Evaluation & Guardrails

March 6, 2026 · View on GitHub

🏥 Quick Return to Emergency Room

You are in a specialist desk.
For full triage and doctors on duty, return here:

WFGY Global Fix Map — main Emergency Room, 300+ structured fixes

WFGY Problem Map 1.0 — 16 reproducible failure modes

Think of this page as a sub-room.
If you want full consultation and prescriptions, go back to the Emergency Room lobby.

Evaluation disclaimer (GlobalFixMap · Eval)
The Eval section describes patterns and tools for building evaluation loops around AI systems.
All example scores, thresholds and labels are illustrative and depend on the local environment in which they were produced.
They should be read as diagnostic hints and design patterns rather than as evidence that any specific model or system has been scientifically validated.
If you adopt these ideas, please re run the evaluations in your own stack, check sensitivity to configuration changes and document the limits of what your numbers actually support.

A hub to prove fixes actually work and won’t regress.
Use this folder when you want to validate that your RAG / LLM pipeline changes are stable, measurable, and reproducible.
The goal is to prevent “double hallucination,” enforce acceptance gates, and keep evaluation pipelines auditable.

What this page is

A compact playbook to evaluate RAG quality and reasoning stability
Drop-in guardrails that catch failures before users see them
CI/CD-ready acceptance targets you can copy directly

When to use

You shipped a fix but cannot show measurable improvement
Answers look plausible but citations or snippets don’t match
Performance flips between seeds, sessions, or agent mixes
Latency tuning silently changes accuracy
Your team disagrees on whether a fix is “actually better”

Open these first

RAG precision/recall spec → eval_rag_precision_recall.md
Latency versus accuracy method → eval_latency_vs_accuracy.md
Cross-agent agreement tests → eval_cross_agent_consistency.md
Semantic stability checks → eval_semantic_stability.md
Why-this-snippet schema → retrieval-traceability.md
Snippet & citation schema → data-contracts.md

Common evaluation pitfalls

Double hallucination → Metrics look good (BLEU, ROUGE) but answers cite the wrong snippet
Recall illusion → Top-k recall seems fine, yet ΔS(question, context) is still unstable
Seed lottery → Success on one random seed hides instability across paraphrases
Hybrid flapping → HyDE + BM25 mixes reorder results differently every run
Over-clamping → Filters enforce tone but fail to fix logical drift
Benchmark mismatch → Eval set ignores OCR noise or multilingual inputs
No trace table → You cannot audit which snippet was cited

Fix in 60 seconds

Adopt acceptance gates
- Retrieval sanity: token overlap ≥ 0.70 to the gold section
- ΔS(question, context) ≤ 0.45 on median across suite
- λ_observe stays convergent across 3 paraphrases
Require citations first
- Enforce cite-then-answer with data-contracts.md
- Log: question, retrieved ids, snippet spans, ΔS, λ
Stability before speed
- Always measure latency vs accuracy before tuning
- See eval_latency_vs_accuracy.md
Cross-agent cross-check
- Run 2 strong models on the same retrieval
- See eval_cross_agent_consistency.md
Regression fence in CI
- Block merges if ΔS median > 0.45 or coverage < 0.70
- See eval_rag_precision_recall.md

Minimal checklist

Trace table saved (citations + snippet spans)
ΔS computed per item; λ recorded at retrieval & reasoning
Coverage ≥ 0.70 to gold snippet
Cross-agent agreement tested
Latency vs accuracy chart archived with run id

Acceptance targets

ΔS(question, context) median ≤ 0.45
λ convergent across 3 paraphrases
Token overlap ≥ 0.70 to gold snippet
No unexplained rank flips on hybrid retrievers
CI blocks merges when targets fail

FAQ

Q: What is ΔS and why does it matter?
A: ΔS measures semantic distance between your query and retrieved context. Values above 0.45 indicate unstable retrieval, even if the snippet looks similar.

Q: Why not just trust BLEU/ROUGE?
A: They score surface similarity, not factual correctness. A fluent but wrong answer can pass BLEU. WFGY gates enforce snippet fidelity.

Q: What does λ_observe mean?
A: λ_observe tracks whether paraphrased queries converge on the same retrieval. Divergence shows instability that will confuse users.

Q: How do I build a trace table?
A: For every eval item, log question, retrieved ids, snippet spans, ΔS, λ_state. This makes your pipeline auditable later.

Q: Do I need a big eval set?
A: No. Start with 20 smoke-test items, including multilingual or noisy samples. Scale up only after you pass basic gates.

Q: What if latency tuning drops accuracy?
A: Always plot latency vs accuracy. Use the knee point of the curve, not the fastest or slowest configuration.

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

Explore More

Layer	Page	What it’s for
⭐ Proof	WFGY Recognition Map	External citations, integrations, and ecosystem proof
⚙️ Engine	WFGY 1.0	Original PDF tension engine and early logic sketch (legacy reference)
⚙️ Engine	WFGY 2.0	Production tension kernel for RAG and agent systems
⚙️ Engine	WFGY 3.0	TXT based Singularity tension engine (131 S class set)
🗺️ Map	Problem Map 1.0	Flagship 16 problem RAG failure taxonomy and fix map
🗺️ Map	Problem Map 2.0	Global Debug Card for RAG and agent pipeline diagnosis
🗺️ Map	Problem Map 3.0	Global AI troubleshooting atlas and failure pattern map
🧰 App	TXT OS	.txt semantic OS with fast bootstrap
🧰 App	Blah Blah Blah	Abstract and paradox Q&A built on TXT OS
🧰 App	Blur Blur Blur	Text to image generation with semantic control
🏡 Onboarding	Starter Village	Guided entry point for new users

If this repository helped, starring it improves discovery so more builders can find the docs and tools.