复核把「首稿没过、重写后已通过」当缺口:report_repair、report_rewrite 这类「改过一次」的信号按同一个报告节点最后一次 report.checked 定档——最终通过的(没有违规;裁判摘要里不再有证据不支持的句子)不进复核的异常清单,复核模型被别的信号叫起来时,说明里另起一节「已经处理好的」写「报告「x」首稿有 N 处没过核对,重写后已全部通过」,不再写成「核对仍未全部通过、可能不准」;最终没过的照旧记缺口。report_repair、report_rewrite 的日志带上处数

This commit is contained in:
ThearchyHelios
2026-10-19 16:43:54 +08:00
parent 43909ea3f0
commit 1e87504e5c
4 changed files with 129 additions and 5 deletions
+48
View File
@@ -428,3 +428,51 @@ def test_agent_field_warnings_are_degraded_signals():
], OK_OUTPUT)
assert [(s.kind, s.severity) for s in signals] == [
("agent_field_mismatch", rv.DEGRADED), ("agent_field_unverified", rv.DEGRADED)]
# --------------------------------------------------------------------------
# 报告「改过一次」的信号按最终结果定档
# --------------------------------------------------------------------------
def _repaired(final_ok: bool) -> list[dict]:
violations = [] if final_ok else [{"code": "bare_number", "message": "裸数字 12"}]
return [
ev(EventType.NODE_STARTED, node_id="write", label="写周报"),
ev(EventType.LOG, node_id="write", level="warn", code="report_repair", count=2, attempt=1,
message="报告有 2 处未通过核对(裸数字 12、裸数字 30),已要求模型重写(第 1 次)"),
ev(EventType.REPORT_CHECKED, node_id="write", ok=final_ok, violations=violations, repairs=1),
]
def test_a_repair_that_finally_passed_is_not_a_gap():
"""首稿没过核对、重写一次之后通过(最后一次 report.checked ok、没有违规):不进异常清单,复核不为它叫模型;
复核被别的信号叫起来时,说明里写「首稿有 N 处没过核对,重写后已全部通过」。"""
rows = _repaired(True)
assert rv.scan(rows, OK_OUTPUT) == []
[note] = rv.notes(rows)
assert note == "报告「写周报」首稿有 2 处没过核对,重写后已全部通过"
prompt = rv.review_prompt("上周产量", "答案", [rv.Signal("tool_error", "工具 x 调用失败", rv.BROKEN)],
settled=rv.notes(rows))
assert "重写后已全部通过" in prompt and "不要写成缺口" in prompt
assert prompt.index("检测到的异常") < prompt.index("已经处理好的")
assert "首稿有 2 处" not in prompt.split("已经处理好的")[0], "处理好了的不在异常清单里"
def test_a_repair_that_still_failed_stays_a_gap():
rows = _repaired(False)
[signal] = rv.scan(rows, OK_OUTPUT)
assert signal.kind == "report_repair" and signal.severity == rv.DEGRADED
assert rv.notes(rows) == []
def test_a_rewrite_counts_by_the_final_judgement():
base = [ev(EventType.NODE_STARTED, node_id="write", label="写周报"),
ev(EventType.LOG, node_id="write", level="warn", code="report_rewrite", count=1,
message="裁判认为 1 句结论证据不支持(……),已退回写作模型仅修改这几句")]
good = [*base, ev(EventType.REPORT_CHECKED, node_id="write", ok=True, violations=[],
judge={"counts": {"supported": 3, "contradicted": 0, "insufficient": 0}})]
bad = [*base, ev(EventType.REPORT_CHECKED, node_id="write", ok=True, violations=[],
judge={"counts": {"supported": 2, "contradicted": 1, "insufficient": 0}})]
assert rv.scan(good, OK_OUTPUT) == [] and rv.notes(good) == ["报告「写周报」有 1 句结论裁判认为证据不支持,改写后都有了依据"]
assert [s.kind for s in rv.scan(bad, OK_OUTPUT)] == ["report_rewrite"] and rv.notes(bad) == []