# B0, B1, and B2 prompt assets Version 1.0. **Hand-authored, untested study assets.** These files define prompts and a proposed assembly contract, not a runner or an executed experiment. They contain no task answers or generated critiques. B2 is **Self-Refine-style**, not a verified reproduction of [Self-Refine](https://arxiv.org/abs/2303.17651). The [experimental protocol](../03-experimental-protocol.md) supplies the comparison conditions, resource accounting, and independent-assessment rules. These assets reuse the completed immutable-task and source-ID format of the [development packets](../development-tasks/README.md), the artifact/history boundary in [05-planner-prompts.md](../05-planner-prompts.md), and its current-plus-two-earlier-artifacts starting policy. The existing treatment prompt, manuscript, and completed task packets remain unchanged. No baseline receives the treatment's six-principle meta-prompt, planner JSON schema, or provenance-linked memory. ## Files and call sequences | Condition | First worker call | Later calls | Artifact selected for assessment | |---|---|---|---| | B0: single-call compute-allocation control | [common-initial-worker.txt](common-initial-worker.txt) | None | Its sole returned deliverable | | B1: generic continuation | Same common initial instruction | [b1-continuation.txt](b1-continuation.txt) once per scheduled subsequent worker turn | Last delivered artifact under the fixed endpoint/budget | | B2: Self-Refine-style | Same common initial instruction | [b2-critique.txt](b2-critique.txt), then [b2-revision.txt](b2-revision.txt) for each subsequent worker turn | Last delivered revision under the fixed endpoint/budget | B2's critique call produces feedback, not a deliverable or a completed worker revision. It consumes resources and must be recorded separately. Ten completed worker turns would mean one initial generation plus nine critique/revision pairs: ten artifact-producing calls and nine critique calls, absent failures or budget exhaustion. B1 would have ten artifact-producing calls. B0 remains exactly one artifact-producing call; making it generate alternatives, critique itself in a second call, or accept a repair prompt would change the condition. No method gets a free discarded attempt. ## Common initial instruction Use the identical common initial template and the identical complete public task packet for B0, B1, B2, and the treatment where the protocol calls for a common first instruction. It accepts any of the three completed development packets without task-specific prompt edits. A prospective caller substitutes only the explicit placeholders; do not manually add hints, computed answers, a preferred decision, or a selected schedule. For B0, assign a prespecified larger single-call generation allowance within the common adaptation-cost cap. Keep the same underlying worker model, task requirements, and source access. The larger allowance is the compute-allocation intervention; it does not relax the task's deliverable word limit or permit additional calls. Record actual consumption because equal resource allowances do not guarantee equal expenditure. Provider limits may prevent B0 from using the entire cap; disclose that rather than manufacturing extra calls. All settings other than the declared generation allowance should match the worker configuration used in the other conditions. ## Placeholder contract The `.txt` files are the complete prompt bodies to render. Explanatory instructions in this README are for the prospective study operator, not additional hidden task instructions. Every placeholder represents a JSON value, not a file path or an instruction to open files: | Placeholder | Required value | |---|---| | `{{TASK_PACKET_JSON}}` | A JSON string encoding the full, unchanged contents of one file under `development-tasks/public/`, including its assignment, provenance, R1–R12, and complete S1–S5 packet | | `{{CURRENT_ARTIFACT_JSON}}` | An object with `artifact_id` and `text`, containing the latest delivered text artifact in full | | `{{EARLIER_ARTIFACTS_JSON}}` | An array of zero, one, or two earlier artifact objects in chronological order, excluding the current artifact | | `{{PREVIOUS_INSTRUCTIONS_JSON}}` | An array of prior worker-instruction objects, each with `instruction_id` and `text`, in chronological order under the common bounded-history policy | | `{{CURRENT_CRITIQUE_JSON}}` | A JSON string encoding the exact current B2 critique, supplied only to the immediately following B2 revision | JSON-encode strings before substitution so quotes and newlines remain unambiguous. Do not wrap already encoded values in extra quotation marks. Task and artifact text is preserved after JSON decoding; source-line labels and requirement IDs are not stripped or rewritten. Empty histories are the literal array `[]`. Never replace a missing task packet or artifact with an assessor answer or a worker success report. An oversized or unavailable required input is a configuration/context problem, not permission to fabricate or summarize it. Each prior worker-instruction `text` contains the exact directive paragraphs from its rendered template, excluding the appended input blocks. This avoids recursively nesting old task/artifact packets inside history. Historical critique text is excluded from these records. Archive full rendered requests outside the permitted context for later audit. Apply the same directive-only history representation to comparison arms; retain the actual task-specific directive text when an arm generates its instructions, rather than replacing it with a template name. Only the public packet and allowed task artifacts enter these placeholders. Do not supply assessor-only guides, criterion scores, external labels, reference answers, tool transcripts, worker reports, private reasoning, repository files, or an external model judge. Source notes can contain conflicting policies or fictional observations; resolve them using the task's explicit requirements and packet precedence, without importing outside facts. ## B1: fixed generic continuation After each scheduled completed worker turn, export the current artifact and render the same B1 template again with updated permitted context. The generic directive is fixed: a human or planner does not choose a new task-specific next action for B1. The worker may revise its own deliverable using ordinary task reasoning. Do not attach an adaptive plan, a critique from another call, a score, or treatment memory. Return a single complete deliverable, even if only a small change is appropriate. An already adequate artifact can be returned unchanged; no correctness claim or special stop token is needed. ## B2: critique visibility without persistent memory At each revision opportunity, construct the allowed context once. Send it to the B2 critique template, then send that same task/artifact/history context to the revision template with the exact returned critique in `{{CURRENT_CRITIQUE_JSON}}`. Do not update the current artifact between those two calls. The revision must see the critique rather than merely a note that a critique exists. The critique is fallible advice generated from the permitted context; it cannot change the immutable assignment or supply new facts. After the revision, export its deliverable as the next current artifact. Keep that pair's critique in the audit record but omit it from future critique and revision inputs. B2 retains only the common bounded artifact/instruction history, not an accumulating feedback buffer, structured lesson list, evidence-linked entries, or a rewritten goal. Any effect of an earlier critique can persist indirectly in a delivered artifact, as ordinary refinement entails. Do not label this absence of a dedicated memory store as absence of all contextual state. By default, critique and revision use the same model snapshot as the worker in separate requests. This follows the same-model role pattern described by Self-Refine while adapting the workflow to this study's text-only boundary. It does not reproduce that paper's task prompts, complete history policy, stopping rules, or all configurations. This package supplies no evidence that the adaptation is equivalent to, or performs like, the published method. ## Proposed reversible defaults These choices make the templates concrete for development; **none is a frozen experimental setting**: - Construct requests from explicit packets without additional implicit chat history. Apply the same worker-context policy to compatible comparison arms before freezing; do not compare this wrapper silently against a treatment with unrestricted conversation access. - Retain the current artifact plus at most the two most recent earlier artifacts, and at most the last twelve worker-instruction directives. In a ten-turn schedule all prior worker directives fit this count bound. Keep immutable task/source content separate and complete. No additional summarization or truncation rule is supplied; if a chosen model cannot fit the complete allowed packet, revise and disclose a common context policy during development before any test runs. - Use up to ten artifact-producing worker turns for B1 and B2, subject to the common total adaptation budget. No quality score, critique confidence, or declaration of completion can extend or shorten that schedule. B0 has one call. - Initially cap each B2 critique at 800 output tokens; a critique's size limit is separate from the public deliverable word limit. Worker-output limits, the common cost cap, model identities, decoding, timeouts, and provider-specific limits must be selected on development data and recorded before an experiment. No numerical cost or worker-token setting is frozen by this package. - Count every critique, revision, initial generation, repeated context read, discarded response, and retry under the common accounting policy. If no budget remains for a B2 revision after its critique, the last existing artifact remains the endpoint and the spent critique still counts. The critique never substitutes for a deliverable. The protocol and manuscript define proposed failure/retry handling. A future configuration must record the chosen policy consistently across arms. B0 does not receive a substantive or schema-repair second call; an incomplete or missing B0 deliverable is recorded under the common failure rule, not converted into a multi-call success. Any permitted transport retry must be disclosed and charged, with a duplicated completion treated as a deviation rather than a free candidate. ## Assessment and status These templates request no scalar scores, criterion pass/fail judgments, pairwise preferences, competing candidate pool, or quality-based acceptance decision. B2's qualitative critique is still evaluation in the ordinary sense; excluding external grading does not remove that judgment. The future study would seal outputs before independent assessment, keeping assessor answers and ratings out of adaptation. No sample filled task response is included and no call sequence has been executed. The templates do not authorize spending API credits or starting experiments.