Apply the fixed B4 acceptance rule to comparison metadata. Do not evaluate artifact quality or invent another vote. This is an adaptation's serialization template, not an extra generative acceptance call required by Feedback Descent. Deterministic application is the default. If used in a model request, charge that call and verify the returned record against the rule. Inputs identify one incumbent and one candidate. Comparison records contain paired judgments on those same artifacts. Require distinct pair IDs and exactly 2, 4, or 6 records, in attempt/order sequence. Attempts must start at 1, be consecutive, and have exactly orders 1 and 2. In each attempt, a_id and b_id must be the two supplied IDs in opposite orders between the two records. Each record must have preference exactly A or B and a nonempty rationale string. Map A to a_id and B to b_id; do not reinterpret the rationale as a vote. If any required record is missing, malformed, null, duplicated, or incorrectly mapped, keep the incumbent, mark the comparison incomplete, and leave active history unchanged. Do not repair or synthesize feedback. Inspect complete attempts chronologically. Both mapped choices agreeing on one artifact form a consensus pair. Stop at the first such pair; any supplied records after it make the input invalid and follow the incomplete rule above. If the consensus favors the candidate, accept it and clear the active feedback buffer (Algorithm 1 reset convention). If it favors the incumbent, retain it and append this proposal plus all its comparison records to the active buffer. If all supplied pairs disagree and fewer than three attempts have occurred, request another swapped-order pair on the same artifacts; leave incumbent and buffer unchanged while waiting. If all three attempts disagree, keep incumbent and append the proposal plus all six records with outcome no_consensus. Never take a majority across attempts, choose a favorite rationale, generate a new artifact for a retry pair, or consult external scores or assessor answers. Return exactly one JSON object with these keys and the indicated value types: { "incumbent_id": , "candidate_id": , "action": <"accept_candidate", "keep_incumbent", or "compare_again">, "selected_id": , "comparison_complete": , "decisive_attempt": , "history_action": <"clear", "append_proposal_feedback", or "unchanged">, "reason": <"candidate_consensus", "incumbent_consensus", "no_consensus", "awaiting_consensus", or "invalid_records"> } Use history_action clear only for acceptance; append_proposal_feedback only for incumbent_consensus or no_consensus; unchanged otherwise. For invalid records, use keep_incumbent with reason invalid_records. Replace the angle- bracket specifications with actual JSON values. Return no Markdown, artifact, rewritten rationale, memory contents, confidence, score, or extra fields. Actual artifact/buffer updates are performed by the operator/adapter using the preserved exact texts, not by this record-generating step. INCUMBENT_ARTIFACT_ID_JSON {{INCUMBENT_ID_JSON}} CANDIDATE_ARTIFACT_ID_JSON {{CANDIDATE_ID_JSON}} COMPARISON_RECORDS_JSON {{COMPARISON_RECORDS_JSON}}