mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 20:21:58 +00:00
ThearchyHelios
e4f2693918
「引用即渲染」：写报告的模型只写引用标记（[[m:指标id]]、[[i:输入字段]]、[[see:…]]），数值由系统按口径卡
填进去，每个数字都挂着一条能复算的证据链。本期是数字层闭环；实体、结论句的概率性证据放在后续几期。
内核（engine/evidence.py，口径卡、出具契约、表达式等）
- 口径卡记录每个指标的来源：原式、代入式（每个输入写成「值 ← 路径」）、复算结果、口径与版本
- 出具契约的数值匹配改为按位置标注；同一个值有几个出处时记为「出处不唯一」并列出候选，不再随便挑一个
- 证据 id 由内容派生，事后对不上就是被改过；零值、结构性数字（列表序号、代码块标签）单独处理
- 口径卡新增 on_missing：缺输入时默认仍是整个节点失败，可以改成记为空值、交给契约判档
报告撰写节点（report）与引用模式的出具契约
- 流式生成时 token 里就是真实数字，不会漏出原始标记；裸数字、不存在的引用当场判违规，交回去重写
（max_repairs，默认 1 次）；仍不过时按 on_violation 失败（正式运行默认）或把违规标在报告里（探索运行默认）
- report.checked 事件记下核对结果，落在封存范围内；fail 模式下仍不过时带 failed，运行记录里画成失败
- 出具契约写了 report_from 就按引用核对：取回文档时复验哈希，独立重算一遍，全部带引用才判完整出具
- 成果字段逐字等于报告正文时标注 _evidence；复核只能加说明，不能改写带证据的答案（前后端各拦一道）
证据接口
- GET /api/runs/{id}/evidence 返回证据图；GET …/evidence/segments/{id} 返回一个片段的证据链
- 只认封存范围内的事件能追到的工件，不读 artifacts 表；封存被改过时不把任何一项标成已封存
- 没有报告的旧运行降级为 none 或 legacy_contract
界面
- EvidenceDoc 按块、句、片段渲染报告，片段是和正文一样的按钮；整份报告只占一个 Tab 位，
←/→ 在片段间移动、↑/↓ 按句、n/N 跳到下一处「无证据」、Enter 打开面板、Esc 关闭并把焦点还回去
- EvidencePanel：宽屏侧边、窄屏底部抽屉、360px 画布右栏里栏内展开；显示指标步骤、代入式、复算结果、
封存状态。正文不是封存的那份报告时明确说出来，不拿别的报告的封存画绿
- 出具横幅多一行「N/N 数字有出处 · 无证据 M」和「定位下一处」；没有 _evidence 的旧回答渲染一字不差
- 画布认得报告撰写节点；Copilot 手册补上口径卡、报告节点、出具契约的写法；新增模板「⑨ 可追溯周报」
验证（沙箱）：tsc 干净，生产构建通过，pytest 1601 passed，check-all 16 项检查全绿（新增 check-evidence）。
88 lines
4.6 KiB
Python
88 lines
4.6 KiB
Python
"""旧契约（按数值回指）的止血：每个数字记下位置，同值的多个指标不再默认算给第一个。
|
|
|
|
真实反例：code 节点解析失败后 13 个数值全部兜底成 0，叙述里 56 个「命中」全部算给了
|
|
同一个指标——`next(...)` 取第一个命中的。界面上那 56 个数都标着「来自指标 A」，
|
|
其实哪个都说不清。位置也没记：同一个 token 出现多次共用一个标记，只能按字符串全局
|
|
匹配，标错地方是必然的。
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from app.engine.issuance import decide_tier, trace_numbers
|
|
|
|
ZEROS = [{"id": f"k{i}", "value": 0.0} for i in range(5)]
|
|
|
|
|
|
def test_same_value_metrics_are_ambiguous_not_first_wins():
|
|
narrative = "订单 0.0 单，退款 0.0 单，投诉 0.0 起，复购 0.0 次，流失 0.0 户"
|
|
report = trace_numbers(narrative, ZEROS)
|
|
assert len(report.matched) == 5 and not report.unmatched
|
|
for hit in report.matched:
|
|
assert hit["metric"] is None, hit
|
|
assert hit["ambiguous"] is True
|
|
assert hit["candidates"] == ["k0", "k1", "k2", "k3", "k4"]
|
|
|
|
|
|
def test_every_number_carries_its_offsets():
|
|
narrative = "本周订单 1,204 单，客单价 35.5 元，另有 7 单；复购 1,204 单"
|
|
report = trace_numbers(narrative, [{"id": "orders", "value": 1204}, {"id": "aov", "value": 35.5}])
|
|
assert [h["token"] for h in report.matched] == ["1,204", "35.5", "1,204"]
|
|
for hit in [*report.matched, *report.unmatched]:
|
|
assert narrative[hit["start"]:hit["end"]] == hit["token"], hit
|
|
# 同一个 token 出现两次，位置各是各的
|
|
assert report.matched[0]["start"] != report.matched[2]["start"]
|
|
assert report.unmatched == [{"token": "7", "context": report.unmatched[0]["context"],
|
|
"start": narrative.index("7 单"), "end": narrative.index("7 单") + 1}]
|
|
|
|
|
|
def test_offsets_for_chinese_numerals_and_suffixes():
|
|
narrative = "客户两千五百家，营收 3.5万，增长三成"
|
|
report = trace_numbers(narrative, [{"id": "c", "value": 2500}, {"id": "r", "value": 35000},
|
|
{"id": "g", "value": 0.3}])
|
|
assert [narrative[h["start"]:h["end"]] for h in report.matched] == ["两千五百", "3.5万", "三成"]
|
|
|
|
|
|
def test_unique_hit_keeps_the_old_shape():
|
|
"""只有一个指标对得上时，形状和以前一样（多了位置），老前端照常能读。"""
|
|
report = trace_numbers("订单 120 单", [{"id": "orders", "value": 120}, {"id": "aov", "value": 35.5}])
|
|
assert report.matched == [{"token": "120", "metric": "orders", "start": 3, "end": 6}]
|
|
|
|
|
|
def test_same_metric_id_from_two_calibers_is_not_ambiguous():
|
|
"""两张口径卡都定义了 orders：出处的指标是同一个，不算出处不唯一。"""
|
|
report = trace_numbers("订单 120 单", [{"id": "orders", "value": 120}, {"id": "orders", "value": 120}])
|
|
assert report.matched == [{"token": "120", "metric": "orders", "start": 3, "end": 6}]
|
|
|
|
|
|
# --------------------------------------------------------------------------
|
|
# decide_tier 的新参数：默认都是空，旧调用一个字不改、结果不变
|
|
# --------------------------------------------------------------------------
|
|
|
|
BASE = {"missing_required": [], "missing_expected": [], "unmatched": [], "strict": False}
|
|
|
|
|
|
def test_old_calls_are_unchanged():
|
|
assert decide_tier(**BASE) == "formal"
|
|
assert decide_tier(**{**BASE, "gaps": ["x"]}) == "degraded"
|
|
assert decide_tier(**{**BASE, "unmatched": [{"token": "1"}], "strict": True}) == "withheld"
|
|
|
|
|
|
@pytest.mark.parametrize("strict, tier", [(False, "degraded"), (True, "withheld")])
|
|
def test_unresolved_references_count_like_unmatched_numbers(strict, tier):
|
|
assert decide_tier(**{**BASE, "strict": strict}, unresolved=[{"ref": "m:gmvx"}]) == tier
|
|
|
|
|
|
def test_claims_are_ignored_without_a_policy():
|
|
assert decide_tier(**BASE, unsupported=[{"unit": "u1"}], uncited_claims=3) == "formal"
|
|
assert decide_tier(**BASE, unsupported=[{"unit": "u1"}], claims_policy="off") == "formal"
|
|
|
|
|
|
def test_claims_policy_degrades_or_withholds():
|
|
assert decide_tier(**BASE, unsupported=[{"unit": "u1"}], claims_policy="judge") == "degraded"
|
|
assert decide_tier(**BASE, uncited_claims=2, claims_policy="require_citation") == "degraded"
|
|
policy = {"policy": "judge", "on_unsupported": "withhold", "on_uncited": "ignore"}
|
|
assert decide_tier(**BASE, unsupported=[{"unit": "u1"}], claims_policy=policy) == "withheld"
|
|
assert decide_tier(**BASE, uncited_claims=[{"unit": "u2"}], claims_policy=policy) == "formal"
|
|
assert decide_tier(**BASE, unsupported=[], uncited_claims=0, claims_policy="judge") == "formal"
|