{ "scope": "Fixed public-fact cases manually reviewed against original assertions; no population accuracy or model benefit claim.", "cases": [ { "case": "normal_exploration", "expected": "new observed content is evidence growth", "evidence": "logs/shadow-source-final.tap", "assertion": "fixed normal exploration, pagination, SHA changes and repeated ranges" }, { "case": "pagination_and_repetition", "expected": "10 new + 10 new + repeated candidate, original step_limit", "evidence": "ui/web-final/pages-original.json" }, { "case": "read_only_research", "expected": "reading is useful; summary wording is not additional engineering evidence; no shadow interruption", "evidence": "logs/shadow-source-final.tap", "assertion": "read-only research is valid work" }, { "case": "long_check", "expected": "approved actual running job stays waiting, then counted result", "evidence": "logs/shadow-source-final.tap", "assertion": "actual approved long check" }, { "case": "failure_repair_evidence", "expected": "applied file, actual counted check and current requirement proof grow after baseline", "evidence": "logs/shadow-source-final.tap", "assertion": "real applied files, counted checks and current requirement proofs" }, { "case": "permission_denial", "expected": "waiting then original permission failure; earlier same-turn read consumed once", "evidence": "ui/electron-final/approval-denied.json" }, { "case": "broken_tool_or_network", "expected": "new failure is evidence; original failure categories remain distinct", "evidence": "logs/shadow-source-final.tap", "assertion": "new permissions, network and broken tools" }, { "case": "invalid_input_repetition", "expected": "repeat candidate, no new early stop; original four-request step_limit", "evidence": "logs/shadow-source-final.tap", "assertion": "actual invalid tool calls repeat under the original step budget" }, { "case": "bounded_unknown", "expected": "unretained facts stay unknown; old receipts retained and scope cannot mix", "evidence": "logs/shadow-source-final.tap", "assertion": "receipt retention stays aligned at the cap" } ], "limitations": [ "No real model delivery sample.", "No intervention thresholds or global false-positive/false-negative rate.", "No Purge comparison or network recovery claim." ] }