mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-21 22:52:02 +00:00
证据面板显示核查状态:指标步骤的双盲复算一行,整句的反方审稿判断,遮罩列的输入只写「已遮罩」
指标步骤多一行双盲复算的结果:一致(带盲算值)、不一致(带盲算值和审核人选了什么,没选的照实说)、无法复算(带原因)、 已跳过(关闭了、口径卡上没有来自查询的输入、这个指标不在复算范围内)。结果都写成字,颜色只表达要不要多看一眼。 点开一整句时,反方审稿的判断复用审批卡上的展示(run/ChallengeSummary,只读它的导出),只列这一句,点开看反方的理由和查询。 取自遮罩列的输入,标签照查询表格里遮罩格的样子写「已遮罩」,输入来源里核对不一致时只说不一致、不写两边的值。 接口字段的类型就近写在面板里,不动 types.ts;界面文案放 EVIDENCE_TEXT。 check-evidence 加一节:六种双盲状态的文字和颜色、深浅色对比度、没复核过的不画;遮罩输入的标签、代入式、遮罩说明、输入来源; 整句面板的反方审稿判断、降档说明、点开取审稿记录看理由、没审到的句子不画、390px 不横向溢出。
This commit is contained in:
@@ -473,6 +473,27 @@ export const EVIDENCE_TEXT = {
|
||||
loading: '正在推断来源…',
|
||||
failed: '未能获取推断的来源',
|
||||
},
|
||||
|
||||
// ---- 期 4:核查状态。指标步骤上的双盲复算、取自遮罩列的输入(反方审稿的判断复用审批卡的 CHALLENGE_TEXT)----
|
||||
dualCheck: {
|
||||
label: '双盲复算',
|
||||
status: { agreed: '一致', disagreed: '不一致', unchecked: '无法复算', skipped: '已跳过' } as Record<string, string>,
|
||||
/** 服务端给了一个界面不认识的状态:照写原值,不归到哪一类 */
|
||||
unknown: (status: string) => `未知状态「${status}」`,
|
||||
/** 盲算的值,服务端按口径卡的格式渲染好的 */
|
||||
blind: (value: string) => `盲算 ${value}`,
|
||||
choice: { a: '审核人采用原流程的结果', b: '审核人采用盲算结果', neither: '审核人判断两边都不对' } as Record<string, string>,
|
||||
noChoice: '审核人没有作出选择',
|
||||
/** 跳过的原因:节点直通(设置或节点里关了、口径卡上没有来自查询的输入),或者这个指标不在复算范围里 */
|
||||
skip: {
|
||||
disabled: '双盲复算已关闭',
|
||||
no_inputs: '口径卡上没有来自查询的输入',
|
||||
not_targeted: '这个指标不在复算范围内',
|
||||
} as Record<string, string>,
|
||||
},
|
||||
/** 指标的输入取自数据源设置了遮罩的列:标签和代入式里不显示值 */
|
||||
maskedInput: '取自设置了遮罩的列,不显示值',
|
||||
maskedMismatch: '与快照不一致;取自设置了遮罩的列,不显示值',
|
||||
} as const
|
||||
|
||||
/**
|
||||
|
||||
@@ -22,10 +22,11 @@ import { ABSENCE_CLAIM_TEXT, EVIDENCE_TEXT, JUDGE_TEXT, MERGE_TEXT, NODE_TYPE_LA
|
||||
import { useCatalog } from '../store/catalog'
|
||||
import { askKey, segmentKey, useEvidence, useExplore, useVerdicts } from '../store/evidence'
|
||||
import type {
|
||||
EvidenceBlock, EvidenceDocData, EvidenceInput, EvidenceMerge, EvidenceOnDemand, EvidenceQuoteSource, EvidenceSeal,
|
||||
EvidenceSegment, EvidenceStep, EvidenceUnit, EvidenceVerdict, EvidenceViolation,
|
||||
ApprovalReport, ChallengeCounts, EvidenceBlock, EvidenceDocData, EvidenceInput, EvidenceMerge, EvidenceOnDemand,
|
||||
EvidenceQuoteSource, EvidenceSeal, EvidenceSegment, EvidenceStep, EvidenceUnit, EvidenceVerdict, EvidenceViolation,
|
||||
} from '../types'
|
||||
import { ArtifactViewer, ResultTable } from './AssistantStream'
|
||||
import { ChallengeSummary } from './ChallengeSummary'
|
||||
import { useEvidenceHost } from './evidenceHost'
|
||||
import { EvidenceProvenance } from './EvidenceProvenance'
|
||||
import { CopyChip } from './Markdown'
|
||||
@@ -71,11 +72,59 @@ import {
|
||||
* 多少。推演不是出具,不写运行记录、不改报告。勾选框只在推演时出现在查询步骤的结果窗口里,只有这个指标用到的行能勾,
|
||||
* 表格底下带着推演结果。
|
||||
*
|
||||
* 核查状态(期 4):指标步骤多一行双盲复算的结果(一致、不一致和审核人选了什么、无法复算、已跳过);点开一整句时
|
||||
* 多一块反方审稿的判断,复用审批卡上的展示(run/ChallengeSummary),能点开看反方的理由和查询。取自遮罩列的输入,
|
||||
* 服务端已把值换成「已遮罩」,标签照样写「已遮罩」,输入来源里只说核对结果、不写值。
|
||||
*
|
||||
* 四种摆法:side 从右侧弹出(宽屏),drawer 从底部抽出(窄屏),inline 在 360px 的
|
||||
* 画布右栏里直接栏内展开,dock 放进页面给的一块常驻位置(记录页的「证据」页签)。
|
||||
* 都不是模态的:开着面板照样能在正文里走。
|
||||
*/
|
||||
|
||||
/**
|
||||
* 期 4:证据接口在指标步骤上给的双盲复算状态(api/evidence.py 的 _dual_check_of)。没有双盲复算复核过这份指标集时
|
||||
* 没有这个键。只有证据面板读它,类型就近写在这里
|
||||
*/
|
||||
interface DualCheckState {
|
||||
status: 'agreed' | 'disagreed' | 'unchecked' | 'skipped' | (string & {})
|
||||
node_id?: string
|
||||
/** 不一致时审核人的选择:a 采用原流程、b 采用盲算、neither 两边都不对。没选的没有 */
|
||||
choice?: string
|
||||
/** 一致、不一致时的盲算值,服务端按口径卡的格式渲染好的 */
|
||||
blind_rendered?: string
|
||||
/** 无法复算:原因(人话);已跳过:disabled / no_inputs / not_targeted */
|
||||
reason?: string
|
||||
}
|
||||
|
||||
/** 期 4:证据接口在句子上给的反方审稿判断(api/evidence.py 的 _challenge_of)。没审到这一句时没有这个键 */
|
||||
interface UnitChallenge {
|
||||
verdict: string
|
||||
node_id?: string
|
||||
/** 审稿记录(challenge_artifact)。服务端保存失败时为 null:判断照样有,理由和查询看不了 */
|
||||
artifact?: string | null
|
||||
/** 实际生效的处置:degrade 被推翻的计入出具缺口;flag 只标注 */
|
||||
on_refuted?: string
|
||||
/** 报告撰写节点的名字 */
|
||||
report_label?: string
|
||||
}
|
||||
|
||||
function dualCheckOf(step?: EvidenceStep): DualCheckState | null {
|
||||
const raw = (step as (EvidenceStep & { dual_check?: unknown }) | undefined)?.dual_check
|
||||
return raw && typeof raw === 'object' && typeof (raw as DualCheckState).status === 'string' ? raw as DualCheckState : null
|
||||
}
|
||||
|
||||
function challengeOf(unit: unknown): UnitChallenge | null {
|
||||
const raw = unit && typeof unit === 'object' ? (unit as { challenge?: unknown }).challenge : undefined
|
||||
return raw && typeof raw === 'object' && typeof (raw as UnitChallenge).verdict === 'string' && (raw as UnitChallenge).verdict
|
||||
? raw as UnitChallenge : null
|
||||
}
|
||||
|
||||
/** 取自遮罩列的输入:服务端把值换成「已遮罩」,masked 写明是哪几列 */
|
||||
function maskedOf(input: EvidenceInput): string[] {
|
||||
const raw = (input as EvidenceInput & { masked?: unknown }).masked
|
||||
return Array.isArray(raw) ? raw.filter((c): c is string => typeof c === 'string' && !!c) : []
|
||||
}
|
||||
|
||||
export type PanelMode = 'side' | 'drawer' | 'inline' | 'dock'
|
||||
export type PanelView = { kind: 'seg'; id: string } | { kind: 'unit'; id: string } | { kind: 'violations' }
|
||||
|
||||
@@ -537,6 +586,8 @@ function ClaimBody({ doc, artifact, runId, runClass, unit, block }: {
|
||||
const cites = (unit.cites ?? []).map((alias) => ({ alias, label: doc.catalog?.[alias]?.label as string | undefined }))
|
||||
// 讲方法、讲结构的句子:SQL 和字段清单直接列在挂的依据下面。正文不是封存的那份时不列(链是另一份报告的)
|
||||
const basis = useMemo(() => (foreign ? null : claimBasis(unit, doc)), [foreign, unit, doc])
|
||||
// 反方审稿对这一句的判断:服务端只认审的就是封存的这一份文档;正文不是那一份时不用
|
||||
const challenge = foreign ? null : challengeOf(answered?.unit)
|
||||
return (
|
||||
<>
|
||||
{foreign && (
|
||||
@@ -550,6 +601,7 @@ function ClaimBody({ doc, artifact, runId, runClass, unit, block }: {
|
||||
onDemand={foreign ? undefined : answered?.unit?.on_demand}
|
||||
answered={foreign ? undefined : answered?.unit?.verdict}
|
||||
pending={!!runId && !!first && (!slot || slot.status === 'loading')} />
|
||||
{challenge && <ChallengePart challenge={challenge} unit={unit} report={report} runId={runId} />}
|
||||
<Part title={JUDGE_TEXT.cites} data-ev-cites="">
|
||||
{cites.length ? (
|
||||
<ul className="space-y-0.5">
|
||||
@@ -577,6 +629,34 @@ function ClaimBody({ doc, artifact, runId, runClass, unit, block }: {
|
||||
)
|
||||
}
|
||||
|
||||
const CHALLENGE_VERDICTS = ['stands', 'weakened', 'refuted', 'inconclusive'] as const
|
||||
|
||||
/**
|
||||
* 反方审稿对这一句的判断(期 4)。复用审批卡上的展示(run/ChallengeSummary):只列这一句,点开看反方的理由和它执行的
|
||||
* 查询。判断是反方模型给的、不是系统核对,那一块自己写明;正式运行里被推翻、按降档处理的,同样写出具降档
|
||||
*/
|
||||
function ChallengePart({ challenge, unit, report, runId }: {
|
||||
challenge: UnitChallenge; unit: EvidenceUnit; report?: string; runId?: string
|
||||
}) {
|
||||
const reports = useMemo<ApprovalReport[]>(() => {
|
||||
const counts: ChallengeCounts = { stands: 0, weakened: 0, refuted: 0, inconclusive: 0 }
|
||||
const known = CHALLENGE_VERDICTS.find((v) => v === challenge.verdict)
|
||||
if (known) counts[known] = 1
|
||||
return [{
|
||||
node_id: report ?? '', label: challenge.report_label || report || '',
|
||||
challenge: {
|
||||
counts, items: [{ unit: unit.id, text: unitText(unit).trim(), verdict: challenge.verdict }],
|
||||
artifact: challenge.artifact ?? null, on_refuted: challenge.on_refuted,
|
||||
},
|
||||
}]
|
||||
}, [challenge, unit, report])
|
||||
return (
|
||||
<div data-ev-challenge={challenge.verdict}>
|
||||
<ChallengeSummary reports={reports} runId={runId ?? ''} />
|
||||
</div>
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* 模型的解释:只对结论句。判定、理由(斜体)、「模型判断 · 模型名 · 非确定」的徽标,封存之后按需追加的
|
||||
* 另写「封存后追加」;到上限没判的写明哪个上限、到哪里调;改写过一次的写明原句。探索运行里还没判过的
|
||||
@@ -1168,6 +1248,8 @@ function MetricPart({ step, entry, inputs, runId, seal, pending, failed, resolve
|
||||
// 所依据的查询对照数据目录查出了错误级的问题:值照算,结果不可靠,出具同样按缺口降档。问题和改法在查询步骤里
|
||||
const sqlFailed = step?.sql_check_failed === true || entry?.sql_check_failed === true
|
||||
const sqlReason = step?.sql_check_reason || entry?.sql_check_reason || EVIDENCE_TEXT.sqlCheckFallback
|
||||
// 期 4:双盲复算复核过这个指标才有
|
||||
const dualCheck = dualCheckOf(step)
|
||||
|
||||
return (
|
||||
<Part title={EVIDENCE_TEXT.metric} data-ev-metric="">
|
||||
@@ -1200,6 +1282,7 @@ function MetricPart({ step, entry, inputs, runId, seal, pending, failed, resolve
|
||||
{upgrade && (
|
||||
<p className="mt-0.5 text-2xs" style={{ color: 'var(--st-waiting)' }} data-ev-caliber-upgrade="">{upgrade}</p>
|
||||
)}
|
||||
{dualCheck && <DualCheckLine check={dualCheck} />}
|
||||
<Integrity step={step} seal={seal} />
|
||||
{missing && <p className="mt-1" style={{ color: 'var(--st-waiting)' }} data-ev-missing="">{EVIDENCE_TEXT.missingValue}</p>}
|
||||
{incomplete && (
|
||||
@@ -1255,6 +1338,35 @@ function MetricPart({ step, entry, inputs, runId, seal, pending, failed, resolve
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* 这个指标的双盲复算结果(期 4),一行:一致(带盲算值)、不一致(带盲算值和审核人选了什么)、无法复算(带原因)、
|
||||
* 已跳过(带原因)。结果都写成字,颜色只表达要不要多看一眼:一致用确定的绿(盲值是代码按原式代回算的、和原值比对的,
|
||||
* 不是模型的判断);不一致而审核人采用了原流程用提醒色,采用盲算、两边都不对、没有选的用失败色(出具计入缺口);
|
||||
* 无法复算用提醒色(同样计入缺口);跳过的中性色
|
||||
*/
|
||||
function DualCheckLine({ check }: { check: DualCheckState }) {
|
||||
const T = EVIDENCE_TEXT.dualCheck
|
||||
const host = useEvidenceHost()
|
||||
const { status } = check
|
||||
const blind = (status === 'agreed' || status === 'disagreed') && check.blind_rendered ? T.blind(check.blind_rendered) : ''
|
||||
const detail = status === 'disagreed' ? (check.choice ? T.choice[check.choice] ?? check.choice : T.noChoice)
|
||||
: status === 'skipped' ? (check.reason ? T.skip[check.reason] ?? check.reason : '')
|
||||
: status === 'unchecked' ? check.reason ?? '' : ''
|
||||
const color = status === 'agreed' ? 'var(--st-done)'
|
||||
: status === 'disagreed' ? (check.choice === 'a' ? 'var(--st-waiting)' : 'var(--st-failed)')
|
||||
: status === 'unchecked' ? 'var(--st-waiting)' : 'var(--text-dim)'
|
||||
return (
|
||||
<p className="mt-1 text-2xs leading-relaxed [overflow-wrap:anywhere]" data-ev-dual-check={status}
|
||||
data-ev-dual-check-choice={status === 'disagreed' ? check.choice ?? 'none' : undefined}>
|
||||
<span className="text-faint">{T.label}:</span>
|
||||
<span className="font-medium" style={{ color }}>{T.status[status] ?? T.unknown(String(status))}</span>
|
||||
{blind && <span className="tnum text-dim">({blind})</span>}
|
||||
{detail && <span className="text-dim">,{detail}</span>}
|
||||
{host.onNode && check.node_id && <> <NodeChip id={check.node_id} /></>}
|
||||
</p>
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* 证据接口逐项复核的结果:eid 能重算、工件哈希复验、重新渲染一致、在封存范围内。
|
||||
* 全过的时候不说(正常态安静);哪一项明确没过就用失败色写出来——这时候这件证据
|
||||
@@ -1283,27 +1395,35 @@ function IntegrityList({ code, items }: { code: string; items: string[] }) {
|
||||
function InputChip({ input, link }: { input: EvidenceInput; link?: InputLink }) {
|
||||
const host = useEvidenceHost()
|
||||
const missing = input.status === 'missing' || input.value == null
|
||||
// 取自遮罩列(期 4):服务端已把值换成「已遮罩」,这里照查询表格里遮罩格的样子写,读着不像一个真值
|
||||
const masked = !missing && maskedOf(input).length > 0
|
||||
// Agent 交来、被截断切开的数组字段:对它计数、求和只算到了取回的那部分
|
||||
const cut = !missing && input.truncated === true
|
||||
const node = input.node_id ? host.nodeLabel?.(input.node_id) || input.node_id : ''
|
||||
const via = input.via ? VIA_TEXT[input.via] ?? nodeTypeLabel(input.via) : ''
|
||||
const from = [node && `来自节点「${node}」`, via && `(${via})`, input.role && ` · ${ROLE_TEXT[input.role] ?? input.role}`,
|
||||
cut && ` · ${EVIDENCE_TEXT.inputTruncated}`].filter(Boolean).join('')
|
||||
cut && ` · ${EVIDENCE_TEXT.inputTruncated}`, masked && ` · ${EVIDENCE_TEXT.maskedInput}`].filter(Boolean).join('')
|
||||
const style = missing || cut ? { color: 'var(--st-waiting)', borderColor: 'var(--st-waiting)' } : { color: 'var(--text)' }
|
||||
const shown = masked ? EVIDENCE_TEXT.masked : valueText(input.value)
|
||||
const body = (
|
||||
<span className="truncate">{valueText(input.value)} ← {input.path ?? NONE}{missing ? '(缺失)' : cut ? '(已截断)' : ''}</span>
|
||||
<span className="truncate">
|
||||
{masked ? <span className="font-sans italic text-faint">{shown}</span> : shown} ← {input.path ?? NONE}
|
||||
{missing ? '(缺失)' : cut ? '(已截断)' : ''}
|
||||
</span>
|
||||
)
|
||||
return (
|
||||
<li className="max-w-full">
|
||||
{link ? (
|
||||
<button type="button" className="chip mono tnum max-w-full transition-colors hover:bg-hover" style={style}
|
||||
data-ev-input={input.path ?? ''} title={`${from ? `${from} · ` : ''}${EVIDENCE_TEXT.gotoQuery(link.alias)}`}
|
||||
aria-label={`${valueText(input.value)} ← ${input.path ?? NONE},${EVIDENCE_TEXT.gotoQuery(link.alias)}`}
|
||||
data-ev-input={input.path ?? ''} data-ev-input-masked={masked ? '' : undefined}
|
||||
title={`${from ? `${from} · ` : ''}${EVIDENCE_TEXT.gotoQuery(link.alias)}`}
|
||||
aria-label={`${shown} ← ${input.path ?? NONE},${EVIDENCE_TEXT.gotoQuery(link.alias)}`}
|
||||
onClick={link.go}>
|
||||
{body}<CornerDownRight size={9} className="shrink-0 opacity-60" aria-hidden />
|
||||
</button>
|
||||
) : (
|
||||
<span className="chip mono tnum max-w-full" data-ev-input={input.path ?? ''} style={style} title={from || undefined}>
|
||||
<span className="chip mono tnum max-w-full" data-ev-input={input.path ?? ''} data-ev-input-masked={masked ? '' : undefined}
|
||||
style={style} title={from || undefined}>
|
||||
{body}
|
||||
</span>
|
||||
)}
|
||||
@@ -1333,7 +1453,7 @@ function SourcesPart({ inputs, linkOf }: { inputs: EvidenceInput[]; linkOf: (inp
|
||||
<Part title={EVIDENCE_TEXT.sources} data-ev-sources="">
|
||||
<ul className="space-y-1.5">
|
||||
{inputs.map((inp, i) => {
|
||||
const said = inputSource(inp)
|
||||
const said = maskedOf(inp).length ? maskedSource(inp) : inputSource(inp)
|
||||
const tone = TONE[said.tone]
|
||||
const link = linkOf(inp)
|
||||
// 查询编号写报告目录里的全局编号(接口补的 query、对上的查询步骤);agent 字段自己的 ref 是
|
||||
@@ -1372,6 +1492,16 @@ function SourcesPart({ inputs, linkOf }: { inputs: EvidenceInput[]; linkOf: (inp
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* 取自遮罩列的输入在「输入来源」里怎么说(期 4):核对结果照说,只是不写值——核对不一致时 inputSource 会把两边的值都写
|
||||
* 出来。另附一句为什么不显示值
|
||||
*/
|
||||
function maskedSource(inp: EvidenceInput): { tone: SourceTone; text: string; reason?: string } {
|
||||
if (inp.status === 'mismatch') return { tone: 'warn', text: EVIDENCE_TEXT.maskedMismatch }
|
||||
const said = inputSource(inp)
|
||||
return { ...said, reason: said.reason ? `${said.reason};${EVIDENCE_TEXT.maskedInput}` : EVIDENCE_TEXT.maskedInput }
|
||||
}
|
||||
|
||||
/**
|
||||
* 查询步骤:SQL(带复制)、结果窗口(被引用的行和格高亮)、窗口说明、打开完整快照、复核。
|
||||
*
|
||||
|
||||
Reference in New Issue
Block a user