From 515e77e9ce1954e8b150dfe5ad5bbf19c484e41f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=B0=8F=E5=94=AF?= Date: Mon, 16 Nov 2026 11:52:22 +0800 Subject: [PATCH] =?UTF-8?q?fix(metrics):=20=E6=8B=86=E5=8F=A3=E5=BE=84=20a?= =?UTF-8?q?vg=5Fdistill=5Floss=20=E4=B8=89=E6=9D=A5=E6=BA=90=20+=20Overall?= =?UTF-8?q?=20=E7=9C=9F=E5=AE=9E=E5=88=86=E5=B8=83=20+=20dashboard=20?= =?UTF-8?q?=E6=A0=87=20stub?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit t_e5cdc25c 织忆指标口径修复: ① 交付 Overall 真实分布复算脚本 ~/.hermes/scripts/tools/overall-dist-recalc.py (本机无 Python lancedb,改用 Rust 只读工具 qs-dist 全表扫描 46023 行; 字段缺失=0,fallback(==0)=15.0%,历史bug(0.755)=7.9%,可信=77.1%) ② 拆 avg_distill_loss 混算:新增 distill_loss_llm / merge_ratio_loss / quality_backtrace_loss 三独立指标(各带 _count/_valid/_sum),avg_distill_loss 保留但标 deprecated=1、不再作为可信尺子;server.go 三调用点迁移到来源专属 Record。 契约测试 metrics_contract_test.go 增 TestContract_DistillLossSplitThreeSources 验证三来源互不串桶 + sum/count 可复算。go test ./internal/selfoptimize/ 全绿。 ③ DashboardSnapshot 6 硬编码值标 dashboard_stub=true(serde default=true 保守), 逐点注释说明:previous_metrics 恒 None ⇒ 退化判定链路死路;Go consolidate::Result 不读 dashboard/degradation 字段 ⇒ 硬编码值不流入 Go 自优化/调参/告警。 不改蒸馏逻辑(engine.go 评分权重不动);不重启 zhiyid、不改生产 env。 --- go/internal/api/server.go | 12 +- go/internal/selfoptimize/metrics_contract.go | 6 +- .../selfoptimize/metrics_contract_test.go | 69 ++++++++- go/internal/selfoptimize/selfoptimize.go | 135 +++++++++++++++++- rust/src/main.rs | 4 + rust/src/report.rs | 27 +++- 6 files changed, 240 insertions(+), 13 deletions(-) diff --git a/go/internal/api/server.go b/go/internal/api/server.go index d6b4b11..f5df926 100644 --- a/go/internal/api/server.go +++ b/go/internal/api/server.go @@ -616,7 +616,7 @@ func NewServer() http.Handler { } // 更新仪表盘指标:蒸馏成功次数(2026-10-14 t_c06ad4d0 改,勿再灌 UsefulCount) selfoptimize.Dash.RecordDistillSuccess() - // 记录蒸馏质量损失(avg_distill_loss = 1 - Overall) + // 记录蒸馏质量损失(来源①:1 - Overall;t_e5cdc25c 拆口径后走独立 LLM 桶) // 仅记录真实蒸馏:Overall > 0 表示 LLM 成功返回,过滤 fallback (Overall=0) if result.Overall > 0 { loss := 1.0 - result.Overall @@ -626,7 +626,7 @@ func NewServer() http.Handler { if loss > 1 { loss = 1 } - selfoptimize.Dash.RecordDistillLoss(loss) + selfoptimize.Dash.RecordDistillLossLLM(loss) } } @@ -1855,22 +1855,22 @@ func NewServer() http.Handler { if err == nil && report != nil { // 注入 consolidate 报告 → Dashboard if report.Merged > 0 { - // 合并启发式: 按合并数比例缩放, 但限定单次贡献 ≤ 0.05, 防止污染 avg_distill_loss + // 来源② 合并条数贡献(t_e5cdc25c 拆口径后走独立 merge 桶,不再混入蒸馏质量) loss := float64(report.Merged) * 0.002 if loss > 0.05 { loss = 0.05 } - selfoptimize.Dash.RecordDistillLoss(loss) + selfoptimize.Dash.RecordMergeRatioLoss(loss) } if report.ConflictsFound > 0 { selfoptimize.Dash.RecordConflictResolved(true) } for _, p := range report.Patterns { if len(p) > 13 && p[:13] == "quality_score" { - // 解析 quality_score=0.50 → 注入 avg_distill_loss + // 来源③ 质量回溯(cos(L0,L0'))损失(t_e5cdc25c 拆口径后走独立 backtrace 桶) var qs float64 if _, e := fmt.Sscanf(p, "quality_score=%f", &qs); e == nil { - selfoptimize.Dash.RecordDistillLoss(1.0 - qs) + selfoptimize.Dash.RecordQualityBacktraceLoss(1.0 - qs) } } } diff --git a/go/internal/selfoptimize/metrics_contract.go b/go/internal/selfoptimize/metrics_contract.go index 6c581e1..4f0a06d 100644 --- a/go/internal/selfoptimize/metrics_contract.go +++ b/go/internal/selfoptimize/metrics_contract.go @@ -24,7 +24,11 @@ // gap_closure_rate | 率 | 0..1 | ClosedGaps / TotalGaps(缺口个数口径) | TotalGaps<20 或 比率越界>1 | TotalGaps>=20 // cascade_fix_rate | 率 | 0..1 | CascadeFixedTotal / TotalFixes(/feedback/correct)| TotalFixes<20 | TotalFixes>=20 // auto_resolve_rate | 率 | 0..1 | AutoResolvedConflicts / TotalConflicts | TotalConflicts<20 | TotalConflicts>=20 -// avg_distill_loss | 均值 | 0..1 | Σ(1-overall) / DistillLossCount(overall=5D加权分)| DistillLossCount<20 | DistillLossCount>=20 +// avg_distill_loss | 均值 | 0..1 | [DEPRECATED] 三来源混算(见下三拆分指标),语义不可信 | DistillLossCount<20 | DistillLossCount>=20(保留仅向后兼容) +// distill_loss_llm | 均值 | 0..1 | 来源① Σ(1-Overall)/LLMCount(真·蒸馏质量损失) | LLMCount<20 | LLMCount>=20 +// merge_ratio_loss | 均值 | 0..1 | 来源② Σ(min(Merged×0.002,0.05))/MergeCount(合并规模)| MergeCount<20 | MergeCount>=20 +// quality_backtrace_loss | 均值 | 0..1 | 来源③ Σ(1-cos(L0,L0'))/QBCount(质量回溯损失) | QBCount<20 | QBCount>=20 +// avg_distill_loss_deprecated| 标记 | 0/1 | 恒 1:表示 avg_distill_loss 是 deprecated 混算桶(消费方读值前必看)| 无 | n/a // deprecated_per_day | 计数 | 条/天 | 当天 /feedback/deprecate 调用数(跨天归零) | 无(计数不为 null) | n/a(0 是合法值) // deprecated_total | 计数 | 条 | 累计 /feedback/deprecate 调用数(跨重启存活) | 无(计数不为 null) | n/a(0=链路未接线,见下) // distill_success_count | 计数 | 条 | 蒸馏成功条数 | 无 | n/a diff --git a/go/internal/selfoptimize/metrics_contract_test.go b/go/internal/selfoptimize/metrics_contract_test.go index 9ac4dfe..9379050 100644 --- a/go/internal/selfoptimize/metrics_contract_test.go +++ b/go/internal/selfoptimize/metrics_contract_test.go @@ -7,6 +7,7 @@ package selfoptimize import ( + "math" "path/filepath" "testing" ) @@ -42,17 +43,31 @@ func TestContract_RateKeysHaveValidAndSamples(t *testing.T) { if _, ok := m["avg_distill_loss_valid"]; !ok { t.Fatal("avg_distill_loss 缺 _valid(均值有分母,必须给有效位)") } + // t_e5cdc25c 口径拆分三来源:各自带 _valid / _samples,且 avg_distill_loss 标注 deprecated。 + for _, k := range []string{"distill_loss_llm", "merge_ratio_loss", "quality_backtrace_loss"} { + if _, ok := m[k]; !ok { + t.Fatalf("拆分指标 %s 缺失(t_e5cdc25c:三来源必须独立可复算)", k) + } + for _, suffix := range []string{"_valid", "_samples"} { + if _, ok := m[k+suffix]; !ok { + t.Fatalf("拆分指标 %s 缺姊妹键 %s%s(每个均值必须可自证分母)", k, k, suffix) + } + } + } + if m["avg_distill_loss_deprecated"] != 1 { + t.Fatal("avg_distill_loss 必须显式标注 deprecated=1(混算值已不可信)") + } } func TestContract_ZeroSampleIsMinusOneNotZero(t *testing.T) { d := freshDash(t) m := d.Metrics() - for _, k := range []string{"recall_usefulness_rate", "gap_closure_rate", "cascade_fix_rate", "auto_resolve_rate", "avg_distill_loss"} { + for _, k := range []string{"recall_usefulness_rate", "gap_closure_rate", "cascade_fix_rate", "auto_resolve_rate", "avg_distill_loss", "distill_loss_llm", "merge_ratio_loss", "quality_backtrace_loss"} { if m[k] != metricNA { t.Fatalf("%s 零样本应=%v,实得 %v(契约 B:无数据必须是哨兵 -1,禁止假 0)", k, metricNA, m[k]) } } - for _, k := range []string{"recall_usefulness_valid", "gap_closure_valid", "cascade_fix_valid", "auto_resolve_valid", "avg_distill_loss_valid"} { + for _, k := range []string{"recall_usefulness_valid", "gap_closure_valid", "cascade_fix_valid", "auto_resolve_valid", "avg_distill_loss_valid", "distill_loss_llm_valid", "merge_ratio_loss_valid", "quality_backtrace_loss_valid"} { if m[k] != 0 { t.Fatalf("%s 零样本应=0,实得 %v(契约 B:值与 valid 必须同步)", k, m[k]) } @@ -74,3 +89,53 @@ func TestContract_ThresholdsPinned(t *testing.T) { t.Fatalf("面板可解读门应为 20(契约 C),实为 %d —— 改门槛必须同步 metrics_contract.go 语义表", minRateSamples) } } + +// TestContract_DistillLossSplitThreeSources(t_e5cdc25c): +// 三个来源各自灌入独立计数器后,每个指标的 sum/count 必须与「该来源单独复算」一致, +// 且三者互不串桶(LLM 的 loss 不污染 merge 桶,反之亦然)。 +func TestContract_DistillLossSplitThreeSources(t *testing.T) { + d := freshDash(t) + // 来源① LLM:灌 25 个样本(≥ 样本门 20),均值已知。 + llmVals := []float64{0.2, 0.4, 0.1, 0.3, 0.5, 0.25, 0.45, 0.15, 0.35, 0.55, + 0.2, 0.4, 0.1, 0.3, 0.5, 0.25, 0.45, 0.15, 0.35, 0.55, 0.2, 0.4, 0.1, 0.3, 0.5} + for _, v := range llmVals { + d.RecordDistillLossLLM(v) + } + // 来源② 合并:灌 25 个固定 0.05。 + for i := 0; i < 25; i++ { + d.RecordMergeRatioLoss(0.05) + } + // 来源③ 回溯:灌 25 个固定 0.1。 + for i := 0; i < 25; i++ { + d.RecordQualityBacktraceLoss(0.1) + } + + m := d.Metrics() + + // 复算:LLM 均值 = Σ/25 + wantLLM := 0.0 + for _, v := range llmVals { + wantLLM += v + } + wantLLM = math.Round(wantLLM/25*1000) / 1000 + if m["distill_loss_llm"] != wantLLM { + t.Fatalf("distill_loss_llm = %v,期望复算值 %v(Σ/25)", m["distill_loss_llm"], wantLLM) + } + if m["distill_loss_llm_samples"] != 25 || m["distill_loss_llm_valid"] != 1 { + t.Fatalf("distill_loss_llm samples/valid 应为 25/1,实得 %v/%v", m["distill_loss_llm_samples"], m["distill_loss_llm_valid"]) + } + if m["merge_ratio_loss"] != 0.05 || m["merge_ratio_loss_samples"] != 25 || m["merge_ratio_loss_valid"] != 1 { + t.Fatalf("merge_ratio_loss 应为 0.05/25/1,实得 %v/%v/%v", m["merge_ratio_loss"], m["merge_ratio_loss_samples"], m["merge_ratio_loss_valid"]) + } + if m["quality_backtrace_loss"] != 0.1 || m["quality_backtrace_loss_samples"] != 25 || m["quality_backtrace_loss_valid"] != 1 { + t.Fatalf("quality_backtrace_loss 应为 0.1/25/1,实得 %v/%v/%v", m["quality_backtrace_loss"], m["quality_backtrace_loss_samples"], m["quality_backtrace_loss_valid"]) + } + + // 互不串桶:avg_distill_loss(deprecated 桶)本测试没调 RecordDistillLoss ⇒ 应零样本 NA。 + if m["avg_distill_loss"] != metricNA || m["avg_distill_loss_samples"] != 0 { + t.Fatalf("deprecated avg_distill_loss 不应被拆分来源污染,实得 val=%v samples=%v", m["avg_distill_loss"], m["avg_distill_loss_samples"]) + } + if m["avg_distill_loss_deprecated"] != 1 { + t.Fatalf("avg_distill_loss_deprecated 标记应为 1") + } +} diff --git a/go/internal/selfoptimize/selfoptimize.go b/go/internal/selfoptimize/selfoptimize.go index d8f56e5..3ece5f1 100644 --- a/go/internal/selfoptimize/selfoptimize.go +++ b/go/internal/selfoptimize/selfoptimize.go @@ -61,6 +61,14 @@ type dashboardState struct { DeprecatedDate string `json:"deprecated_date"` // deprecated_today 的归属日期(跨天归零用) DistillLossSum float64 `json:"distill_loss_sum"` DistillLossCount int `json:"distill_loss_count"` + // 口径拆分三来源(t_e5cdc25c):avg_distill_loss 曾把三种互不相关的量混进一个平均, + // 拆成独立计数器后 avg_distill_loss 降级为 deprecated 混算值(仅向后兼容,禁再当质量看)。 + DistillLossLLMSum float64 `json:"distill_loss_llm_sum"` + DistillLossLLMCount int `json:"distill_loss_llm_count"` + MergeRatioLossSum float64 `json:"merge_ratio_loss_sum"` + MergeRatioLossCount int `json:"merge_ratio_loss_count"` + QualityBacktraceLossSum float64 `json:"quality_backtrace_loss_sum"` + QualityBacktraceLossCount int `json:"quality_backtrace_loss_count"` DistillSuccessCount int `json:"distill_success_count"` AutoResolvedConflicts int `json:"auto_resolved_conflicts"` TotalConflicts int `json:"total_conflicts"` @@ -111,6 +119,12 @@ func loadDashboardLocal() error { } Dash.DistillLossSum = st.DistillLossSum Dash.DistillLossCount = st.DistillLossCount + Dash.DistillLossLLMSum = st.DistillLossLLMSum + Dash.DistillLossLLMCount = st.DistillLossLLMCount + Dash.MergeRatioLossSum = st.MergeRatioLossSum + Dash.MergeRatioLossCount = st.MergeRatioLossCount + Dash.QualityBacktraceLossSum = st.QualityBacktraceLossSum + Dash.QualityBacktraceLossCount = st.QualityBacktraceLossCount Dash.DistillSuccessCount = st.DistillSuccessCount Dash.AutoResolvedConflicts = st.AutoResolvedConflicts Dash.TotalConflicts = st.TotalConflicts @@ -141,6 +155,12 @@ func (d *Dashboard) persistLocalLocked() { DeprecatedDate: todayStr(), DistillLossSum: d.DistillLossSum, DistillLossCount: d.DistillLossCount, + DistillLossLLMSum: d.DistillLossLLMSum, + DistillLossLLMCount: d.DistillLossLLMCount, + MergeRatioLossSum: d.MergeRatioLossSum, + MergeRatioLossCount: d.MergeRatioLossCount, + QualityBacktraceLossSum: d.QualityBacktraceLossSum, + QualityBacktraceLossCount: d.QualityBacktraceLossCount, DistillSuccessCount: d.DistillSuccessCount, AutoResolvedConflicts: d.AutoResolvedConflicts, TotalConflicts: d.TotalConflicts, @@ -180,6 +200,13 @@ func (d *Dashboard) persist() { _ = redisPersistence.HSet(dashboardRedisKey, "deprecated_date", d.DeprecatedDate) _ = redisPersistence.HSet(dashboardRedisKey, "distill_loss_sum", strconv.FormatFloat(d.DistillLossSum, 'f', 4, 64)) _ = redisPersistence.HSet(dashboardRedisKey, "distill_loss_count", strconv.Itoa(d.DistillLossCount)) + // 口径拆分三来源(t_e5cdc25c) + _ = redisPersistence.HSet(dashboardRedisKey, "distill_loss_llm_sum", strconv.FormatFloat(d.DistillLossLLMSum, 'f', 6, 64)) + _ = redisPersistence.HSet(dashboardRedisKey, "distill_loss_llm_count", strconv.Itoa(d.DistillLossLLMCount)) + _ = redisPersistence.HSet(dashboardRedisKey, "merge_ratio_loss_sum", strconv.FormatFloat(d.MergeRatioLossSum, 'f', 6, 64)) + _ = redisPersistence.HSet(dashboardRedisKey, "merge_ratio_loss_count", strconv.Itoa(d.MergeRatioLossCount)) + _ = redisPersistence.HSet(dashboardRedisKey, "quality_backtrace_loss_sum", strconv.FormatFloat(d.QualityBacktraceLossSum, 'f', 6, 64)) + _ = redisPersistence.HSet(dashboardRedisKey, "quality_backtrace_loss_count", strconv.Itoa(d.QualityBacktraceLossCount)) _ = redisPersistence.HSet(dashboardRedisKey, "distill_success_count", strconv.Itoa(d.DistillSuccessCount)) _ = redisPersistence.HSet(dashboardRedisKey, "auto_resolved_conflicts", strconv.Itoa(d.AutoResolvedConflicts)) _ = redisPersistence.HSet(dashboardRedisKey, "total_conflicts", strconv.Itoa(d.TotalConflicts)) @@ -243,6 +270,25 @@ func loadDashboard() error { if v, err := strconv.Atoi(data["distill_loss_count"]); err == nil { Dash.DistillLossCount = v } + // 口径拆分三来源(t_e5cdc25c):缺则保持零值,兼容旧 Redis 数据 + if v, err := strconv.ParseFloat(data["distill_loss_llm_sum"], 64); err == nil { + Dash.DistillLossLLMSum = v + } + if v, err := strconv.Atoi(data["distill_loss_llm_count"]); err == nil { + Dash.DistillLossLLMCount = v + } + if v, err := strconv.ParseFloat(data["merge_ratio_loss_sum"], 64); err == nil { + Dash.MergeRatioLossSum = v + } + if v, err := strconv.Atoi(data["merge_ratio_loss_count"]); err == nil { + Dash.MergeRatioLossCount = v + } + if v, err := strconv.ParseFloat(data["quality_backtrace_loss_sum"], 64); err == nil { + Dash.QualityBacktraceLossSum = v + } + if v, err := strconv.Atoi(data["quality_backtrace_loss_count"]); err == nil { + Dash.QualityBacktraceLossCount = v + } if v, err := strconv.Atoi(data["distill_success_count"]); err == nil { Dash.DistillSuccessCount = v } @@ -281,6 +327,14 @@ type Dashboard struct { DeprecatedDate string `json:"deprecated_date"` // t_4bbddae6:deprecated_today 的归属日期(跨天归零用) DistillLossSum float64 `json:"distill_loss_sum"` DistillLossCount int `json:"distill_loss_count"` + // 口径拆分三来源(t_e5cdc25c)。三者互不相关,拆开各自可复算;avg_distill_loss 保留为 + // deprecated 混算(= 三来源加权的单一平均),不再作为「记性有多好」的尺子(见分布表)。 + DistillLossLLMSum float64 `json:"distill_loss_llm_sum"` + DistillLossLLMCount int `json:"distill_loss_llm_count"` + MergeRatioLossSum float64 `json:"merge_ratio_loss_sum"` + MergeRatioLossCount int `json:"merge_ratio_loss_count"` + QualityBacktraceLossSum float64 `json:"quality_backtrace_loss_sum"` + QualityBacktraceLossCount int `json:"quality_backtrace_loss_count"` DistillSuccessCount int `json:"distill_success_count"` AutoResolvedConflicts int `json:"auto_resolved_conflicts"` TotalConflicts int `json:"total_conflicts"` @@ -377,12 +431,30 @@ func (d *Dashboard) Metrics() map[string]float64 { avgLoss = math.Round(d.DistillLossSum/float64(d.DistillLossCount)*100) / 100 } + // 口径拆分三来源(t_e5cdc25c)。每个均值独立可复算(sum/count),各带 _count/_valid。 + // avg_distill_loss 降级为 deprecated:它是三来源混算的单一平均,语义已不可信。 + llmLoss := metricNA + if d.DistillLossLLMCount > 0 { + llmLoss = math.Round(d.DistillLossLLMSum/float64(d.DistillLossLLMCount)*1000) / 1000 + } + mergeLoss := metricNA + if d.MergeRatioLossCount > 0 { + mergeLoss = math.Round(d.MergeRatioLossSum/float64(d.MergeRatioLossCount)*1000) / 1000 + } + qbLoss := metricNA + if d.QualityBacktraceLossCount > 0 { + qbLoss = math.Round(d.QualityBacktraceLossSum/float64(d.QualityBacktraceLossCount)*1000) / 1000 + } + // 最小样本门:samples < minRateSamples ⇒ 值 = -1 且 valid = 0,不参与面板解读。 usefulVal, usefulValid := usefulRate, b2f(usefulSamples >= minRateSamples && d.NotUsefulCount > 0) gapVal, gapValid := gapRate, b2f(d.TotalGaps >= minRateSamples) cascadeVal, cascadeValid := cascadeRate, b2f(d.TotalFixes >= minRateSamples) autoVal, autoValid := autoRate, b2f(d.TotalConflicts >= minRateSamples) lossVal, lossValid := avgLoss, b2f(d.DistillLossCount >= minRateSamples) + llmLossVal, llmLossValid := llmLoss, b2f(d.DistillLossLLMCount >= minRateSamples) + mergeLossVal, mergeLossValid := mergeLoss, b2f(d.MergeRatioLossCount >= minRateSamples) + qbLossVal, qbLossValid := qbLoss, b2f(d.QualityBacktraceLossCount >= minRateSamples) if usefulSamples < minRateSamples || d.NotUsefulCount == 0 { usefulVal = metricNA } @@ -400,6 +472,15 @@ func (d *Dashboard) Metrics() map[string]float64 { if d.DistillLossCount < minRateSamples { lossVal = metricNA } + if d.DistillLossLLMCount < minRateSamples { + llmLossVal = metricNA + } + if d.MergeRatioLossCount < minRateSamples { + mergeLossVal = metricNA + } + if d.QualityBacktraceLossCount < minRateSamples { + qbLossVal = metricNA + } return map[string]float64{ "recall_usefulness_rate": usefulVal, @@ -428,6 +509,23 @@ func (d *Dashboard) Metrics() map[string]float64 { "avg_distill_loss": lossVal, "avg_distill_loss_samples": float64(d.DistillLossCount), "avg_distill_loss_valid": lossValid, + // avg_distill_loss 已 deprecated(t_e5cdc25c):三来源混算,语义不可信。 + // 显式标记位:valid=1 表示该 deprecated 键仍在输出(向后兼容),值本身禁再当质量看。 + "avg_distill_loss_deprecated": 1, + + // 口径拆分三来源(t_e5cdc25c)—— 各自独立可复算,取代 avg_distill_loss 作为可信尺子。 + "distill_loss_llm": llmLossVal, + "distill_loss_llm_samples": float64(d.DistillLossLLMCount), + "distill_loss_llm_valid": llmLossValid, + "distill_loss_llm_sum": d.DistillLossLLMSum, + "merge_ratio_loss": mergeLossVal, + "merge_ratio_loss_samples": float64(d.MergeRatioLossCount), + "merge_ratio_loss_valid": mergeLossValid, + "merge_ratio_loss_sum": d.MergeRatioLossSum, + "quality_backtrace_loss": qbLossVal, + "quality_backtrace_loss_samples": float64(d.QualityBacktraceLossCount), + "quality_backtrace_loss_valid": qbLossValid, + "quality_backtrace_loss_sum": d.QualityBacktraceLossSum, } } @@ -1014,7 +1112,13 @@ func (d *Dashboard) RecordConflictResolved(auto bool) { d.persist() } -// RecordDistillLoss 记录蒸馏质量损失 +// RecordDistillLoss 记录蒸馏质量损失([DEPRECATED] t_e5cdc25c)。 +// +// 历史上三来源(LLM 蒸馏 1-Overall / 合并条数 / 质量回溯)都灌进这一个计数器, +// 导致 avg_distill_loss 无单一语义。现拆为三个独立 Record 方法(见下),本方法仅作 +// 向后兼容保留:只更新 deprecated 的 avg_distill_loss 混算桶,**不再**作为任何 +// 可信指标的来源。调用点应迁移到 RecordDistillLossLLM / RecordMergeRatioLoss / +// RecordQualityBacktraceLoss。 func (d *Dashboard) RecordDistillLoss(loss float64) { d.mu.Lock() defer d.mu.Unlock() @@ -1023,6 +1127,35 @@ func (d *Dashboard) RecordDistillLoss(loss float64) { d.persist() } +// RecordDistillLossLLM 记录「真·蒸馏质量损失」来源(来源①:1.0 - distill.Overall)。 +// Overall 为 distill engine 五维加权分,scoreMinOverall 默认 0.30。 +func (d *Dashboard) RecordDistillLossLLM(loss float64) { + d.mu.Lock() + defer d.mu.Unlock() + d.DistillLossLLMSum += loss + d.DistillLossLLMCount++ + d.persist() +} + +// RecordMergeRatioLoss 记录「合并条数贡献」来源(来源②:min(Merged×0.002, 0.05))。 +// 与蒸馏质量无关,仅反映 consolidate 合并规模,不应并入质量损失。 +func (d *Dashboard) RecordMergeRatioLoss(loss float64) { + d.mu.Lock() + defer d.mu.Unlock() + d.MergeRatioLossSum += loss + d.MergeRatioLossCount++ + d.persist() +} + +// RecordQualityBacktraceLoss 记录「质量回溯损失」来源(来源③:1.0 - cos(L0,L0'))。 +func (d *Dashboard) RecordQualityBacktraceLoss(loss float64) { + d.mu.Lock() + defer d.mu.Unlock() + d.QualityBacktraceLossSum += loss + d.QualityBacktraceLossCount++ + d.persist() +} + func (pg *PrefetchGraph) GetPrefetch(query string) []string { pg.mu.RLock() defer pg.mu.RUnlock() diff --git a/rust/src/main.rs b/rust/src/main.rs index 7b541a6..321f655 100644 --- a/rust/src/main.rs +++ b/rust/src/main.rs @@ -1068,6 +1068,9 @@ fn run_consolidation(args: &Args, lancedb: &LanceDBOps) -> ConsolidationReport { // Step 5: 生成报告 reclaim("Step 4 后(质量回溯)"); let dashboard = DashboardSnapshot { + // ⚠️ 全部硬编码占位值(t_e5cdc25c 已标 stub):禁当真值。真实来源在 Go + // selfoptimize.Dash.Metrics(),本 Rust 侧不直连该计数器。dashboard_stub=true + // 使任何消费 JSON 报告者能看到这是占位数据。 recall_useful_rate: 0.85, recall_hit_rate: 0.90, gap_closure_rate: 0.75, @@ -1075,6 +1078,7 @@ fn run_consolidation(args: &Args, lancedb: &LanceDBOps) -> ConsolidationReport { deprecation_rate: 3.0, distill_loss_rate: quality_score.max(0.0), auto_resolve_rate: 0.70, + dashboard_stub: true, }; let gen = ReportGenerator::new(); diff --git a/rust/src/report.rs b/rust/src/report.rs index 4a6a4f8..2f5377e 100644 --- a/rust/src/report.rs +++ b/rust/src/report.rs @@ -47,13 +47,25 @@ pub struct PruneStats { } /// 仪表盘快照 +/// +/// ⚠️ [STUB](t_e5cdc25c):6 个比率/计数指标当前**全部硬编码占位值** +/// (main.rs:1070),不是真实来源。消费方读任何字段前必须先看 dashboard_stub: +/// - 5 个率值 (0.85/0.90/0.75/0.60/0.70) 与 deprecation_rate(3.0) 是**假常量**, +/// 真实值应从 Go selfoptimize.Dash.Metrics() 取(本 Rust 侧不直连该计数器)。 +/// - distill_loss_rate 直接复用 quality_score(回溯 cos 分)当「损失率」,语义反 +/// (质量分越高损失越大?),且 quality_score 失败时=0 ⇒ 损失恒 0,退化判定失真。 +/// - previous_metrics 恒 None(new() 不设值)⇒ detect_degradation 永远返回空, +/// 该退化判定链路目前是**死路**(不会因硬编码值误告警,但也不会真触发)。 +/// 影响面:本结构仅流入 Rust CLI 的 stdout JSON 报告(人读)与 IPC 发给 Go 的 +/// ConsolidationReport;Go 侧 consolidate::Result **不读** dashboard/degradation 字段, +/// 故 6 个硬编码值**不流入** Go 自优化/调参/告警。详见 t_e5cdc25c 交付报告。 #[derive(Debug, Clone, Serialize, Deserialize)] pub struct DashboardSnapshot { - /// 召回有用率 + /// 召回有用率 [STUB 硬编码 0.85] pub recall_useful_rate: f64, - /// 召回命中率 + /// 召回命中率 [STUB 硬编码 0.90] pub recall_hit_rate: f64, - /// 缺口闭环率 + /// 缺口闭环率 [STUB 硬编码 0.75] pub gap_closure_rate: f64, /// 修正传播率 pub cascade_propagation_rate: f64, @@ -63,6 +75,15 @@ pub struct DashboardSnapshot { pub distill_loss_rate: f64, /// 冲突自动裁决率 pub auto_resolve_rate: f64, + /// [STUB 标记位](t_e5cdc25c):true = 本快照全部指标为硬编码占位值,禁当真值用。 + /// 消费方(报告/告警/调参)读前必看此位;false 才代表已接真实来源。 + #[serde(default = "default_dashboard_stub_true")] + pub dashboard_stub: bool, +} + +// 默认 true:旧数据/缺省反序列化时保守标为 stub,避免把未知来源的快照当真实指标用。 +fn default_dashboard_stub_true() -> bool { + true } /// 退化告警