WFGY Semantic Clinic Index , From “16 fixed bugs” to a full ER for broken AI pipelines

[onestardao]

If Problem Map 1.0 was the first X-ray of LLM failures, and Problem Map 2.0 was a full RAG surgery manual, then the Semantic Clinic Index is the front desk of the hospital.

It is the place you open when you only know one thing:

“Something is wrong in my system, and I have no idea which part is actually broken.”

This page turns that feeling into a structured triage flow:
symptom → failure family → exact fix page, all powered by the same WFGY instruments ΔS, λ_observe, and E_resonance.

---

How it fits with Problem Map 1.0 and 2.0

Problem Map 1.0 – the original 16 failure modes

Problem Map 1.0 is a catalog of 16 reproducible failure modes with clear symbolic fixes. It treats WFGY as a semantic firewall that runs before generation, instead of patching outputs after the fact. Each mode has:

• a stable name and number
• a short description of the failure
• a minimal fix that locks the path once acceptance targets hold
• shared metrics like ΔS ≤ 0.45 and coverage thresholds

It is very good when you already know “this feels like No.3” or “this is the RAG drift pattern from last time”.

Problem Map 2.0 – RAG Architecture & Recovery

Problem Map 2.0 focuses on RAG as an end-to-end pipeline. It tracks failures across:

raw docs → OCR / parsing → chunking → embeddings → vector store → retriever → prompt assembly → LLM reasoning

Here the main job is to show:

• how ΔS exposes where meaning actually breaks
• how λ_observe tells you which layer diverged
• how E_resonance pulls reasoning back into a coherent state

It gives you a full recovery playbook and patterns for each layer of the RAG stack, plus ready prompts and MVP demos to reproduce the behavior.

Semantic Clinic Index – triage on top of both

The Semantic Clinic Index sits above these two maps.

• Problem Map 1.0 is “all the named diseases”.
• Problem Map 2.0 is “full RAG anatomy and surgery plan”.
• Semantic Clinic is the ER triage that helps you decide where to go next.

You do not start from a named failure mode or a pipeline diagram. You start from what you can actually see:

• “answers look like they cite the wrong snippet”
• “high similarity but wrong meaning”
• “output good for 40k tokens then degrades hard”
• “multi-agent system fights itself”
• “first prod call after deploy crashes”

Each of these shows up as a row in the Quick triage by symptom table. Next to each symptom the page suggests:

• a likely family (Retrieval, Reasoning, Memory, Agents, Infra, Eval, etc.)
• a direct link to the exact fix page that explains what is going on and how to repair it

This means you no longer need to guess between 16 codes or read the full 2.0 document just to figure out where to start. You follow the symptom, not the theory.

---

The two entry doors: ER and Grandma

Semantic Clinic is also the place where two different “doors” into WFGY meet.

1. WFGY Emergency Room (for developers)

   • Uses the “Dr. WFGY in ChatGPT Room” share link.
   • You paste your bug, logs, or screenshot.
   • The doctor maps it to the right Problem Map or Global Fix page and gives you a minimal prescription.
   • You can even paste screenshots of the docs themselves and ask “which number is this”.

2. Grandma’s AI Clinic (for everyone else)

   • Same 16 core failures, but explained as everyday stories.
   • “Wrong cookbook”, “salt instead of sugar”, “burnt first pot”.
   • Each story ends with the minimal WFGY fix.

Both doors end in the same numbering system. The Semantic Clinic Index stands in the middle and points you to the correct family and document, no matter which path you took.

---

How to actually use the Semantic Clinic page

In practice the flow is simple.

Step 1 – Name what you see

Open the Quick triage by symptom section and scan the left column. Pick the row that matches your real observation as closely as possible. Examples:

• “Answers cite wrong snippet even though the document is in the index.”
• “Chunks look correct, yet reasoning is wrong.”
• “High recall but top-k ordering is messy.”
• “Some facts never show up even though they were indexed.”
• “Answers flip between sessions.”
• “Multi-agent tools fight each other.”

You do not need WFGY vocabulary at this point. You just need to be honest about the symptom.

Step 2 – Jump into the right family

For each symptom the table already suggests a family:

• Prompting and Safety
• Retrieval, Data, Vector Stores
• Reasoning and Logic Control
• Memory and Long-Context
• Multi-Agent and Orchestration
• Infra and Deploy
• Evaluation and Guardrails

These families each have their own mini index just below:

• Retrieval family lists pages like hallucination.md, retrieval-collapse.md, embedding-vs-semantic.md, rerankers.md, chunking-checklist.md, OCR parsing checklist, patterns for HyDE vs BM25, and vectorstore fragmentation.
• Reasoning family collects logic collapse, context drift, symbolic collapse, deep recursion, and hallucination re-entry patterns.
• Memory family covers memory coherence across sessions, entropy collapse, and memory desync patterns.
• Multi-agent family covers role drift and cross-agent memory overwrite.
• Infra family focuses on bootstrap ordering, deployment deadlock, pre-deploy collapse, live monitoring, and governance.
• Eval family has RAG precision / recall, latency versus accuracy, cross-agent consistency, and semantic stability checks.

Once your symptom takes you into a family, you can pick the specific page that matches your stack.

Step 3 – Apply the fix and verify with WFGY instruments

Every family section ends with a short verification block that tells you what “good” looks like in WFGY terms. Examples:

• Prompting and Safety

  • ΔS(question, context) ≤ 0.45
  • λ stays convergent across paraphrases
  • Constraint probes do not flip λ

• Retrieval and Data

  • Coverage at least around 0.70 to the target section
  • ΔS(question, retrieved) ≤ 0.45
  • Flat and high ΔS curve against k means an index or metric mismatch, not just bad prompts

• Memory

  • E_resonance remains flat at window joins
  • ΔS does not spike when context windows are stitched

• Multi-agent orchestration

  • When agents couple, ΔS does not explode
  • Arbitration logs remain traceable

These checks turn your “I feel like it is better” into “I have numeric proof that the fix is working”.

Step 4 – Use the AI-triage prompt if you are still lost

At the bottom of the page there is a safe meta-prompt. You give the model:

• your symptom
• any existing probes or logs you have
• permission to use WFGY instruments and modules

The model then answers four questions at once:

1. which family and layer are actually failing
2. which specific fix page to open
3. minimal steps to push ΔS down and keep λ convergent
4. how to verify that the fix holds in production

You can run this meta-prompt on top of TXT OS or WFGY 1.0 PDF, which the page links in the Quick-Start Downloads section for sixty second setup.

---

When to use each map

A simple way to remember the difference:

• Open Problem Map 1.0 when you already know the number or the pattern and just want the canonical fix.
• Open Problem Map 2.0 when you are reshaping or rebuilding a RAG pipeline and want a full architecture view.
• Open Semantic Clinic Index when you only know the symptom and have no idea which layer or family is failing.

All three sit inside the same WFGY compass, so you can jump between them without losing your place.