TITLE: TEST One FPCC

Run an actual goal-orchestrator judging job for the Fren Pet costume contest. Dispatch twenty separate goal-task worker items, each producing one independent visual judging report, and compile their real results. Do not simulate judges within one response. First inspect whether the controller supplied .imd/reads/goal.json, a native episode, and sufficient item/episode budget. Do not create, modify or spoof controller state. If this context is missing, write both required outputs with status unsupported_goal_context, zero dispatched/completed judges, null winners and the exact missing capability; stop without simulating a panel. If native context is present, follow its episode protocol and write decisions to the controller-designated path; use artifacts/goal/decisions.json only if this objective must supply the path.

Eligible costumes A-I: https://fren-pet-h2026-collage.rckmtlshared.chatgpt.site/H2026-collage-A-I.png . Layout A,B,C / D,E,F / G,H,I. Normal-pet reference: https://fren-pet-h2026-collage.rckmtlshared.chatgpt.site/Fren-Pet-normal-appearance-reference.png . The reference is not an entry.

Each judge independently acts as a guest judge at a community costume contest and ranks all nine entries by overall impression, without numerical scores. Guidelines: creativity, convincing costume transformation and purposeful execution; favor visible costume-related props, coordinated pieces and inventive use beyond ordinary headwear. Fully covering the original pet is allowed. Ordinary hat-only or visibly poor designs are less competitive. Ignore intrinsic pet traits, rarity, backgrounds, labels and screenshot quality. Judge visible results, not guessed effort or human popularity. Each report gives one concrete visual observation per entry, its complete preference order, and brief explanations of its first and second choices. If either image cannot be inspected, abstain explicitly; never fabricate observations.

Use one research batch of 20 items keyed judge-01 through judge-20, one ballot per item, with no access to other ballots while judging. Do not use consensus replicas or Oracle quorum. Use native canary checks if required. Preserve every real report. Record worker/model identities only from actual platform metadata, otherwise unknown; item labels do not prove distinct workers. Disclose repeated workers/models, missing judges and abstentions separately. Never replace dissenting votes. Merge without adding a compiler vote. Most first-choice votes selects first place. Remove the unique winner from each valid preference order and tally the highest remaining choice for second place. If first place ties, report tied leaders and leave second place unresolved; report a second-place tie without inventing a tiebreak. Publish individual ballots, vote totals, first/second results, uncertainty and short explanations in artifacts/goal/contest-report.md and artifacts/goal/ballots.json. Distinguish planned, dispatched, completed and valid-ballot counts. Do not open additional paid orders, purchase anything, deploy contracts or send transactions.