LLM Evaluation Integrity

Eval rules that produced 55+ auditable runs and zero retracted numbers.

Re-derive every number twice, check confounds before crediting a change, compare on identical items, handle contamination, store complete outputs, keep metric definitions stable, and pick the cheapest decisive experiment.

Why it's different

A 0.74 → 0.98 parse rate jump that wasn't the retry loop
Same-item comparisons only
Every rule tied to a real incident

Try asking

“report eval results”“model or prompt comparison”“ablation study”

Details

Works with: Claude.ai, Claude Desktop, Claude Code, Codex and other SKILL.md agents
Needs: None
Pack: LLM Evaluation & Fact-Checking (€39)
Also covers: llm evaluation, model evaluation best practices, AI benchmark integrity

Same pack

Also in LLM Evaluation & Fact-Checking

LLM Output Validation

Instructions don't bind models. Checks in code do.

No setup neededView →

Fact-Check Before Sending

Catch the wrong number, name or organisation before it reaches an inbox.

No setup neededView →