Eval Observability — λ_observe

🧭 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.

Evaluation disclaimer (λ_observe)
λ_observe is a heuristic stability signal defined inside the WFGY framework.
It helps you notice drift and volatility but is not a formal guarantee of long term robustness.

A core probe for evaluating semantic convergence across multiple seeds, paraphrases, and retrieval variations.
While ΔS measures semantic distance, λ_observe captures stability vs divergence of reasoning paths.

Why λ_observe matters

Detect fragile reasoning: Even when ΔS looks safe, λ divergence indicates unstable chains.
Identify paraphrase sensitivity: If λ flips across harmless rewordings, the system is brittle.
Audit retrieval randomness: Different seeds producing opposite λ signals reveal weak schema.
Ensure eval reproducibility: Stable λ means tests repeat reliably under small perturbations.

λ state encoding

Symbol	Meaning	Example failure
→	Forward convergence, stable path	Same citations and reasoning across paraphrases
←	Backward collapse, early abort	Tool call retries, empty citations
<>	Split state, partial divergence	One paraphrase cites correct snippet, others miss
×	Total collapse	Random answers, no citation alignment

Acceptance targets

Convergence rate ≥ 0.80 across 3 paraphrases × 2 seeds.
No × states tolerated in gold-set eval.
Split states (<>): ≤ 10% of test cases acceptable.
Forward (→) must dominate stable runs.

Evaluation workflow

Run triple paraphrase probe
Ask the same question three ways. Collect λ states.
Repeat with two seeds
Track variance.
Roll-up stats
Compute convergence ratio, collapse frequency, divergence rate.
Escalation
If λ <0.80 or × >0%, run root-cause: schema audit, retriever split, prompt ordering.

Example probe schema

{
  "query_id": "Q42",
  "runs": [
    {"paraphrase": 1, "seed": 123, "λ": "→"},
    {"paraphrase": 2, "seed": 123, "λ": "→"},
    {"paraphrase": 3, "seed": 123, "λ": "<>"},
    {"paraphrase": 1, "seed": 456, "λ": "→"},
    {"paraphrase": 2, "seed": 456, "λ": "×"},
    {"paraphrase": 3, "seed": 456, "λ": "→"}
  ]
}

Common pitfalls

Only measuring ΔS → misses hidden divergence.
Seed-fixed eval → looks stable but fragile in production.
Ignoring split states → small divergence often grows into collapse.
No per-query logs → averages hide catastrophic single failures.

Reporting recommendations

λ distribution table: % of →, ←, <>, ×.
Convergence trend: chart over time by eval batch.
Drift alerts: trigger if convergence <0.80 or × appears.
Correlation: track ΔS vs λ to spot mixed failures.

🔗 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.

5.5 KiB Raw Blame History Unescape Escape