{"manifest_version":1,"served_method":"Zeno v0.57 live via POST /api/public/v1/check (selected sources corrected, others keep their reference)","candidate":{"run_id":"v057-main-l40sx1-20261007T1138Z","repo":"ZenoDivergent/zeno-divergent-v057","revision":"edc32c3618a36bc792834cb98cb55e8bf70f783e","feature_schema":"v0451-features-1","status":"served","gate_passed":true,"hf_card":"https://huggingface.co/ZenoDivergent/zeno-divergent-v057","note":"Gate-passed per source on dev and the later confirmation period; every other source returns its reference, labelled."},"training_version":"v0.80","surfaces":{"site":{"zeno_check":"/check","watch_it_learn":"/live-read","use_zeno":"/check","arena":"/arena","failure_watch":"/watch"},"api":{"contract":"zeno30-assess/1.5.0","check":"POST /api/public/v1/check","model_identity":{"served_method":"Zeno v0.57 live via POST /api/public/v1/check (selected sources corrected, others keep their reference)","candidate_run_id":"v057-main-l40sx1-20261007T1138Z","candidate_repo":"ZenoDivergent/zeno-divergent-v057","candidate_gate_passed":true,"training_version":"v0.80","note":"gate-passed per source; live model at /api/public/v1/check"}},"mcp":{"endpoint":"POST /api/public/v1/mcp","tool":"zeno_check","contract":"zeno30-assess/1.5.0"},"hf_card":{"url":"https://huggingface.co/ZenoDivergent/zeno-divergent-v057","label":"Download the served model"}},"paper":{"kind":"architecture","version":"0.1","date":"2026-10-02","url":"/architecture","addendum":[{"date":"2026-10-04","version":"v0.45","change":"Task context (asset, location, target, horizon, units, season) added as one hashed label; tested by a matched no-context version."},{"date":"2026-10-04","version":"v0.45.2","change":"A source that gets worse falls back to its own reference; other sources keep their gains (no whole-model veto)."},{"date":"2026-10-04","version":"v0.45.3","change":"Each source records whether it is served as its reference, a reduced correction or the full one; forecast, warning and 90% coverage are reported separately; panel sizes recorded; forecasters never assumed independent."},{"date":"2026-10-19","version":"v0.77r","change":"Research only, not served: audited re-check of the v0.77 CPU tree experiment (time-valid averages, matched calibration, question-level weights); the earlier ForecastBench gain was withdrawn."},{"date":"2026-10-19","version":"v0.78","change":"Research only, proposed and not trained: neural panel encoder with per-pair history, trust and shared-failure heads, judged against the v0.77r tree and four matched controls."},{"date":"2026-10-19","version":"v0.78s","change":"Research only: 50 neural slots trained on one H200; final scoring crashed and the models were lost, so there is no test result. Logged numbers are selection-week scores. Correctness-only successor v0.78t prepared, not run."},{"date":"2026-10-20","version":"v0.80","change":"Research only, not served: relation-aware panel network trained on one H200 (454,490 cases, 16 sources). Flu and energy passed both test periods; audit defects (future-dated forecasts, energy outcome clock, question-level grouping) make them candidates until v0.80.1 reproduces them."}]},"limits":"Source-specific later-data results do not transfer automatically. Standard mode keeps the named reference where a correction did not earn selection; experimental answers are labelled untested."}