# The FDE Evidence Ladder — Self-Assessment (v1.0)

By Adam Boudjemaa · https://adam-boudjemaa.com/fde-evidence-ladder · CC BY 4.0

Score your AI system by finding the **highest rung for which you have the verifiable
artifact — and have every rung below it too.** A rung is earned by an artifact you can
show, not a claim you can make.

| Rung | You have earned it when… | Artifact to point at | ✅ ? |
|---|---|---|---|
| 0 · Demo-grade | It produces plausible answers on a clean sample | A working demo | ☐ |
| 1 · Grounded retrieval | Answers cite retrieved sources and retrieval quality is measured | RAG + retrieval eval (recall/precision) | ☐ |
| **2 · Verified citation (byte-level)** ◆ | Every claim opens to an exact source span | Byte-level citation links | ☐ |
| **3 · Faithfulness-checked (NLI)** ◆ | The claim is *entailed* by the cited span | NLI check blocking contradiction/neutral | ☐ |
| **4 · Numerically consistent** ◆ | Every figure reconciles against source numbers | Numeric-consistency check | ☐ |
| **5 · Abstention-capable** ◆ | It declines/escalates when evidence is thin | Evidence-gated abstention path | ☐ |
| 6 · Evaluated & red-teamed | A regression eval + adversarial coverage gate changes | Eval harness + injection tests | ☐ |
| 7 · Governed & operable | Scoped access, observability, cost caps, recovery paths | Access control, audit, cost caps, escalation | ☐ |
| 8 · Regulator-defensible & owned | Controls map to a recognized structure; one owner | NIST AI RMF / ISO 42001 mapping + named owner | ☐ |

◆ = the distinctive verification spine (Rungs 2–5).

**Your score: ___ / 8**

Reading it:
- **0–1**: a promising demo. The prototype-to-production gap is ahead of you.
- **2–5**: you are doing the verification work most teams skip. This is where trust is earned.
- **6–8**: production-grade and, at 8, defensible in a regulated setting.

Honest note: rungs 1 and 6–8 are standard engineering practice. Rungs 2–5 are the part that
separates a system you can trust with a real decision from one you can only demo.