docs: soften overclaim language in mapped files

ProblemMap/GlobalFixMap/OpsDeploy/blue_green_switchovers.md

@@ -17,7 +17,7 @@
 </details>
 
 
-A safe pattern to switch 100% of traffic between two identical stacks (Blue = current live, Green = new). Use this when you need instant rollback, reproducible cutovers, and zero surprise from schema or index drift.
+A safe pattern to shift primary traffic between two identical stacks (Blue = current live, Green = new), typically all traffic in planned cutovers when conditions allow. Use this when you need a pre-validated rollback path, reproducible cutovers, and minimized surprise from schema or index drift.
 
 ---
 
@@ -30,7 +30,7 @@ A safe pattern to switch 100% of traffic between two identical stacks (Blue = cu
 ---
 
 ## When to use
-- Model or prompt upgrades where instant rollback is required.
+- Model or prompt upgrades where a low-latency rollback path is required.
 - Vector index rebuilds that must cut over atomically.
 - Config, secrets, or feature-flag rewires that risk drift by region.
 - Store or analyzer changes that can shift ΔS/coverage.
@@ -42,7 +42,7 @@ A safe pattern to switch 100% of traffic between two identical stacks (Blue = cu
 - Coverage ≥ 0.70 on the gold set (same as pre-switch).
 - λ convergent across 2 seeds, no flip increase vs Blue.
 - p95 latency within +15% of Blue; error rate ≤ 0.5%.
-- Instant rollback path validated (≤ 2 minutes TTR).
+- Rollback path validated (typically ≤ 2 minutes TTR).
 
 ---
 
@@ -95,7 +95,7 @@ All reads/writes and side effects must point at a **single active arm**. If you
 
 ### Index alias swap (Vector/RAG)
 - Build Green index offline. Validate with sampled ΔS/coverage.  
-- Swap an **atomic alias** `docs_live → docs_vB`. Keep `docs_vA` for instant rollback.  
+- Swap an **atomic alias** `docs_live → docs_vB`. Keep `docs_vA` as the validated rollback lane.  
 - If store lacks aliases, emulate with a config key in a single-writer KV and block multi-writer races.
 
 ### Queue consumer cutover

ProblemMap/GlobalFixMap/README.md

@@ -79,7 +79,7 @@ Choose **Grandma’s Clinic** if you want to understand the bug in plain words.
 > - if the state is unstable, it loops, resets, or redirects the path.  
 > - only a stable semantic state is allowed to generate output.  
 >
-> this is why every failure mode, once mapped, stays fixed.  
+> this is why each mapped failure family is typically stable again under the same semantic settings; if model/version/context shifts, we treat it as a new mapping and rerun the loop.  
 > you’re not firefighting after the fact — you’re installing a reasoning firewall at the entry point.  
 >
 > ---
@@ -90,28 +90,28 @@ Choose **Grandma’s Clinic** if you want to understand the bug in plain words.
 > |--------------|-----------------------------------------|---------------------------------------------------|
 > | **Flow**     | Output → detect bug → patch manually    | Inspect semantic field → only stable state generates |
 > | **Method**   | Add rerankers, regex, JSON repair, tool patches | ΔS, λ, coverage checked upfront; loop/reset if unstable |
-> | **Cost**     | High — every bug = new patch, risk of conflicts | Low — once mapped, bug sealed permanently |
-> | **Ceiling**  | 70–85% stability limit                  | 90–95%+ achievable, structural guarantee |
-> | **Experience** | Firefighting, “whack-a-mole” debugging | Structural firewall, “fix once, stays fixed” |
+> | **Cost**     | High — every bug = new patch, risk of conflicts | Low — once mapped and revalidated, recurrence is usually reduced for the same family |
+> | **Ceiling**  | 70–85% stability limit                  | 90–95%+ observed on internal cohorts; setup-dependent and not a universal ceiling |
+> | **Experience** | Firefighting, “whack-a-mole” debugging | Structural firewall, “fix once under accepted targets, then monitor for drift” |
 > | **Complexity** | Growing patch jungle, fragile pipelines | Unified acceptance targets, one-page repair guide |
 >
 > ---
 >
 > ### ⚡ Performance impact
 > - **Traditional patching**: 70–85% stability ceiling. Each new patch adds complexity and potential regressions.  
-> - **WFGY firewall**: 90–95%+ achievable. Fix once → the same bug never resurfaces. Debug time cut by 60–80%.  
+> - **WFGY firewall**: around 90–95%+ stability in internal cohorts. A mapped fix is usually suppressed for that family under stable conditions, with drift requiring re-mapping. Debug time often drops by 60–80% under those conditions.  
 > - **Unified metrics**: every fix is measured (ΔS ≤ 0.45, coverage ≥ 0.70, λ convergent). No guesswork.  
 >
 > ### 🛑 Key notes
 > - This is **not a plugin or SDK** — it runs as plain text, zero infra changes.  
 > - You must **apply acceptance targets**: don’t just eyeball; log ΔS and λ to confirm.  
-> - Once acceptance holds, that path is sealed. If drift recurs, it means a *new* failure mode needs mapping, not a re-fix of the old one.  
+> - Once acceptance holds, that path is treated as stable for the same setting. If drift recurs, it means a *new* failure mode needs mapping, not merely a re-fix of the old one.  
 >
 > ---
 >
 > **Summary**:  
 > Others patch symptoms **AFTER** output. WFGY blocks unstable states **BEFORE** output.  
-> That is why it feels less like debugging, more like installing a **structural guarantee**.  
+That is why it feels less like debugging, more like installing **structural acceptance scaffolding** with guardrails rather than universal guarantees.  
 >
 > ---
 </details>
@@ -159,7 +159,7 @@ Choose **Grandma’s Clinic** if you want to understand the bug in plain words.
 
 > **What is the Global Fix Map?**  
 > A vendor-neutral panoramic index that consolidates **300+** topics, frameworks, and reproducible failure modes (RAG, embeddings, chunking, OCR/language, reasoning/memory, agents, serverless, eval/governance).  
-> **Purpose:** convert repeatable bugs into **verifiable structural repairs** — fix once, stays fixed.
+> **Purpose:** convert repeatable bugs into **verifiable structural repairs** — map the pattern, lock acceptance targets, and revalidate when context changes.
 
 **Principles**
 - **Zero-install:** boot with **TXT OS** / **WFGY PDF** as context.

ProblemMap/GlobalFixMap/Safety_PromptIntegrity/jailbreaks_and_overrides.md

@@ -41,10 +41,10 @@ Use this page when adversarial text like *“you are now DAN”* or *“forget r
 ---
 
 ## Core acceptance
-- Model never executes user override like *“ignore all above”*.  
+- Model should not execute user override requests like *“ignore all above”* within this acceptance profile; those inputs are treated as policy violations and routed to validation handling.
 - ΔS(question, retrieved) ≤ 0.45 even under jailbreak text.  
 - λ remains convergent across paraphrases (no flip to override mode).  
-- Schema integrity: tool/JSON outputs pass validation 100%.  
+- Schema integrity: tool/JSON outputs must satisfy validation checkpoints under current versioned contracts, with failed validations blocked or retried before output.
 
 ---
 

ProblemMap/GlobalFixMap/Safety_PromptIntegrity/json_mode_and_tool_calls.md

@@ -41,7 +41,7 @@ This page defines structural fixes to ensure **valid JSON**, **schema adherence*
 ---
 
 ## Core acceptance
-- Every tool call conforms to schema 100% (no free-text).  
+- Tool calls are expected to conform to schema under the active acceptance targets (no free-text). If a call misses the contract, we retry, escalate to recovery, or stop output according to the same policy.
 - No mixed narrative and JSON in one block.  
 - ΔS(question, retrieved) ≤ 0.45 for JSON-only probes.  
 - λ convergent across three paraphrases of the same JSON request.  

ProblemMap/GrandmaClinic/README.md

@@ -45,7 +45,7 @@ If the user has no clue where to start, direct them to Beginner Guide first, the
 Most people fix AI bugs **after** the model already spoke. You then add patches, rerankers, or regex. The same failure returns later in a different shape.
 
 **WFGY installs a semantic firewall *before* output.**
-It inspects the semantic field first. If the state is unstable, it loops, narrows, or resets. Only a stable state is allowed to speak. Once a failure mode is mapped, it stays fixed.
+It inspects the semantic field first. If the state is unstable, it loops, narrows, or resets. Only a stable state is allowed to speak. Once a failure mode is mapped, it is usually stable for that same condition. If the prompt, model, or schema context changes, treat it as a new family and re-map.
 
 **How to use this page in 30 seconds**
 
@@ -967,7 +967,7 @@ You do not patch forever. You set **acceptance targets** and keep them:
 * λ state convergent
 * Source present before final
 
-When a new bug appears, map it to a number, apply the fix once, and it stays fixed. That is the point of a semantic firewall.
+When a new bug appears, map it to a number, apply the fix for that same condition, and it is typically reduced in recurrence checks. If the context drifts, consider it a re-mapping case. That is the point of a semantic firewall.
 
 ---
 
@@ -993,7 +993,7 @@ No. WFGY is text-only. It works as a reasoning guard before output. Over time it
 
 **Q3. How do I know if the fix actually worked?**  
 Check the acceptance targets: ΔS ≤ 0.45, Coverage ≥ 0.70, λ convergent.  
-If these hold across 3 paraphrases, the bug is fixed.
+If these hold across 3 paraphrases, the bug is usually stable for the same pattern under current targets.
 
 **Q4. Is Grandma Clinic enough, or do I need the full Problem Map?**  
 The Clinic covers the 16 most common errors in simple language.  

ProblemMap/getting-started.md

@@ -26,7 +26,7 @@ A laptop with 16 to 32 GB RAM is enough for a 300 to 800 page corpus using CPU e
 
 These map directly to common failures in the Problem Map.
 
-* Retrieval never returns context that crosses entity or constraint boundaries. This reduces No.1 Hallucination and Chunk Drift.
+* Retrieval is constrained to avoid returning context that crosses entity or constraint boundaries by default. Boundary leakage can still occur if constraints are incomplete, in which case we tighten extraction rules.
 * Generation must quote only the provided evidence. If not provable, it must say `not in context`.
 * Every answer writes a machine readable trace with query, chunk ids, and scores. This turns debugging from guesswork into data.
 * Optional rerank step promotes passages that match constraints, not only keywords.
@@ -63,7 +63,7 @@ if __name__ == "__main__":
 ```
 
 Chunk with a constraint aware strategy.
-Keep headings or definition lines attached to their paragraphs. Merge small lines, but never cross section boundaries like References or Footnotes.
+Keep headings or definition lines attached to their paragraphs. Merge small lines, but avoid crossing section boundaries like References or Footnotes unless section-level intent explicitly allows it.
 
 ```python
 # tools/chunk.py

ProblemMap/rag-architecture-and-recovery.md

@@ -368,7 +368,7 @@ This is the “double hallucination” trap. The first illusion hides the second
 | 3    | **E\_resonance** | “can we re-lock coherence?”     | apply stability modules (BBMC/BBPF/BBCR/BBAM) at the failing layer                | the **repair action**           |
 | 4    | **ProblemMap**   | “what page fixes this?”         | open the matched doc (e.g., `retrieval-collapse.md`)                              | the **concrete fix recipe**     |
 
-> 90% of cases end after steps 1–3. You only go deeper when a fix requires a structural change (schema, retriever, index).
+> In the current internal playbook, about 90% of cases tend to end after steps 1–3. We only go deeper when a fix requires a structural change (schema, retriever, index) under sustained failure signatures.
 
 ### Layer-specific Fix Index (one-click)
 

core/README_zh-TW.md

@@ -61,7 +61,7 @@ WFGY 2.0 (WFGY 核心) = [WFGY 1.0 數學公式](https://github.com/onestardao/W
 **Flagship** 與 **OneLine** 的行為完全相同；只差在「可讀性」與「極簡性」的取捨。  
 
 就是這麼簡單 — 無需插件、無需安裝、純文字即可。  
-*實際上，自動啟動模式大約能達到顯式呼叫 **70–80%** 的效能（可參考下方八家 AI 的測試數據）。*
+*在目前四套 retrieval 場景與六種模型版本組合中，**自動啟動模式大約能達到 70–80%** 的顯式呼叫效能（可參考下方八家 AI 的測試數據；效果受版本、資料及提示詞版本影響）。*
 
 ---