半導體跨站測試覆蓋優化 SOPSemiconductor Cross-Stage Test Coverage Optimization SOP

0目的與範圍Purpose & Scope

當同一顆 die 在 WT（CP）與 FT 兩站、甚至在 RT／HT／LT 多個溫度 corner 下被重複測試相同或高度相關的項目時，若能證明「後站不會抓到前站漏抓的 fail」，即可將該測項自後站移除或降頻抽測，達成 Test Time Reduction（TTR）。 When the same die is tested for identical or highly correlated items at both WT (CP) and FT — often across multiple temperature corners (RT/HT/LT) — proving that "the downstream stage never catches a fail the upstream stage missed" lets that test be removed or sampled downstream, achieving Test Time Reduction (TTR).

本文件的目的是把這個判斷「模板化、流程化」，避免僅憑幾百顆樣本「沒看到 fail」就下結論的常見統計誤區。核心主張：只看 Bin（Pass／Fail）層級不夠，必須下探到 Parameter（實際量測值）層級，才能看見「兩站都 Pass，但 margin 正在被吃掉」的隱性風險，也才能用遠小於 Bin-level 所需的樣本數，取得具科學依據的結論。 This document turns that judgment into a repeatable template and process, avoiding the common statistical fallacy of concluding "safe to remove" from a few hundred zero-fail samples. Core argument: Bin-level (Pass/Fail) data alone is not enough — analysis must go down to the Parameter level (actual measured values) to expose the hidden risk of "both stages Pass, but margin is eroding," and to reach a scientifically defensible conclusion with a far smaller sample size than Bin-level analysis requires.

1名詞定義Definitions

2流程總覽Process Overview

主流程分七個 Phase（Phase 0–6），前一 Phase 的輸出是後一 Phase 的輸入。Phase 0 是不需要資料就能先判斷的「硬性守則」（規格強制項目、零覆蓋保護、量測系統有效性），先擋掉不該評估的候選；Phase 3 則是本 SOP 的統計核心，把 Bin-level、Parameter-level 與（適用時的）溫度 Corner 線性驗證拆成並行的證據線。額外的 Phase 7（虛線／Optional）不是循序步驟，而是客訴事件觸發時才啟動的查證程序。 The main process runs across seven phases (Phase 0–6), each phase's output feeding the next. Phase 0 applies hard rules that need no data at all (spec-mandated items, the zero-coverage guard, measurement-system validity) to screen out candidates that should never be evaluated for removal; Phase 3 is the statistical core, splitting Bin-level, Parameter-level, and (where applicable) temperature-corner linearity validation into parallel lines of evidence. The extra Phase 7 (dashed / Optional) isn't a sequential step — it's a verification procedure that only runs when a customer complaint triggers it.

3Phase 0 — 前置強制檢查Phase 0 — Hard Gates

在投入任何資料分析之前，先用三個不需要統計證據的「硬性守則」篩掉不該評估移除的候選 —— 這些守則不會被任何統計結果推翻，任一項不通過即直接保留。 Before investing in any data analysis, screen out candidates that should never be evaluated for removal using three hard rules that need no statistical evidence at all — no statistical result can override these; failing any one of them means the item is kept, full stop.

3.1 規格／驗證標準強制項目3.1 Spec / Qualification-Mandated Items

先查客戶 datasheet／spec 書、PPAP 承諾項目，或適用的驗證標準（如 AEC-Q100、JEDEC）測項清單 —— 若該測項被這些文件明確、逐項寫死要求執行，則無論 Incremental Catch 或相關性統計結果多乾淨，都不可移除，因為這已經是契約或驗證等級的承諾，不是工程風險判斷可以覆蓋的範圍。 Check the customer datasheet/spec, PPAP commitments, or the applicable qualification standard's test list (e.g., AEC-Q100, JEDEC) first. If the item is explicitly, individually mandated by one of these documents, it cannot be removed no matter how clean the Incremental Catch or correlation statistics look — that is a contractual or qualification-level commitment, outside the scope of an engineering risk judgment.

3.2 零覆蓋守則3.2 Zero-Coverage Guard

Redundancy 分析的前提是移除的必須是重複的覆蓋，而不是唯一的覆蓋。無論統計證據多強，決策都不可以讓任一 spec 參數在整個流程（所有站別 × 所有溫度 corner）中變成完全沒有測試點覆蓋。 Redundancy analysis only makes sense when what's being removed is duplicate coverage, never the only coverage. No matter how strong the statistical evidence, a decision may never leave any spec parameter with zero test coverage across the entire flow (every stage × every temperature corner).

3.3 溫度 Corner 移除與線性假設3.3 Temperature-Corner Removal & the Linearity Assumption

跨站（WT↔FT）冗餬與跨溫度 corner（RT/HT/LT）冗餘是兩個不同的問題，不能互相替代。若候選是「移除某個溫度 corner 的測試」（例如認為 RT 已涵蓋、HT／LT 可省略），必須額外驗證該參數在溫度間的關係，不能只憑「數值看起來隨溫度平滑變化」就假設線性成立 —— 部分電性參數（如漏電流）隨溫度呈指數關係而非線性，兩點內插／外插在這類參數上會嚴重低估風險。這條證據線在 Phase 3（§6.3）中與 Bin-level、Parameter-level 並列，僅在候選涉及溫度 corner 移除時適用。 Cross-stage (WT↔FT) redundancy and cross-temperature-corner (RT/HT/LT) redundancy are two different questions and cannot substitute for each other. If the candidate is "remove testing at one temperature corner" (e.g., assuming RT already covers it, so HT/LT can be skipped), the temperature relationship must be separately validated — never assume linearity just because values "look smooth across temperature." Some electrical parameters (e.g., leakage current) follow an exponential, not linear, relationship with temperature, and two-point interpolation/extrapolation badly underestimates risk for those. This evidence line sits alongside Bin-level and Parameter-level in Phase 3 (§6.3), and applies only when the candidate involves removing a temperature corner.

3.4 量測系統有效性（Contact／Repeatability／儀器極限）3.4 Measurement System Validity (Contact / Repeatability / Instrument Limits)

在信任任何相關係數或 Cpk 之前，必須先確認量測系統本身是有效的 —— 否則相關性可能只是反映「兩站共享同樣的量測雜訊」，或者反過來，真正的實體冗餘被雜訊蓋過而誤判為不相關。三個常見盲點： Before trusting any correlation coefficient or Cpk, first confirm the measurement system itself is valid — otherwise a correlation might just reflect "both stages sharing the same measurement noise," or conversely, real physical redundancy gets buried under noise and wrongly read as unrelated. Three common blind spots:

• Contact 品質：探針卡／socket 接觸不良（接觸電阻異常、需要 retouch）會在資料中製造假的變異。務必監控 retest rate、contact resistance 等指標，異常批次要先排除或重測，不能直接拿來做冗餘分析。 Contact quality: poor probe-card/socket contact (elevated contact resistance, needing a retouch) manufactures fake variance in the data. Monitor retest rate, contact resistance, and similar indicators; exclude or retest anomalous lots before using them for redundancy analysis.
• 測試程式 Repeatability：同一顆 die 在同一站、同一條件下重複測試，量測值的變異（Gauge R&R 中的 repeatability 分量）若太大，會直接壓低任何跨站相關係數的上限 —— 這種情況下低 r 不代表「兩站不相關」，可能只是「量測本身不夠穩定」。 Test-program repeatability: retest the same die at the same station under the same conditions — if the measurement variance (the repeatability component of a Gauge R&R study) is too large, it caps how high any cross-stage correlation can possibly read. A low r here doesn't mean "the two stages are unrelated" — it may just mean "the measurement itself isn't stable enough."
• 測試項特性與儀器極限：部分參數天生量測雜訊大（如極低漏電流量測接近儀器解析度極限），或該儀器的量測範圍／解析度本身就無法穩定分辨 die 間的真實差異 —— 這種情況下，相關性分析的結論本來就不可靠，不該用來支持移除決策。 Test-item characteristics and instrument limits: some parameters are inherently noisy to measure (e.g., ultra-low leakage current near the instrument's resolution floor), or the instrument's range/resolution simply can't reliably distinguish real die-to-die differences. In that case the correlation analysis's conclusion is unreliable to begin with and shouldn't be used to support a removal decision.

上面的 %R&R 門檻要「信任」，本身也需要足夠的重複測試次數才能估得準 —— 這裡常用反查 Chi-Square（reverse chi-square）的方法：母體變異 σ² 的信賴區間可以用樣本變異 s²（n 次重複量測算出）與卡方分布連結；要反過來問「要多少次重複量測，才能讓估出來的 σ 夠精準」，就是把這個關係式倒過來解 n。 Trusting the %R&R threshold above requires enough repeat measurements to estimate it precisely in the first place — a common method here is reverse (inverse) chi-square sample sizing: the confidence interval for a population variance σ² connects to the sample variance s² (from n repeat measurements) via the chi-square distribution. Asking "how many repeats do I need for a precise-enough σ estimate" means solving that relationship backward for n.

4Phase 1 — 測項盤點與配對Phase 1 — Test Inventory & Mapping

建立「測試對應矩陣」，列出 WT／FT、各溫度 corner 下的所有測項，並標註其類型。這一步決定了後面能不能做 Parameter-level 分析 —— 純 Bin-only 的測項（如某些 digital functional test）天生沒有 parameter 可看，只能走 Bin-level 途徑；量測型測項（parametric）則兩條途徑都要走。 Build a "Test Mapping Matrix" listing every test item across WT/FT and all temperature corners, tagged by type. This step determines whether Parameter-level analysis is even possible downstream — purely Bin-only items (e.g., certain digital functional tests) have no parameter to examine and can only follow the Bin-level path; parametric items must follow both paths.

5Phase 2 — 資料萃取與 Die-level 追溯Phase 2 — Data Extraction & Die-Level Traceability

從 STDF 的 PTR（Parametric Test Record，量測值）與 FTR（Functional Test Record，pass/fail）萃取資料，並透過 Wafer ID + X,Y 座標（WT 端）對應到封裝序號／Lot-Unit ID（FT 端），做到 die-level 一對一配對，而不是只在 lot 層級做良率比較。 Extract data from STDF's PTR (Parametric Test Record, measured values) and FTR (Functional Test Record, pass/fail), then map Wafer ID + X,Y coordinates (WT side) to package serial number / Lot-Unit ID (FT side) to achieve die-level one-to-one pairing, rather than a coarse lot-level yield comparison.

• 沒有 die-level 追溯，就只能做 lot-level 的粗略良率相關性，無法算出「這一顆」在 WT 端的 margin 是否預示了它在 FT 端會出問題。 Without die-level traceability, you're limited to a rough lot-level yield correlation — you cannot determine whether "this specific die's" WT margin predicted an FT problem.
• 建議資料涵蓋期間至少橫跨數個 lot、不同時段（避免單一批次的偶發現象被誤判為通則），並記錄晶圓廠／測試機台／探針卡等潛在共變因子。 Data should span at least several lots across different time periods (so a one-off event in a single batch isn't mistaken for a general pattern), and should record potential confounders such as fab, tester, and probe card.
• Parametric 測項務必保留原始量測值（非僅 Pass/Fail flag），這是 Phase 3 做 Parameter-level 分析的前提。 Parametric items must retain the raw measured value (not just a Pass/Fail flag) — this is a prerequisite for the Phase 3 Parameter-level analysis.

6Phase 3 — 統計分析方法Phase 3 — Statistical Analysis Methods

這是整份 SOP 的核心：Bin-level 分析回答「有沒有看到 fail」，Parameter-level 分析回答「margin 有沒有在流失」，溫度 Corner 線性分析回答「未測的溫度點是否仍被可靠預測」。三者互補，而非其中一種可以取代另一種；溫度 Corner 線性分析僅在候選涉及移除溫度 corner 時才需要（見 Phase 0 §3.3）。 This is the core of the SOP: Bin-level analysis answers "did we see a fail?"; Parameter-level analysis answers "is margin eroding?"; temperature-corner linearity analysis answers "is the untested temperature point still reliably predicted?" All three are complementary — none replaces another; the temperature-corner line is only required when the candidate involves removing a temperature corner (see Phase 0 §3.3).

6.1 Bin-level（屬性資料）分析6.1 Bin-Level (Attribute Data) Analysis

對候選 pair 統計 Incremental Catch：WT Pass、FT 該測項 Fail 的顆數。若在觀察樣本中 Incremental Catch = 0，只能說「在這個樣本數下沒觀察到」，還不能說「風險為零」—— 這正是 Bin-level 方法的統計弱點：它是二元資料（attribute data），要把「零事件」轉換成「風險上界」，需要非常大的樣本數。 For each candidate pair, count Incremental Catch: units that Pass at WT but Fail this item at FT. An observed Incremental Catch of 0 only means "not observed at this sample size" — not "zero risk." This is the statistical weakness of Bin-level methods: because the data is binary (attribute data), converting a "zero-event" observation into a risk upper bound requires a very large sample size.

6.2 Parameter-level（變數資料）分析 —— 本 SOP 的重點6.2 Parameter-Level (Variables Data) Analysis — The Focus of This SOP

Bin-level 只看「有沒有跨過 spec limit」，看不到「距離 spec limit 還有多遠、這個距離是不是在縮小」。同樣是「WT、FT 都 Pass」的兩顆 die，一顆 margin 很寬、一顆貼著 FT limit，Bin-level 資料完全無法區分，但這正是後續製程漂移或環境應力（如封裝、burn-in、溫度循環）造成漏測的高風險族群。 Bin-level data only tells you whether a value crossed the spec limit, not how far from the limit it sits or whether that distance is shrinking. Two dies that both "Pass at WT and FT" can be indistinguishable at the Bin level — one with wide margin, one sitting right against the FT limit — yet the latter is exactly the high-risk population most likely to escape after downstream process drift or environmental stress (packaging, burn-in, temperature cycling).

(a) 相關性驗證 — Pearson / Fisher z(a) Correlation Validation — Pearson / Fisher z

對同一 parameter 在 WT、FT 兩站的量測值做相關係數 r 檢定，確認兩站量測確實反映同一物理量，且相關性穩定高於可接受下限 r_min（而非只看單次樣本的 r 剛好夠高）。前提是 Phase 0 §3.4 的量測系統有效性已經確認——量測雜訊過大時，算出來的 r 偏低不代表兩站真的不相關，只代表量測系統蓋過了訊號。 Test the correlation coefficient r between WT and FT measured values for the same parameter, confirming the two stages truly reflect the same physical quantity and that the correlation reliably exceeds an acceptable floor r_min (not just that a single sample's r happens to clear the bar). This assumes Phase 0 §3.4's measurement-system validity has already been confirmed — when measurement noise is too large, a low computed r doesn't mean the two stages are truly unrelated, only that the measurement system is drowning out the signal.

(b) Margin-to-limit（guardband）趨勢分析(b) Margin-to-Limit (Guardband) Trend Analysis

將每顆量測值換算成「距離 spec limit 幾個 σ」，畫出 WT margin 與 FT margin 的散佈圖與時間趨勢（如圖 6.2）。重點不是平均值夠不夠遠，而是尾端（tail）有沒有子族群持續逼近 limit —— 這類子族群常與特定站位、特定 lot、特定測試機台相關，Bin-level 的加總統計會把它們平均掉而看不見。 Convert each measured value into "distance from the spec limit in σ," then plot WT margin against FT margin and its trend over time (see Fig. 6.2). What matters isn't whether the average sits far enough away, but whether a tail sub-population keeps closing in on the limit — such sub-populations often correlate with a specific site, lot, or tester, and get averaged away and hidden in aggregate Bin-level statistics.

(c) Cpk／Ppk 站間比較(c) Cross-Stage Cpk/Ppk Comparison

(d) 變數抽樣法所需樣本數遠小於屬性抽樣(d) Variables Sampling Needs Far Fewer Samples Than Attribute Sampling

這是 Parameter-level 分析背後的統計原理：屬性資料（Pass/Fail）每顆只回傳 1 bit 資訊，要把「零事件」轉換成低 DPPM 等級的信心上界，天然需要極大樣本；變數資料（實際量測值）每顆回傳連續資訊（均值與變異數都在估計中被使用），可用統計容忍區間（statistical tolerance interval）方法，以遠小的樣本數（實務上常見 n = 30～200 等級）對母體比例做同等信心水準的宣稱。方法與 k-factor 查表請參考 ASTM E2586 / ISO 16269-6（見附錄）；k-factor 背後「用卡方分布反查所需樣本數」的推導方式，見 §3.4 的反查 Chi-Square 說明。 This is the statistical principle behind Parameter-level analysis: attribute data (Pass/Fail) carries only 1 bit per unit, so converting a zero-event observation into a low-DPPM confidence bound inherently requires a huge sample. Variables data (actual measured values) carries continuous information per unit — both mean and variance are used in the estimate — so a statistical tolerance interval method can make an equivalent-confidence population claim with a far smaller sample (commonly n = 30–200 in practice). See ASTM E2586 / ISO 16269-6 for the method and k-factor tables (Appendix); for how those k-factors trace back to reverse-solving the chi-square relationship for sample size, see §3.4's reverse chi-square note.

(e) 全矩陣相關性 — 同站跨溫度 vs 跨站同溫度(e) Full Correlation Matrix — Same-Stage Cross-Temperature vs. Cross-Stage Same-Temperature

候選配對不該只鎖定「WT vs FT、同一溫度」這一組。同一參數實際上最多有六個量測點（WT-RT／WT-HT／WT-LT／FT-RT／FT-HT／FT-LT），應該用 (a) 的方法對所有存在的配對都算一次相關係數，畫出完整的相關矩陣，原因有二： Candidate pairs shouldn't be limited to "WT vs FT at the same temperature." A given parameter has up to six measurement points (WT-RT/WT-HT/WT-LT/FT-RT/FT-HT/FT-LT); use the method in (a) to compute a correlation coefficient for every pair that actually exists, and build the full correlation matrix. Two reasons:

1. 同站跨溫度（如 WT-RT 對 WT-HT）本身可能就有很高的相關性 —— 這代表的是「溫度 corner 冗餘」而非「跨站冗餘」。這類配對即使 r 很高，也必須額外通過 §6.3 的線性／TC 模型驗證才能核准移除，因為相關係數高不代表關係是線性的，也不代表兩端物理機制相同。 Same-stage, cross-temperature pairs (e.g., WT-RT vs WT-HT) can themselves be highly correlated — that's temperature-corner redundancy, not cross-stage redundancy. Even with a high r, such pairs must still additionally pass the §6.3 linearity/TC-model validation before removal is approved, since a high correlation coefficient neither implies a linear relationship nor that both ends share the same underlying physics.
2. 跨站同溫度（如 WT-RT 對 FT-RT）才是 (a)–(d) 處理的標準案例。矩陣中相關性偏低的配對，無論同站跨溫或跨站同溫，都應直接排除移除候選，不需要進到 Phase 4。 Cross-stage, same-temperature pairs (e.g., WT-RT vs FT-RT) are the standard case that (a)–(d) already cover. Any pair in the matrix with low correlation — same-stage or cross-stage — is disqualified as a removal candidate and does not proceed to Phase 4.

(f) Outlier Detection — 個體層級異常偵測(f) Outlier Detection — Individual-Level Anomaly Screening

高相關係數、良好線性都只是「母體層級」的證據，不保證沒有個別 die 是離群值 —— 樣本數一大，少數 outlier 對整體 r 或 R² 的影響會被稀釋，異常反而被平均掉而看不見。核准移除前，應在殘差上額外跑一次個體層級的異常偵測： A high correlation coefficient and good linearity are only population-level evidence — they don't guarantee no individual die is an outlier. As sample size grows, a handful of outliers get diluted in their effect on the overall r or R², averaging the anomaly away until it's invisible. Before approving removal, run an individual-level outlier-detection pass on the residuals:

• NNR（Nearest Neighbor Residual）：把每顆 die 的量測值與晶圓上鄰近 die 的平滑預測值比較，殘差過大視為空間異常（典型的 Good-Die-in-Bad-Neighborhood 現象），能抓到單靠站間或溫度相關性看不出來的局部異常。 NNR (Nearest Neighbor Residual): compare each die's measured value against a smoothed prediction from its spatial neighbors on the wafer; a large residual flags a spatial anomaly (the classic "Good Die in Bad Neighborhood" pattern) — catching local anomalies that a stage- or temperature-correlation check alone would miss.
• DPAT（Dynamic Part Average Testing）：不用固定 spec limit，改用每個 lot／wafer 自己的統計分布（mean ± kσ）動態計算限制；即使數值仍在 spec 內，只要明顯偏離同批分布就會被標記。 DPAT (Dynamic Part Average Testing): instead of a fixed spec limit, compute limits dynamically per lot/wafer from that batch's own statistics (mean ± kσ); a die can still be flagged for sitting well outside its own batch's distribution even while inside spec.
• 雙變量／多變量異常偵測（如 Mahalanobis 距離）：單一參數看起來正常，不代表兩個相關參數的「組合」正常 —— 用 Mahalanobis 距離同時看多個維度，才能抓到只有在聯合分布上才顯形的離群點。 Bivariate/multivariate outlier detection (e.g., Mahalanobis distance): a value can look normal on each parameter individually yet be abnormal in the combination of two correlated parameters — Mahalanobis distance looks across multiple dimensions at once to catch outliers that only show up in the joint distribution.

這條檢查獨立於相關係數與線性驗證：任何一種方法標記出的 outlier 都應該讓該候選回頭複查，而不是被一個看起來很高的整體 r 值掩蓋過去。 This check is independent of the correlation coefficient and linearity validation. Any outlier flagged by any of these methods should send the candidate back for review, rather than being smoothed away by a reassuringly high aggregate r.

(g) 主動收緊前站 Limit（Guardbanding）作為 Enabler(g) Actively Tightening the Upstream Limit as an Enabler

有時候相關性不是天生就「夠高」，但仍然可以透過收緊前站（如 WT）的測試 limit，主動換取移除後站測試的空間。原理：用迴歸模型從前站量測值推算後站預期值，再加上預測區間（prediction interval）的安全緩衝；只要收緊後的前站 limit 距離這個緩衝夠遠，通過收緊後前站測試的 die，其後站預測值就能以目標信心水準保證落在後站 spec 之內。 Sometimes the raw correlation isn't naturally "high enough" on its own, but removal headroom can still be actively created by tightening the upstream (e.g., WT) test limit. The principle: use the regression model to project the expected downstream value from the upstream measurement, then add a safety buffer sized by the prediction interval. As long as the tightened upstream limit sits far enough beyond that buffer, any die passing the tightened upstream test is guaranteed — at the target confidence level — to land inside the downstream spec.

這個做法不是免費的：收緊 limit 會讓原本在舊 limit 內、新 limit 外的 die 在前站就被判退，即使它們原本在後站測試也會 Pass。決策時必須把「犧牲的前站良率成本」跟「移除後站測試省下的 test time」放在同一個 ROI 天平上比較（見 Phase 4 ROI 參考），而不是只看統計上站不站得住腳。當相關性 r 已經很高、殘差很小時，收緊的幅度通常很小、犧牲的良率可忽略；相關性普通時，才需要認真評估這筆交易划不划算。 This isn't free: tightening the limit means dies that would have passed the old limit but fall outside the new one get rejected upstream — even though they would have passed the downstream test. The decision must weigh "upstream yield sacrificed" against "test time saved by removing the downstream test" on the same ROI scale (see the Phase 4 ROI reference), not just whether the statistics hold up. When r is already high and the residual is small, the required tightening is usually tiny and the sacrificed yield negligible; when correlation is only moderate, this trade-off needs careful evaluation before committing to it.

6.3 溫度 Corner 冗餘分析（線性／TC 模型）6.3 Temperature-Corner Redundancy (Linearity / TC Model)

僅在候選屬於「移除某個溫度 corner 的測試」時才需要這條證據線（見 Phase 0 §3.3、§6.2(e)）；若候選單純是跨站（WT↔FT）冗餘、溫度 corner 不變，此節不適用。 This evidence line is only required when the candidate is "remove testing at one temperature corner" (see Phase 0 §3.3 and §6.2(e)); if the candidate is purely cross-stage (WT↔FT) redundancy with the temperature corner unchanged, this section does not apply.

線性假設成立後，用該模型從已測 corner 推算未測 corner 的預測區間（prediction interval），確認預測值距離該 corner 的 spec limit 仍有足夠 margin（同 §6.2(b) 的 guardband 概念，只是把「跨站」換成「跨溫度」）；同時比較已測 corner 與（歷史或抽測）未測 corner 的相關係數 r，樣本數估算沿用 §6.2(a) 的 Fisher z 公式。 Once linearity holds, use the model to project a prediction interval from the tested corner(s) to the untested corner, and confirm the projected value still has adequate margin to that corner's spec limit (the same guardband concept as §6.2(b), with "cross-stage" swapped for "cross-temperature"). Also compare the correlation coefficient r between the tested corner and a historical or sampled untested corner; estimate the required sample size with the same Fisher z formula as §6.2(a).

6.4 判讀準則6.4 Interpretation Criteria

所有適用的證據線都要通過，才視為統計上站得住腳：Every applicable evidence line must pass before the conclusion is statistically sound:

1. Bin-level：觀察期間內 Incremental Catch = 0，且樣本數達到該 Tier 要求的信心水準（見 6.1 表）。 Bin-level: Incremental Catch = 0 over the observation window, with sample size meeting the confidence level required for that Tier (see the table in 6.1).
2. Parameter-level：r ≥ 該 Tier 的 r_min（若原始 r 不足，可改用 §6.2(g) 收緊前站 limit 後重新滿足此條件）；margin 趨勢無逼近 limit 的尾端子族群；Cpk(FT) 與 Cpk(WT) 差異在可接受範圍內；相關矩陣中同站跨溫與跨站同溫配對皆已檢視（§6.2(e)）；殘差通過個體層級異常偵測，無 NNR／DPAT／雙變量離群點（§6.2(f)）。 Parameter-level: r ≥ the Tier's r_min (if the raw r falls short, §6.2(g)'s tightened upstream limit can be used to satisfy this condition instead); no tail sub-population trending toward the limit; the Cpk(FT)–Cpk(WT) gap is within an acceptable range; every same-stage cross-temperature and cross-stage same-temperature pair in the correlation matrix has been reviewed (§6.2(e)); residuals pass individual-level outlier detection with no NNR/DPAT/bivariate outliers (§6.2(f)).
3. 溫度 Corner 線性（僅溫度 corner 候選適用）：線性假設經迴歸驗證成立，未測 corner 的預測區間距 spec limit 仍有足夠 margin。 Temperature-corner linearity (applies only to temperature-corner candidates): the linearity assumption is confirmed by regression, and the untested corner's prediction interval still clears its spec limit with adequate margin.

任一適用條件不成立 → 保留該測項，進入下一輪觀察期；全部適用條件皆成立 → 進入 Phase 4 依風險分級決定移除策略。 If any applicable condition fails → keep the test and continue observing; if every applicable condition holds → proceed to Phase 4 to decide the removal strategy by risk tier.

7Phase 4 — 風險分級與決策準則Phase 4 — Risk Tiering & Decision Rules

不是所有測項都要求同一信心水準。建議依「測項性質」與「終端應用」（消費性電子 vs. 車用／醫療等高可靠度應用，後者風險容忍度應大幅收緊）分三級： Not every test item warrants the same confidence level. Tier by test nature and end application (consumer electronics vs. automotive/medical high-reliability applications, where risk tolerance should be tightened substantially):

8Phase 5 — Pilot 導入與監控Phase 5 — Pilot Rollout & Monitoring

即使統計驗證通過，也不建議直接 100% 移除，先以可回退的方式導入： Even after statistical validation passes, avoid an outright 100% removal — roll out in a reversible way first:

• Skip-lot / Skip-test：先以固定比例（如 1-in-N）抽測取代全測，而非直接關閉，觀察 3–6 個月的實際 escape 狀況。 Skip-lot / Skip-test: Replace full testing with a fixed sampling ratio (e.g., 1-in-N) first, rather than switching the test off outright, and observe actual escape behavior for 3–6 months.
• SPC 監控：對該測項的 margin／Cpk 建立管制圖，設定觸發線；一旦趨勢異常，自動恢復 100% 測試。若該測項是透過 §6.2(g) 收緊前站 limit 才核准移除，PCM／WAT 與此參數的相關性監控為必要項目，不是選配 —— 製程層級的變異通常會先反映在 PCM/WAT，而不是等到 WT/FT 資料才看得出來。 SPC monitoring: Build control charts for the item's margin/Cpk with trigger limits; automatically restore 100% testing if the trend goes out of control. If the item was approved for removal via the §6.2(g) tightened-upstream-limit route, monitoring the PCM/WAT correlation for this parameter is required, not optional — process-level variation typically shows up in PCM/WAT before it's visible in WT/FT data.
• 定期重新驗證：製程節點轉換、封裝廠變更、測試機台汰換等事件，皆應觸發該測項的重新驗證，不可視為一次性決策。 Periodic re-validation: process node changes, assembly house changes, tester replacement, and similar events should all trigger re-validation of the item — this is never a one-time decision.

9Phase 6 — 正式變更、文件化與回退機制Phase 6 — Formal Change, Documentation & Rollback

• Pilot 期間無異常 → 提報正式 ECN／測試流程變更，附上 Phase 3 的統計證據（Bin-level 表 + Parameter-level 圖表）與 Phase 5 的 pilot 紀錄。 No anomalies during the pilot → file a formal ECN / test flow change, attaching the Phase 3 statistical evidence (Bin-level table + Parameter-level charts) and the Phase 5 pilot record.
• 建立「回退觸發條件」清單（如：客戶端出現對應 fail mode、SPC 管制圖超出管制界線、良率異常事件），一旦觸發即自動恢復該測項，不需重新走完整審核流程；「客戶端出現對應 fail mode」的具體查證程序見 Phase 7。 Maintain a list of rollback trigger conditions (e.g., the corresponding fail mode appears in the field, an SPC control chart goes out of limits, a yield excursion), so hitting any of them automatically restores the test without a full re-approval cycle; see Phase 7 for the concrete verification procedure behind "the corresponding fail mode appears in the field."
• 所有決策需留存於變更歷史紀錄（見附錄範本），供未來稽核與新進工程師追溯決策脈絡。 Every decision must be retained in the change history log (see Appendix template) so future audits and new engineers can trace the reasoning.

10Phase 7（Optional）— 客訴快速查證與覆蓋回補Phase 7 (Optional) — Customer Complaint Quick Verification & Coverage Recovery

這個 Phase 是事件觸發，不是循序執行的步驟，所以不畫在 Section 2 的七階段管線圖裡 —— 它是 Phase 6 回退觸發條件中「客戶端出現對應 fail mode」這一項的具體操作程序，目標是縮短「客訴發生」到「查明是否誤移除、並恢復覆蓋」之間的時間。若組織尚未有正式的客訴（CQC）流程可對接，此 Phase 可暫不導入，故標示為 Optional。 This phase is event-triggered, not a sequential step, so it isn't drawn into Section 2's seven-phase pipeline — it's the concrete operating procedure behind Phase 6's rollback trigger "the corresponding fail mode appears in the field," aimed at shrinking the time between a complaint landing and confirming whether a test was wrongly removed and restoring coverage. If the organization has no formal CQC process to plug into yet, this phase can be deferred — hence marked Optional.

1. CQC 進線：從客訴中萃取出對應的 fail mode／電性參數／bin。 CQC intake: extract the associated fail mode / electrical parameter / bin from the complaint.
2. 比對決策紀錄：用該參數查詢 §11.2 的決策紀錄（Risk & Decision Log）—— 這個測項目前是「保留」「抽測」還是「已移除」？ Cross-check the decision log: look up that parameter in the §11.2 Risk & Decision Log — is this test item currently "kept," "sampled," or "removed"?
3. 若為已移除或抽測 → 立即觸發 Phase 6 的回退條件，恢復 100% 測試，不等待正式審核流程；根因調查與統計覆核可以並行進行，不應拖延恢復覆蓋的時間。 If removed or sampled → immediately trigger the Phase 6 rollback, restoring 100% testing without waiting for a formal review cycle; root-cause investigation and statistical re-validation can run in parallel and should not delay restoring coverage.
4. 若查無對應測項（真正的覆蓋缺口）→ 這比「移除錯誤」更嚴重，代表整個流程從未涵蓋這個 fail mode，應立即評估新增測項，並回頭檢視 Phase 0 §3.2 零覆蓋守則當初是否有漏判。 If no corresponding test item exists at all (a genuine coverage gap) → this is more serious than a wrong removal — it means the flow never covered this fail mode. Immediately evaluate adding a new test item, and revisit whether Phase 0 §3.2's zero-coverage guard missed something.
5. 回寫紀錄：把這次 CQC 事件與處置結果補進決策紀錄，作為未來稽核與流程覆核的依據。 Log it back: record the CQC event and its resolution in the decision log, as evidence for future audits and process reviews.

11附錄Appendix

11.1 範本欄位 — 測試對應矩陣（Test Mapping Table）11.1 Template Fields — Test Mapping Table

11.2 範本欄位 — 決策紀錄（Risk & Decision Log）11.2 Template Fields — Risk & Decision Log

11.3 公式參考卡11.3 Formula Reference Card

11.4 延伸參考11.4 Further References

• Clopper–Pearson exact binomial confidence interval — 屬性資料信賴區間的精確解法。Clopper–Pearson exact binomial confidence interval — the exact-form confidence interval for attribute data.
• ASTM E2586 — Standard Practice for Calculating and Using Tolerance Factors for Statistical Tolerance Intervals.
• ISO 16269-6:2014 — Statistical interpretation of data, Part 6: Determination of statistical tolerance intervals.
• Nearest Neighbor Residual（NNR）— 以晶圓上鄰近 die 的平滑預測值篩選空間異常，常見於 Good-Die-in-Bad-Neighborhood 篩選文獻。Nearest Neighbor Residual (NNR) — screens spatial anomalies using a smoothed prediction from neighboring die on the wafer map; widely discussed in Good-Die-in-Bad-Neighborhood screening literature.
• Dynamic Part Average Testing（DPAT）— AEC-Q001 PAT 概念的延伸，以每 lot／wafer 自身分布動態計算限制而非固定 spec limit。Dynamic Part Average Testing (DPAT) — an extension of the AEC-Q001 PAT concept that computes limits dynamically from each lot/wafer's own distribution rather than a fixed spec limit.
• Mahalanobis, P. C. (1936) — 多變量離群偵測的基礎距離量測，將 z-score 概念推廣到相關變數的聯合分布。Mahalanobis, P. C. (1936) — the foundational distance measure for multivariate outlier detection, generalizing the z-score concept to the joint distribution of correlated variables.
• ILAC-G8:2019 — Guidelines on Decision Rules and Statements of Conformity，guardband／收緊 limit 決策規則的參考依據。ILAC-G8:2019 — Guidelines on Decision Rules and Statements of Conformity, a reference for guardband/limit-tightening decision rules.
• Cohen, J. (1988), Statistical Power Analysis for the Behavioral Sciences — 相關係數檢定力與樣本數公式（Fisher z）。power and sample-size formulas for correlation coefficients (Fisher z).
• AEC-Q001 — Guidelines for Part Average Testing（PAT），車用半導體 outlier 管控參考方法。AEC-Q001 — Guidelines for Part Average Testing (PAT); a reference method for outlier control in automotive semiconductors.