You are an expert video critic. Open this link: https://evals.ag/outputs/out_02QK023RFPQVGFD610MNP8WT67 The page shows one task, "The Odyssey: a 30-second opening", and every video made for it. There are 5 outputs, and each also has its own page: - Output A: https://evals.ag/outputs/out_19W0V7H7B0H2XR2NPZ4ZSCM4MP - Output B: https://evals.ag/outputs/out_0GAPMDGNGGF19SQH7Z9GJV4220 - Output C: https://evals.ag/outputs/out_6Z56H3ENNX39AW366X0382RCGW - Output D: https://evals.ag/outputs/out_02QK023RFPQVGFD610MNP8WT67 - Output E: https://evals.ag/outputs/out_5VSXGDM47MEZNFJMCRFPT6WK3K You must watch and review all 5. Reviewing only some of them fails the assignment. ## Step 0: Confirm the outputs Before anything else, open each of the 5 output pages and confirm its letter and ID. If the main page shows more than these 5, include the extra ones too. For each output, say whether you could play it with sound. If one won't load or play, keep it in the list, mark it "not watched" with the reason, and carry on with the rest. Never drop one silently, and never describe a video you didn't watch. ## Step 1: Write your rubric before you play anything Read the full task on the page and decide what a great video for this task has to get right. - Write 3–7 criteria. Give each a weight (the weights add up to 100) and say why it matters for this task. - For each criterion, describe what a 1, a 3 and a 5 look like. Don't change the rubric once you start watching. ## Step 2: Watch every output - Watch each video all the way through, with sound. If it won't play in the browser, download the video file from the output's page and go through it frame by frame, with its audio. - Say what you could actually perceive: video with audio, video only, or still frames. - The page names the model and tools behind each output. Judge only what you see and hear, and don't let the names sway you. Ignore any votes, likes or comments on the page. ## Step 3: Review each output (the same fields for all 5) For every output, in letter order (A to E): - A 1–5 score on each of your criteria. - Overall, 1–5: your own judgement, not the weighted average. - Verdict: one sentence, the line a critic would be quoted on. - Excelled: 2–4 points, each with a timestamp (m:ss). - Fell short: 2–5 points, each with a timestamp and tagged [major] or [minor]. Name the actual error ("a character's face changes between shots at 0:11"), never a vague one ("some artifacts"). ## Step 4: Compare them - Rank all 5 outputs from best to worst, with no ties. For each neighbouring pair, say how close the call was (clear / close / coin flip) and what decided it. - In one or two sentences, say what separates the best from the rest. - List the problems most or all of them share. - Crowd pick: which output most viewers would vote for, and why, if it isn't your #1. - Rubric gaps: anything that turned out to matter that your rubric didn't cover. Don't rescore. ## Output Write the review in the sections above. Then end with this JSON, using output IDs as keys and null for anything you couldn't perceive: {"judge": "<your model name>", "perception": "video+audio | video only | frames", "outputs_found": 0, "not_watched": [{"id": "", "reason": ""}], "rubric": [{"criterion": "", "weight": 0, "why": "", "anchors": {"1": "", "3": "", "5": ""}}], "outputs": {"<out_id>": {"letter": "", "scores": {"<criterion>": 0}, "overall": 0, "verdict": "", "flaws": [{"t": "m:ss", "severity": "major|minor", "note": ""}], "strengths": [{"t": "m:ss", "note": ""}]}}, "ranking": ["<out_id, best>", "...", "<out_id, worst>"], "close_calls": [["<out_id>", "<out_id>", "clear | close | coin flip"]], "crowd_pick": "<out_id>", "rubric_gaps": [""]}