WFGY/ProblemMap/GlobalFixMap/RAG/eval_drift.md

Eval Drift — Guardrails and Fix Pattern

🧭 Quick Return to Map

You are in a sub-page of RAG.
To reorient, go back here:

• RAG — retrieval-augmented generation and knowledge grounding
• 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 (RAG drift)
Drift signals here are measured inside specific RAG pipelines and datasets.
They are debugging indicators, not proof that a system will stay stable in all real workloads.

---

When evaluation metrics swing unpredictably across runs (precision, recall, ΔS, coverage) even though the data and index appear unchanged.
This signals eval drift: your evaluation harness is not structurally stable.

---

Open these first

• Recovery overview: RAG Architecture & Recovery
• Retrieval knobs: Retrieval Playbook
• Eval acceptance rules: Eval — Quality & Readiness Gates
• Precision & recall contract: Eval RAG Precision/Recall
• Ordering control: Bootstrap Ordering
• Pre-deploy safety: Predeploy Collapse

---

Core acceptance

• ΔS(question, retrieved) ≤ 0.45 across 3 paraphrases
• Coverage ≥ 0.70 per target section
• λ convergent on 2 seeds, stable across runs
• Variance of metrics ≤ 0.05 across replays

---

Typical symptoms → exact fix

Symptom	Likely cause	Open this
Precision/recall varies ±0.20 each run	eval harness non-deterministic	Eval Precision/Recall
Identical queries give different metrics	bootstrap not fenced	Bootstrap Ordering
Eval metrics collapse on fresh deploy	index not fully warmed	Predeploy Collapse
Coverage < 0.50 despite gold answers	embedding or chunk drift	Embedding ≠ Semantic, Chunking Checklist

---

Fix in 60 seconds

1. Lock seeds
   Fix random seeds at retrieval, reranker, and eval harness layers.

2. Fence bootstrap
   Require VECTOR_READY==true and index hash match before eval begins.

3. Replay 3 paraphrases
   Eval the same question with 3 paraphrases. Require ΔS variance < 0.05.

4. Cross-seed check
   Run two seeds. λ must remain convergent across both.

5. Regression gate
   Ship only if coverage ≥ 0.70 and precision/recall stable within 0.05.

---

Copy-paste eval harness snippet

def eval_guardrails(question, retrieved, gold):
    ds_qr = deltaS(question, retrieved)
    ds_rg = deltaS(retrieved, gold)

    assert ds_qr <= 0.45, "ΔS drift detected"
    assert coverage(retrieved, gold) >= 0.70, "Coverage too low"
    assert lambda_state(retrieved) in {"→","←","<>"} , "λ divergent"

    return {
        "ΔS_qr": ds_qr,
        "ΔS_rg": ds_rg,
        "coverage": coverage(retrieved, gold),
        "λ": lambda_state(retrieved)
    }

---

Diagnostic probes

• Re-run variance test: run eval 5 times, log precision/recall. Stddev >0.05 → unstable harness.
• Anchor comparison: compare ΔS to gold anchor vs decoy. If both similar, re-embed.
• Deploy warm-up: log VECTOR_READY and index hash before eval.

---

🔗 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 + ”
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	Canonical framework entry point	View
Problem Map	Diagnostic map and navigation hub	View
Tension Universe Experiments	MVP experiment field	View
Recognition	Where WFGY is referenced or adopted	View
AI Guide	Anti-hallucination reading protocol for tools	View

If this repository helps, starring it improves discovery for other builders.