{
  "name": "The FDE Evidence Ladder",
  "version": "1.0",
  "published": "2026-07-21",
  "author": "Adam Boudjemaa",
  "url": "https://adam-boudjemaa.com/fde-evidence-ladder",
  "license": "CC BY 4.0",
  "definition": "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.",
  "novelty_statement": "Most individual rungs are known engineering. What is claimed as original is (1) treating evidence strength as a graded ladder rather than a binary, (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. The synthesis and the proof are claimed, not the primitives.",
  "rungs": [
    { "n": 0, "name": "Demo-grade", "tier": "standard", "claim": "Plausible answers, no grounding guarantee.", "artifact": "A working demo on a clean, curated sample." },
    { "n": 1, "name": "Grounded retrieval", "tier": "standard", "claim": "Answers are grounded in retrieved sources, and retrieval quality is measured.", "artifact": "A RAG pipeline plus a retrieval eval (recall/precision on a reference set)." },
    { "n": 2, "name": "Verified citation (byte-level)", "tier": "distinctive", "claim": "Every claim maps to an exact, retrievable source span the reader can open.", "artifact": "Byte-level citation links, not a vague 'sources' footer." },
    { "n": 3, "name": "Faithfulness-checked (NLI)", "tier": "distinctive", "claim": "The claim is entailed by the cited span, not merely near it.", "artifact": "A natural-language-inference check that blocks contradiction and neutral matches." },
    { "n": 4, "name": "Numerically consistent", "tier": "distinctive", "claim": "Every figure is reconciled against the source numbers.", "artifact": "A numeric-consistency check." },
    { "n": 5, "name": "Abstention-capable", "tier": "distinctive", "claim": "On insufficient or conflicting evidence, the system declines or escalates instead of guessing.", "artifact": "An evidence-gated abstention path." },
    { "n": 6, "name": "Evaluated & red-teamed", "tier": "standard", "claim": "An offline eval suite and a reference dataset gate regressions; adversarial cases are covered.", "artifact": "A regression eval harness plus prompt-injection / adversarial coverage." },
    { "n": 7, "name": "Governed & operable", "tier": "standard", "claim": "Access is scoped, the system is observable, cost is capped, and failures have a defined recovery path.", "artifact": "Access control + data-scope limiting, audit trail, cost caps / model routing, human escalation." },
    { "n": 8, "name": "Regulator-defensible & owned", "tier": "standard", "claim": "Controls map to recognized structures and one accountable owner stands behind the deployment.", "artifact": "Controls mapped to structures like the NIST AI RMF or ISO/IEC 42001, plus disclosure and a named owner." }
  ],
  "scoring": {
    "method": "Assign the highest rung for which the system has the verifiable artifact for that rung AND every rung below it.",
    "range": "0-8"
  }
}