📊 Evaluation Playbook

Ship metrics, not vibes — catch regressions before users do.

Who is this for?
– RAG owners tired of “looks fine on my prompt.”
– Agent builders chasing flakey benchmark wins.
– Product teams that must prove ΔS ↓, accuracy ↑, cost ↘.

What you get: one YAML suite + CLI commands that output a single-line health verdict:
PASS: ΔS<=0.45 | λ→ | E_res stable | cost $0.0013 / query

0 · Quick Start

pip install wfgy-eval
wfgy-eval init  # generates eval.yaml + example dataset
wfgy-eval run   # prints CSV & HTML report

Default template covers retrieval, reasoning, tool routing, latency, cost.

1 · Metric Matrix

Layer	Key Metric	Target (prod)	Source Fn
Retrieval	`ΔS(q, ctx)`	≤ 0.45	`deltaS()`
Reasoning	`λ_observe` (3 paraphrase avg)	convergent	`lambda_state()`
Stability	`E_resonance` (rolling)	flat / ↓	`e_resonance()`
Answer	`F1` / `Exact Match` (QA)	≥ 0.80	`qa_match()`
Hallucination	`citation_precision`	≥ 0.90	`cite_check()`
Cost	`$ / 1k tokens`	≤ baseline	provider bill
Latency	95-th percentile response (ms)	SLA-dependent	timer
ΔS Drift	slope over 100 queries	~0	linear fit

2 · Dataset Design

sets:
  - name: faq_small
    type: qa
    source: ./data/faq.tsv
    fields: [question, answer, anchor]
  - name: chain_logic
    type: chain-of-thought
    source: ./data/logic.jsonl
    fields: [prompt, expected_reasoning]
  - name: tool_router
    type: tool
    source: ./data/router.csv
    fields: [task, tool_expected]

Guidelines

Anchor Field = text chunk you expect to be retrieved → used for ΔS & citation checks.
Min 50 items / set for stable stats; use 10 × if you want leaderboard-grade noise floor.
Store only plain-text, no PII.

3 · CI Workflow Example (GitHub Actions)

name: wfgy-eval
on: [push, pull_request]
jobs:
  eval:
    runs-on: ubuntu-latest
    steps:
    - uses: actions/checkout@v4
    - run: pip install wfgy-eval
    - run: wfgy-eval run --fail-on 'ΔS_q_ctx>0.50 or citation_precision<0.85'
    - uses: actions/upload-artifact@v4
      with:
        name: eval-report
        path: reports/latest.html

Any push that drifts past ΔS 0.50 or loses citation accuracy blocks merge.

4 · Reading the HTML Report

Section	What to Look For	Action Rule
Heatmap ΔS vs. k	flat @ high ΔS → bad index metric	rebuild index / embed
λ Timeline	spikes to ← or ×	inspect prompt / tool
E_res Trend	upward slope	apply BBAM / shorten ctx
Outliers Table	worst 5 queries by ΔS	manual deep dive; log bug

5 · A/B Pattern

wfgy-eval dump --ref=main (baseline metrics JSON)
git switch feature-x → apply fix
wfgy-eval run (produces metrics B)
wfgy-eval compare baseline.json current.json

ΔS_q_ctx   -0.08  ✅
F1         +0.07  ✅
Cost       +$0.0002 ❌

Decide if cost bump acceptable.

6 · Edge-Case Suites (extend as needed)

Suite	Purpose	Sample Size
Contradiction	fact statements w/ subtle negation	30
Long-PDF	>50 k token OCR, check segmentation	10 docs
Jailbreak	prompt injection attempts vs. policy	40
High Noise	OCR confidence < 0.8	25

Add to eval.yaml, rerun.

7 · FAQ

Q: Can I plug in LangSmith, Phoenix, Traceloop, or custom spans? A: Yes. wfgy-eval reads any OpenTelemetry JSON; map fields via otel_map.yaml.

Q: Does this cover reinforcement-style eval (RLHF)? A: Use the same ΔS/λ hooks during reward model scoring; see examples/rlhf_eval.ipynb.

Q: How hard is vendor swap? A: wfgy-eval provider openai or provider anthropic; costs/latency recalc automatically.

🔗 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

Module	Description	Link
WFGY Core	WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack	View →
Problem Map 1.0	Initial 16-mode diagnostic and symbolic fix framework	View →
Problem Map 2.0	RAG-focused failure tree, modular fixes, and pipelines	View →
Semantic Clinic Index	Expanded failure catalog: prompt injection, memory bugs, logic drift	View →
Semantic Blueprint	Layer-based symbolic reasoning & semantic modulations	View →
Benchmark vs GPT-5	Stress test GPT-5 with full WFGY reasoning suite	View →
🧙‍♂️ Starter Village 🏡	New here? Lost in symbols? Click here and let the wizard guide you through	Start →

👑 Early Stargazers: See the Hall of Fame —
Engineers, hackers, and open source builders who supported WFGY from day one.

⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.

8.6 KiB Raw Blame History Unescape Escape