# B4: Feedback Descent text-artifact assets Version 1.0. **Hand-authored, untested study assets; defaults are unfrozen.** These templates implement a proposed task-local adaptation, not a verified reproduction or evidence of baseline performance. They contain no task answers or executed trajectories. ## Evidence and scope The existing [related-work map](../../06-evidence-and-related-work.md) identifies Feedback Descent as a close comparator. The primary source inspected is Lee, Boen, and Finn, [*Feedback Descent: Open-Ended Text Optimization via Pairwise Comparison*, arXiv:2511.07919v2](https://arxiv.org/abs/2511.07919v2), including [§2 and Algorithm 1](https://arxiv.org/html/2511.07919v2#S2), [§4.1](https://arxiv.org/html/2511.07919v2#S4.SS1), and [Appendix C.1–C.3](https://arxiv.org/html/2511.07919v2#A3). The [PDF](https://arxiv.org/pdf/2511.07919) also confirms the algorithm below. Read on 2026-11-12; no implementation was executed. | Source finding | Consequence here | |---|---| | §2 proposes an artifact from the incumbent and comparative rationale history; preference controls replacement. | Generate deliverables directly, retaining feedback from rejected proposals. | | Algorithm 1 clears history on acceptance; nearby prose omits that reset, and C.1's SVG description accumulates winning rationales. | Follow Algorithm 1's reset, explicitly; cross-description consistency remains unresolved. | | §4.1's SVG comparisons use both orders, retry inconsistent judgments, and discard unresolved candidates. | Adapt that consistency gate to text; default to three total paired attempts, since retry-count interpretation is not fully specified. | | Prompt experiments use training feedback and validation accuracy; Appendix C describes hypothesis/tagging/statistical analysis. | This reference-free adaptation does not reproduce that scored experiment. | | Appendix C.3 supplies domain-specific templates. | The templates here are newly authored, not verbatim published prompts. | The generic loop and the domain implementations are not interchangeable specifications. Exact historical prompts/settings, rationale handling across inconsistent comparisons, and a universal cross-domain history policy are not established by the inspected sources. The choices below make this adaptation usable without resolving those questions by assertion. No novelty, superiority, convergence, or transfer claim follows. ## Templates and outputs | Asset | Inputs | Output/binding | |---|---|---| | Existing [common-initial-worker.txt](../common-initial-worker.txt) | Complete public task packet | Initial deliverable `x0` | | [proposal.txt](proposal.txt) | Same packet, incumbent, comparison-feedback buffer | One complete candidate deliverable; bind its exact exported text to `CANDIDATE_ARTIFACT_JSON` | | [comparison-rationale.txt](comparison-rationale.txt) | Same packet, anonymously labeled A/B artifacts | JSON `{ "preference": "A" or "B", "rationale": string }` | | [acceptance.txt](acceptance.txt) | Pair IDs and comparison records | JSON action plus incumbent/history update instructions | **A proposal call is the worker's candidate execution.** It returns a deliverable, not an instruction requiring another worker rollout. There is no free proposal stage or extra execution hidden in the accounting. Acceptance is a deterministic rule by default; its standalone template can describe/serialize that rule for an operator or adapter. A model-based acceptance serializer is optional, must be checked against the rule, and incurs additional charged calls. It must never become a new quality judge. ## Placeholder contract Substitute JSON values once, never file paths. Encode strings once without extra quotation marks; preserve decoded text, requirement IDs, and source labels exactly. Use explicit fresh requests, without implicit chat history. | Placeholder | Required value | |---|---| | `{{TASK_PACKET_JSON}}` | JSON string containing one entire unchanged `development-tasks/public/` packet, including assignment, public requirements, provenance, and complete authorized sources | | `{{INCUMBENT_ARTIFACT_JSON}}` | `{ "artifact_id": string, "text": string }`, the retained deliverable | | `{{FEEDBACK_HISTORY_JSON}}` | Chronological array of the feedback entries defined below; initially `[]` | | `{{ARTIFACT_A_JSON}}`, `{{ARTIFACT_B_JSON}}` | Objects with only `{ "text": string }`; incumbent/candidate identity is withheld from the comparator | | `{{INCUMBENT_ID_JSON}}`, `{{CANDIDATE_ID_JSON}}` | Distinct JSON strings identifying the two artifacts, used only for acceptance metadata | | `{{COMPARISON_RECORDS_JSON}}` | Array of records for this proposal only: `{ "attempt": integer, "order": 1 or 2, "a_id": string, "b_id": string, "preference": "A" or "B" or null, "rationale": string or null }` | The operator wraps the proposal's raw deliverable as `CANDIDATE_ARTIFACT_JSON = { "artifact_id": assigned_id, "text": exact_exported_text }`. This is an output binding, not an additional input placeholder or request to generate an answer in this package. Each feedback entry has exactly this assembly structure: `{ "candidate_artifact": { "artifact_id": string, "text": string }, "incumbent_id": string, "comparisons": [comparison records], "outcome": "incumbent_consensus" or "no_consensus" }`. Preserve all comparison texts and A/B-to-ID mappings from that proposal, including inconsistent paired attempts. This schema and retention of inconsistent rationales are **adaptations**. Do not summarize them into asserted truths, invent a consensus rationale, or add the treatment's principles/provenance-memory mechanism. On acceptance, clear the active buffer as Algorithm 1 specifies; archive everything outside model context. Every input must derive solely from the public packet, delivered artifacts, B4's own judgments, and their assembly metadata. **No assessor-only files, hidden scores/criteria, supported-answer sheets, reference solutions, validation labels, external grades, private reasoning, tool traces, worker reports, or repository content may enter any template.** Public arithmetic facts are allowed. B4's qualitative preferences are internal evaluation, not independent assessment. Feedback can be mistaken and cannot amend a public requirement. ## Proposed call sequence 1. **Initialize:** render the existing common initial-worker template unchanged with one public packet. Keep its deliverable as incumbent; set the feedback buffer to `[]`. Use the same worker snapshot/settings as compatible arms. No task-specific starting hints. 2. **Propose:** render `proposal.txt` with the packet, incumbent, and complete active buffer. Export one candidate artifact. Keep the incumbent unchanged while comparisons are pending. 3. **Compare in both orders:** run `comparison-rationale.txt` twice with the same two artifacts and packet, swapping A/B on the second call. Randomize the first orientation using a logged seed as an unfrozen default. The comparator sees neither IDs/method names nor earlier preferences, history, or candidate-generation text. Store its exact rationale and preference with the external mapping. 4. **Apply acceptance:** map both choices to artifact IDs. If both prefer the candidate, replace the incumbent and clear the buffer. If both prefer the incumbent, retain it and append the candidate plus all its comparison records to the buffer. If they disagree, repeat the two-order comparison on the **same** artifacts, up to three total paired attempts. Stop at the first consistent pair; do not take a majority across attempts. If all three pairs disagree, discard the candidate from selection, retain the incumbent, and append its complete comparison record as `no_consensus`. The active buffer remains unchanged while another pair is pending. 5. **Continue and seal:** repeat with the retained incumbent and updated buffer until the fixed endpoint or common resource cap. Archive every artifact, request, comparison, update, failure, and usage entry. The final B4 endpoint is the retained incumbent, potentially from an earlier iteration; do not substitute the latest rejected candidate or choose a historical winner using assessor scores. Seal the registered batch before independent assessment; no subsequent ratings return to adaptation. The two-order gate is a declared text adaptation of the SVG procedure, not the IFBench accuracy gate. Forced A/B preferences remain fallible even when order-consistent. Rationale history is B4-specific methodological state: it can contain more rejected artifacts than the continuation arms' current-plus-two window. Do not silently truncate it to that window or describe these history contents as identical across arms. All arms remain within public task/artifact channels and common resource accounting; a strict identical-history variant would need separate specification before freezing. ## Defaults, failures, and resource use Unfrozen development defaults: one task, one candidate per iteration, up to nine proposals after the common initial call, no quality-based early stopping, three total paired comparison attempts per candidate, and Algorithm 1's reset-on-acceptance rule. The finite schedule bounds the active buffer to at most nine proposal entries; no model-written summary or cross-task memory is added. Preserve the complete packet and active entries. If they exceed context limits, record an incomplete run rather than silently truncating; change the declared policy during development before testing. Require a complete task artifact and comparator JSON with exactly the two documented keys, an A/B preference, and a nonempty rationale. Missing/malformed comparison output ends that iteration as incomplete: retain the incumbent and leave the active buffer unchanged; archive partial feedback outside adaptation. Missing artifacts likewise cannot be promoted. Do not manually improve proposals or use a human quality gate. Parsing and failure rules are adapters, not claimed paper settings. A transport retry, if later authorized under a frozen retry policy, repeats the identical request and counts all consumed resources; it is distinct from the planned extra paired judgments after disagreement. No uncharged repair calls or regeneration of disliked candidates. For `k` completed proposal iterations and `a_i` paired attempts for proposal `i`, with no failures/retries and deterministic acceptance: initial calls = 1, candidate executions = `k`, comparator calls = `2 × sum(a_i)`, total = `1 + k + 2 × sum(a_i)`. With `1 ≤ a_i ≤ 3`, nine proposals cost **28–64 calls**, including ten artifact-producing worker calls. All rejected/no-consensus candidates and comparisons count. Optional model-based acceptance after each pair adds `sum(a_i)` calls, making the analogous range **37–91**. These are arithmetic bounds, not measured usage or results. Any additional planner, execution, rationale-summarization, or repair call changes the condition and must be charged and disclosed. Apply the [protocol's](../../03-experimental-protocol.md) common total adaptation-cost cap as well as the proposed worker-call ceiling; rejected proposals consume that ceiling. Reserve allowance for a complete iteration under chosen per-call caps before starting. If interrupted, charge partial work and retain the last fully selected incumbent, marking incompletion. Record usage and billed cost by role, including repeated context reads, exposed reasoning tokens, retries, and discarded responses. Independent assessment cost is reported separately as study expense. Model identities, decoding, output allowances, context limits, cost caps, retry policy, and seeds remain unfrozen. The comparator may use a separate request to the same model family as the worker as a reversible development default; it is not an independent final assessor. Tune and freeze configurations with matched development allowances before test. Direct deliverable refinement cannot establish family-policy transfer, and excluding native training/validation signals prevents claiming reproduction of the paper's prompt-optimization results. Only prompt/document assets were authored. No models, experiments, production changes, paid calls, or submissions are authorized or performed.