// FRAMEWORK · v1.0
The FDE Evidence Ladder
The FDE Evidence Ladder is a maturity model that grades how strongly an AI system can back up what it says. A system climbs from demo-grade output to a regulator-defensible deployment, one rung at a time, where each rung is earned by a verifiable artifact, not a claim.
By Adam Boudjemaa · v1.0 · Published 2026-07-21 · CC BY 4.0
The nine rungs
A system earns a rung with a verifiable artifact, not a promise. Rungs 2 to 5, markedSPINEare the distinctive verification stack. The rest are standard practice, included so the model is complete and honestly labeled.
- Rung 8Regulator-defensible & ownedSTANDARD PRACTICE
Controls map to recognized structures and one accountable owner stands behind the deployment.
Controls mapped to structures like the NIST AI RMF or ISO/IEC 42001, plus disclosure and a named owner.
- Rung 7Governed & operableSTANDARD PRACTICE
Access is scoped, the system is observable, cost is capped, and failures have a defined recovery path.
Access control + data-scope limiting, audit trail, cost caps / model routing, human escalation.
- Rung 6Evaluated & red-teamedSTANDARD PRACTICE
An offline eval suite and a reference dataset gate regressions; adversarial cases are covered.
A regression eval harness plus prompt-injection / adversarial coverage. Standard practice.
- Rung 5Abstention-capableDISTINCTIVE SPINE
On insufficient or conflicting evidence, the system declines or escalates instead of guessing.
An evidence-gated abstention path. The honesty rung: saying 'I don't have enough to answer.'
- Rung 4Numerically consistentDISTINCTIVE SPINE
Every figure is reconciled against the source numbers.
A numeric-consistency check. Decisive in finance, where a wrong number is a real problem.
- Rung 3Faithfulness-checked (NLI)DISTINCTIVE SPINE
The claim is entailed by the cited span, not merely near it.
A natural-language-inference check that blocks contradiction and neutral matches.
- Rung 2Verified citation (byte-level)DISTINCTIVE SPINE
Every claim maps to an exact, retrievable source span the reader can open.
Byte-level citation links, not a vague 'sources' footer. Open the span, see the sentence.
- Rung 1Grounded retrievalSTANDARD PRACTICE
Answers are grounded in retrieved sources, and retrieval quality is measured.
A RAG pipeline plus a retrieval eval (recall/precision on a reference set).
- Rung 0Demo-gradeSTANDARD PRACTICE
Plausible answers, no grounding guarantee.
A working demo on a clean, curated sample. The state most pilots never leave.
What is actually new here (and what isn't)
Most individual rungs are known engineering. What is mine is (1) treating evidence strength as a graded ladder rather than a binary 'we do RAG', (2) shipping the Rung 2-5 verification stack together in a regulated-finance production system, solo and forward-deployed, and (3) tying every rung to a reproducible artifact. I claim the synthesis and the proof, not the primitives.
Related work: others have used a graded "evidence ladder" idea for AI, mostly to stratify how strongly an LLM-as-judge evaluation is supported (E0-E5 style eval-confidence rungs). This framework is different in scope: it grades a whole deployed system, its spine is a shipped citation-to-abstention verification stack, and it is written from a forward-deployed, regulated-finance deployment rather than an eval-methodology lens.
Use it
Score your own system 0 to 8. Everything is free under CC BY 4.0; attribution is appreciated (see "cite" below).
Cite this framework
Boudjemaa, Adam. "The FDE Evidence Ladder," v1.0. adam-boudjemaa.com, 2026-07-21. https://adam-boudjemaa.com/fde-evidence-ladder. Licensed CC BY 4.0.
Changelog
- v1.0 (2026-07-21) - First public release: 9 rungs (0-8), self-assessment, and machine-readable spec.