WFGY/ProblemMap/GlobalFixMap/Eval_Observability/variance_and_drift.md

Eval Observability — Variance and Drift

🧭 Quick Return to Map

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

• Eval_Observability — evaluation metrics and system observability
• 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.

Variance and drift checks detect when evaluation scores are unstable across runs or when semantic meaning slowly shifts without clear boundary failures.
These probes prevent "false confidence" in benchmarks by catching hidden instability.

---

Why variance and drift matter

• Variance: Scores fluctuate heavily depending on seed, paraphrase, or retriever order. Averages hide the volatility.
• Drift: Performance declines slowly across sessions, data refreshes, or version bumps. Looks fine short-term but collapses long-term.
• Silent regressions: Systems pass local tests but fail in production due to unmonitored entropy rise.

---

Acceptance targets

• Variance (σ/μ) ≤ 0.15 across 3 seeds and 3 paraphrases.
• Drift slope: Δscore per batch ≤ 0.02 absolute over 5+ eval windows.
• No monotonic downward slope longer than 3 consecutive windows.
• Drift alerts fire if ΔS average increases ≥ 0.10 compared to gold anchors.

---

Detection workflow

1. Collect runs across seeds

   • At least 3 seeds, 3 paraphrases.
   • Log ΔS, λ, coverage, citations.

2. Compute variance

   • Calculate σ/μ for each metric.
   • High variance = unstable eval → rerun with schema locks.

3. Track drift over time

   • Compare eval batch N vs N-1.
   • Plot moving average.
   • Alert if slope exceeds tolerance.

4. Root-cause analysis

   • If variance high → check retriever metrics, random seeding, rerankers.
   • If drift detected → audit embeddings, re-chunk, verify data refresh.

---

Common pitfalls

• Single-run evals: Hides high variance. Always run multi-seed.
• Averages without spread: Mean looks fine, variance reveals collapse.
• Ignoring slow drift: Short tests OK, but 1–2 weeks later accuracy dies.
• Cross-store drift: One vector DB stable, another drifts. Must track both.

---

Example reporting schema

{
  "metric": "ΔS",
  "seed_runs": [0.38, 0.42, 0.44],
  "variance_ratio": 0.14,
  "drift_slope": +0.03,
  "alert": true
}

---

Fix modules to open

• Retriever instability → Rerankers
• Embedding mismatch → embedding-vs-semantic.md
• Fragmentation drift → vectorstore-fragmentation.md
• Prompt instability → context-drift.md

---

🔗 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	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.