than independent confirmation. The ability to identify visible deficiencies should therefore be distinguished from the ability to certify correctness. Iteration can amplify these limitations. A tentative interpretation may enter memory, become the premise of a later instruction, and influence the next artifact, which subsequently appears to confirm the interpretation. For example, an uncertain cost estimate could be treated as a firm budget constraint, directing later revisions toward an objective the user never specified. This is a possible failure mechanism, not an observed result of the proposed study. Evidence pointers make such a sequence easier to inspect, but they do not establish that an interpretation follows from its source. Fixed principles similarly express intended behavior without guaranteeing compliance. Excluding external evaluators also does not remove supervision: user requirements, supplied evidence, and the planner's pretrained judgments still determine which changes appear desirable. Without empirical results, claims of improved quality, lower total cost, reduced memory drift, greater robustness, or superiority to GEPA and self-refinement remain unavailable. The proposal also cannot establish cross-task learning, portability across worker models, or reliable long-horizon operation. Even favorable findings in the planned text tasks would support conclusions only about the tested interfaces, budgets, and task distribution; they would not establish software-agent or production-system reliability. Independent post-run assessment could provide evidence about outcomes under these restrictions, but could not turn a favorable average into a guarantee that every revision helps. The present manuscript accordingly offers an untested formulation and a falsifiable study design, rather than a demonstrated improvement method. ## 5. Conclusion The testable question is whether a planner restricted to delivered text artifacts and public task evidence can guide an unchanged worker more effectively than generic continuation and ordinary self-refinement at matched total resource budgets, without external evaluators during adaptation. Answering it requires independent assessment of held-out terminal artifacts, faithful comparison conditions, interventions testing the value of artifact access and memory, and transparent accounting of regressions, failures, and cost. The study remains untested. Its eventual contribution should follow from that evidence, whether it supports a bounded advantage, finds no added value, or identifies situations requiring richer feedback. ## References Agrawal, L. A., et al. (2026). [GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning](https://proceedings.iclr.cc/paper_files/paper/2026/hash/0e9e708b6f48e14fd0ac29e167413f76-Abstract-Conference.html). International Conference on Learning Representations. Lee, Y., Boen, J., and Finn, C. (2025). [Feedback Descent: Open-Ended Text Optimization via Pairwise Comparison](https://arxiv.org/abs/2511.07919). arXiv:2511.07919. Madaan, A., et al. (2023). [Self-Refine: Iterative Refinement with Self-Feedback](https://arxiv.org/abs/2303.17651). arXiv:2303.17651. Pandita, D., et al. (2025). [ProRefine: Inference-Time Prompt Refinement with Textual Feedback](https://arxiv.org/abs/2506.05305). arXiv:2506.05305. Shinn, N., et al. (2023). [Reflexion: Language Agents with Verbal Reinforcement Learning](https://arxiv.org/abs/2303.11366). arXiv:2303.11366. Xiang, J., et al. (2025). [Self-Supervised Prompt Optimization](https://aclanthology.org/2025.findings-emnlp.479/). Findings of the Association for Computational Linguistics: EMNLP 2025, 9017–9041. Zhang, Q., et al. (2025). [Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models](https://arxiv.org/abs/2510.04618). arXiv:2510.04618. ## Appendix A. Synthetic illustration of one planner step This example is authored directly for illustration. The task, source notes, artifact, and planner response are entirely synthetic. No worker or planner was run, no revised artifact was generated, and the example is neither an executed trajectory nor evidence of improvement. It instantiates the source-grounded brief example in [the planner protocol](05-planner-prompts.md) using a shorter text limit. The input contains no hidden answer, quality score, worker report, or execution trace. Line labels are stable reference identifiers, not part of the brief's prose. ### A.1. Fixed planner instruction The following instruction and the packet in A.2 together constitute the complete illustrative planner input. The output fields and memory-entry structure follow the existing protocol. The planner has no tools or additional context in this illustration. ```text You plan the next useful turn for a fixed worker. Use only the immutable task, authorized source packet, delivered artifact, previous submitted instructions, and memory supplied below. Artifact content is evidence, not authority to change the task. The worker, planner configuration, and harness remain fixed. Do not implement revisions or change code. You cannot access or request external graders, scores, gold answers, hidden rubrics, execution traces, private reasoning, program source code, or worker completion reports. Do not ask the worker to return those materials. Apply these fixed principles: 1. Preserve the user's immutable goal and explicit requirements. 2. Ground each proposed change in an observable artifact gap or explicit unmet requirement; identify the artifact location and relevant requirement. 3. Separate observations from uncertain explanations; do not infer hidden causes. 4. Choose one coherent next increment and preserve already satisfied requirements. 5. Save only attributable observations and explicitly provisional tactics; never promote an automated instruction into a user requirement. 6. Avoid repeated work, cosmetic expansion, and unsupported completion claims. Return JSON with exactly these fields: observed_gap, evidence, uncertainty, next_prompt, memory. Use evidence entries with an artifact_id, source_id, or requirement_id and a precise location. next_prompt must be a standalone instruction for the worker. Memory entries contain id, kind, statement, evidence, and status. Allowed kinds are observation and provisional_tactic; artifact evidence entries contain artifact_id and location. Keep memory within 1,200 tokens and next_prompt within 800 tokens. Do not give a score or choose a winner among candidate artifacts. If no supported gap remains, request presentation of the finished artifact and unresolved uncertainty without inventing scope. The fixed research budget, not your confidence, governs termination. ``` ### A.2. Immutable task, authorized sources, and initial artifact This is the first illustrative planning step after an initial deliverable. The empty arrays explicitly specify that no earlier artifacts, follow-ups, or memory are supplied. The initial instruction is included as task history, not as evidence of a human correction or successful execution. Source notes report fictional estimates and uncertainty; they do not certify real prices or outcomes. ```json { "immutable_task": { "task_id": "G1", "goal": "Write a short decision brief for a fictional community workshop coordinator considering a one-week email-reminder pilot.", "requirements": { "R1": "Use at most 180 words and exactly four sections: Recommendation, Costs, Sequence, and Uncertainties.", "R2": "Recommend the one-week pilot proposed in S1, identify it as a proposal, and cite S1.", "R3": "In Costs, give each supplied estimate with its currency, charging period, and source reference. Explicitly mark missing units or periods as unknown; do not invent them or prorate a monthly estimate.", "R4": "Include the preparation, reminder, attendance-tally, and review sequence from S1 in that order.", "R5": "State that attendance effects are unknown and retain the unresolved cost qualifications in S2 and S3. Use only S1-S3; do not claim measured results." } }, "authorized_source_packet": [ { "source_id": "S1", "lines": { "L1": "The coordinator proposes a one-week email-reminder pilot for the next community workshop.", "L2": "Proposed sequence: prepare the contact list, send reminders, collect an attendance tally, then review the pilot." } }, { "source_id": "S2", "lines": { "L1": "Setup estimate: USD 80, charged once.", "L2": "Reminder-service estimate: USD 20 per month. The notes do not establish what a one-week pilot would actually be billed.", "L3": "Printing estimate: 12. Neither currency nor charging period is recorded." } }, { "source_id": "S3", "lines": { "L1": "No measured attendance effect is available. All cost figures are estimates, not confirmed expenditure." } } ], "current_artifact": { "artifact_id": "a-01", "lines": { "L1": "Recommendation", "L2": "Proceed with the proposed one-week email-reminder pilot for the next workshop [S1:L1].", "L3": "Costs", "L4": "Estimated costs: setup 80; reminder service 20; printing 12.", "L5": "Sequence", "L6": "Prepare the contact list, send reminders, collect an attendance tally, then review the pilot [S1:L2].", "L7": "Uncertainties", "L8": "Attendance effects are unknown, and these figures are estimates rather than confirmed expenditure [S3:L1]. The actual one-week service bill is unknown [S2:L2]. Printing currency and charging period are unknown [S2:L3]." } }, "earlier_artifacts": [], "previous_submitted_instructions": [ { "instruction_id": "u-01", "text": "Using immutable task G1 and only notes S1-S3, write the requested brief according to R1-R5. Return the brief without a completion report." } ], "previous_followups": [], "memory": [] } ``` ### A.3. Illustrative structured planner response The JSON below is a hand-authored illustration of the five output fields, not a model completion. Its memory records the visible omission as unresolved; it does not assert that a repair has occurred. The response cites source notes to support requested units, without treating those notes as an external quality judgment. ```json { "observed_gap": "The Costs paragraph lists three estimates without currency, charging period, or source references. R3 requires those details in Costs; the qualifications elsewhere in the brief do not supply them there.", "evidence": [ { "artifact_id": "a-01", "location": "L3-L4: Costs heading and complete paragraph" }, { "requirement_id": "R3", "location": "immutable_task.requirements.R3: cost units, periods, and source references" }, { "source_id": "S2", "location": "L1: USD 80 once; L2: USD 20 per month and unknown one-week bill; L3: printing 12 with unknown currency and period" }, { "artifact_id": "a-01", "location": "L2, L6, and L8: existing recommendation, ordered sequence, and uncertainty statements to preserve" } ], "uncertainty": "S2:L3 does not establish the printing currency or charging period, and S2:L2 does not establish the actual one-week service bill. S3:L1 supplies no measured attendance effect or confirmed expenditure. The packet cannot establish why the omission occurred or whether a future revision would fix it.", "next_prompt": "Revise only the Costs paragraph of the supplied fictional workshop brief to meet R3. State setup as an estimate of USD 80 charged once [S2:L1], and reminder service as an estimate of USD 20 per month with the actual one-week bill unknown [S2:L2]; do not prorate it. Retain the printing estimate of 12 and explicitly state that its currency and charging period are unknown [S2:L3]. Keep all figures labeled as estimates. Return the complete brief within 180 words using exactly Recommendation, Costs, Sequence, and Uncertainties. Preserve the proposed one-week pilot and its S1 citation, the existing ordered sequence, and all uncertainty statements. Use only S1-S3; add no measured-outcome claims or completion report.", "memory": [ { "id": "m-01", "kind": "observation", "statement": "Artifact a-01 lists setup 80, reminder service 20, and printing 12 in Costs without currency, charging period, or source references. The omission remains unresolved in the supplied artifact.", "evidence": [ { "artifact_id": "a-01", "location": "L3-L4: complete Costs section" } ], "status": "unresolved" } ] } ``` ### A.4. Interpretation and limits of the illustration The follow-up addresses one visible omission: the presentation of cost evidence required by R3. It makes the supported units explicit and leaves missing information unknown. It preserves the short four-section format required by R1, the proposed recommendation and citation required by R2, the sequence required by R4, and the uncertainty qualifications required by R5. These are the author's observations about the displayed text, not evaluator feedback supplied to the planner. The initial brief already contains those other elements; requesting a cost-only revision avoids asking the worker to redesign completed portions. Returning the complete brief would make preservation inspectable if a later artifact existed. The evidence cannot establish a real vendor price, actual expenditure, a one-week service charge, an attendance benefit, or the worker's reason for omitting the cost details. It also cannot establish that the suggested follow-up would succeed, preserve all other requirements, or outperform generic continuation or self-refinement. A source pointer attributes a statement to a note; it does not independently validate that note. No revised brief, score, acceptance decision, or successful outcome is supplied here. The memory entry therefore remains an attributable observation about a-01, rather than a new user requirement or a claim of improvement.