Verification, Citations, and Abstention

#ai #verification #citations #uncertainty

Verification, Citations, and Abstention

Fact-check status

Reviewed claim by claim on 2026-09-02 against the primary or official sources identified inline as S1 and S2. Recommendations are design guidance unless a cited standard makes them normative.

Key jargon

Term	Plain-language meaning
Verification	Checking a claim against evidence or an executable test. S1
Citation	A pointer to the source supporting a specific claim. S1
Abstention	Explicitly declining to assert an answer when evidence is insufficient. S1
Calibration	How well expressed confidence matches actual correctness. S1

Key concepts

A citation is useful only when the cited source directly supports the nearby claim. S1
A reliable workflow permits uncertainty and abstention instead of rewarding confident guesses. S1

Concept map

flowchart LR
    A["Identify important claims"] --> B["Trace each to evidence"]
    B --> C["Test where possible"]
    C --> D["Accept correct or abstain"]

Diagram semantics verified against S1 and S2 on 2026-09-02.

Verification ladder

Risk	Appropriate check
Low-stakes wording	Human read-through S1
Structured extraction	Schema and sample comparison S1
Factual research	Open cited sources and verify claim support S1
Calculation	Recompute with deterministic code or calculator S1
Code	Review, tests, lint, and isolated execution S1
High-stakes advice	Qualified human and current authoritative guidance S1
External action	Preview, permission check, confirmation, and audit record S1

A citation is useful only if it exists, is authoritative enough, is current enough, and directly supports the nearby claim. Citation quantity is not evidence quality. S1

Abstention contract

Tell the system when to say: S1

“The supplied evidence does not establish this.” S1
“These sources conflict.” S1
“This changed recently and needs verification.” S1
“I cannot safely perform that action without approval.” S1

Exercise

Ask for five claims about a topic you know. Require claim-level citations, then open each source and label it supported, partially supported, or unsupported.