11 KiB
Eval Benchmarking — Protocols, Targets, and Reporting
🧭 Quick Return to Map
You are in a sub-page of Eval.
To reorient, go back here:
- Eval — model evaluation and benchmarking
- 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 desk within a ward.
If you need the full triage and all prescriptions, return to the Emergency Room lobby.
Evaluation disclaimer (benchmarking)
All scores and examples on this page are scenario specific debug signals.
They are not an official leaderboard or scientific proof and do not show that one model is always better.
Use them as local guidance for your own stack and re run the setup when you change models, data or prompts.
This page defines a clean, repeatable way to benchmark your pipeline and prove that a fix actually improved behavior. It uses the same WFGY instruments as everywhere else: ΔS for semantic stress, λ_observe for stability, and E_resonance for coherence over long windows.
Open these first
- RAG map and recovery path: RAG Architecture & Recovery
- Eval playbook and gates: Eval Playbook · Regression Gate
- ΔS and λ instruments: deltaS_thresholds.md · lambda_observe.md
- Coverage and drift: coverage_tracking.md · variance_and_drift.md
- Gold construction: Goldset Curation
- Precision and recall: Eval RAG Precision/Recall
- Public benchmark page: Benchmark vs GPT-5
Acceptance targets
Benchmark runs are accepted when all of the following pass:
- Precision ≥ 0.80 on cited snippets
- Recall ≥ 0.70 to target sections
- ΔS(question, cited) ≤ 0.45 for 80 percent of pairs
- λ remains convergent across three paraphrases and two seeds
- Run to run variance ≤ 0.10 for precision and recall
- No regression versus previous accepted run by more than 3 percent on any metric without a documented goldset change
Benchmark protocols
Protocol A: A versus A+WFGY
Purpose is to prove the benefit of the WFGY layer with the same base model and the same data.
- Same dataset, prompts, and retriever
- Arm 1 baseline without WFGY
- Arm 2 with WFGY Core and the Problem Map instruments
- Compare precision, recall, ΔS distribution, λ stability, latency
Protocol B: Cross model control
Purpose is to show that gains are not tied to a single provider.
- Choose two or more providers from your production shortlist
- Keep gold, retriever, and prompts constant
- Run baseline and WFGY arms per provider
- Report deltas within provider and also pooled across providers
Protocol C: Stress and stability
Purpose is to surface brittleness that simple single shot tests will hide.
- For each question, run three paraphrases and two seeds
- Expand k values in retrieval to 5, 10, 20
- Record λ states per step and ΔS histograms
- Accept only when variance and flip rates are within thresholds
Dataset design
- Use at least 50 questions spanning three difficulty bands
- Each question has gold snippets with offsets and token ranges
- Include adversarial distractors that look semantically close in the same index
- Mixed language tests require tokenizer checks and casing constraints
- For long context tasks, mark the join points for E_resonance probes
See the construction details in Goldset Curation.
Metrics you must report
- RAG: precision, recall, ΔS mean and p90, λ flip rate, coverage
- Reasoning: correction stickiness after one steer, hallucination re-entry rate
- Latency: median and p90 per step (retrieve, rerank, reason)
- Stability: variance across paraphrases and seeds
- Cost: normalized tokens or API units per correct answer
Targets and field definitions are aligned with the pages linked in the Open section above.
JSONL reporting schema
Each benchmark row is one question run in one arm. Use JSONL for easy diffing.
{
"suite": "v1_rag_core",
"protocol": "A",
"arm": "baseline" ,
"provider": "openai",
"model": "gpt-4o-mini-2025-07",
"question_id": "q_042",
"paraphrase": 2,
"seed": 13,
"k": 10,
"precision": 0.86,
"recall": 0.72,
"coverage": 0.74,
"ΔS_avg": 0.38,
"ΔS_p90": 0.47,
"λ_state_seq": ["→","→","→"],
"λ_flip_rate": 0.0,
"latency_ms": { "retrieve": 120, "rerank": 45, "reason": 930 },
"tokens": { "in": 1850, "out": 420 },
"hallucination_reentry": false,
"notes": "meets thresholds"
}
For aggregation, compute means and p90 per protocol and arm, then produce deltas for A vs A+WFGY and for each provider in Protocol B.
Minimal 60 second run
-
Pick 10 questions from the goldset with citations.
-
Run Protocol A comparing baseline vs WFGY on a single provider.
-
Record JSONL and compute precision, recall, ΔS, λ stability.
-
If any acceptance target fails, route to the right fix page:
- Wrong meaning with high similarity → embedding-vs-semantic.md
- Messy ordering → rerankers.md
- No trace or mixed sources → retrieval-traceability.md
Common pitfalls and how to avoid them
-
Goldset ambiguity Two valid snippets exist but only one is labeled. Expand gold with alt spans. See Goldset Curation.
-
Tokenizer and casing drift Mixed language corpora collapse precision. Apply the multilingual checklist and keep analyzers consistent. See Data Contracts and Rerankers.
-
Hidden index skew Flat high ΔS across k suggests metric or normalization mismatch. Rebuild index and verify with a small canary set. See RAG Playbook and Embedding vs Semantic.
-
Prompt header instability λ flips when the header order changes. Lock schema and clamp variance with BBAM.
-
Eval leakage Using dev answers in prompts inflates metrics. Keep a holdout split and rotate keys between runs.
Publishing results
When you publish, include:
- Protocol tables with acceptance ticks
- ΔS histograms and λ flip rates per arm
- Precision and recall bars with error bands across paraphrases
- A short narrative mapping any failures to the exact Problem Map pages you used to fix them
- A link to your JSONL and the goldset diffs
Public examples and figures live here: Benchmark vs GPT-5
🔗 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 based tension engine |
| Engine | WFGY 2.0 | Production tension kernel and math engine for RAG and agents |
| Engine | WFGY 3.0 | TXT based Singularity tension engine, 131 S class set |
| Map | Problem Map 1.0 | Flagship 16 problem RAG failure checklist and fix map |
| Map | Problem Map 2.0 | RAG focused recovery pipeline |
| Map | Problem Map 3.0 | Global Debug Card, image as a debug protocol layer |
| Map | Semantic Clinic | Symptom to family to exact fix |
| Map | Grandma’s Clinic | Plain language stories mapped to Problem Map 1.0 |
| Onboarding | Starter Village | Guided tour for newcomers |
| App | TXT OS | TXT semantic OS, fast boot |
| App | Blah Blah Blah | Abstract and paradox Q and A built on TXT OS |
| App | Blur Blur Blur | Text to image with semantic control |
| App | Blow Blow Blow | Reasoning game engine and memory demo |
If this repository helped, starring it improves discovery so more builders can find the docs and tools.
