D/Discover AI Now
Home / Compare / Evidence Boundary Framework
Decision framework

Point AI at the right evidence boundary before asking it for an answer.

The first tool decision should not be “which model is smartest?” It should be “where does the truth for this task live?”

Decision matrix

Truth lives inYour closed source set

PDFs, policies, meeting notes, internal documents.

Start with NotebookLM / bounded document workflow
Truth lives inThe current public web

Recent announcements, market facts, product changes, public evidence.

Start with a source-oriented web research workflow
Truth lives inBoth internal + external

Internal policy compared against current public evidence.

Use a two-stage research workflow
Truth requiresExecution

Code behavior, SQL output, API behavior, calculations or reproducible tests.

Verify by running the system, not by asking the model again
Truth requiresHuman authority

Policy approval, business intent, ownership, legal interpretation or unpublished decisions.

Insert explicit human approval gates
Truth is unclearDo not synthesize yet

First identify missing evidence and conflicting sources.

Use evidence-first research prompts

Strict source-verification prompt

Answer only from the evidence boundary defined below. QUESTION [QUESTION] ALLOWED EVIDENCE [LIST DOCUMENTS / URLS / SYSTEM OUTPUTS / DATASETS] BOUNDARY - Do not use outside knowledge unless explicitly allowed. - If the answer is not supported by the allowed evidence, say "Not established by the supplied evidence." - For every consequential claim, provide the source and exact section/page/record when available. - Distinguish direct evidence from inference. - If sources conflict, show both positions and explain the conflict. - Do not assign a numeric confidence score unless a scoring method is explicitly supplied. QUOTE-CHECK For each key conclusion: 1. show the supporting source reference, 2. provide a short supporting excerpt or exact record locator when permitted, 3. explain how that evidence supports the conclusion, 4. flag any gap between the evidence and the wording of the claim. OUTPUT 1. Answer 2. Evidence map 3. Conflicts 4. Unsupported / unknown items 5. Verification steps
Verify an AI answer

Why this matters

Tool choice changes the failure mode. A closed-document task can be weakened by unnecessary web material. A current-events question can be weakened by a static document-only workflow. And questions about actual software behavior often require execution rather than another round of language-model reasoning.

Verification funnel

01Define the evidence boundary

Internal only, public only, mixed, executable system or human authority.

02Choose the retrieval path

Use a tool that can access the evidence you actually trust.

03Force traceability

Require source references for consequential claims and separate inference from direct evidence.

04Quote-check a sample

Manually inspect several important claims against the original source.

05Escalate unknowns

If evidence cannot establish the answer, route the question to a human or executable verification path.

Avoid false precisionNumeric “confidence scores” generated by a model can look rigorous without being calibrated. Prefer explicit evidence status: supported, inferred, conflicting or not established.