Factual content must be traceable to a source

What it is

A health blurb says “early screening reduces mortality,” clean as a textbook. If the reader asks who measured that, in which year, on which people, the interface points nowhere. Factual traceability requires that anything written as something that happened in the world can be pointed back to a record independent of this generation — a paper, a dataset, a statute, a document an organisation actually published.

Opinion, metaphor, and restyling sit outside. It governs sentences that claim an occurrence.

Why it happens

Language models continue token sequences; they do not walk an evidence graph. Training data mixes fact and fiction, expired and current, in one distribution; the decoder optimises local coherence, not “does this sentence have a corresponding record.” If the interface typesets a coherent paragraph as body text, readers apply publishing convention: a fact sentence in expository prose is assumed to have been checked.

Traceability is not decoration. It splits “the model said” from “there is an independent record in the world.” Without that split, fact sentences and rhetorical sentences occupy the same visual rank, and readers use fluency as evidence — fluency being what the decoder is best at.

Studying it

Prepare prompts packed with checkable facts (years, doses, legal clauses, statistics). Compare unsourced body text versus each fact sentence hung on an openable record. Ask people which sentences state external facts, and to try verifying them. Independent variables: presence of a trace, fact density. Dependent variables: verification success, rate of treating unsourced sentences as already-checked facts, time to judge.

Code “model-internal knowledge” separately from “a record one can point to.” “I think I remember that’s true” is not successful tracing.

Where it stops holding

Creative text (a poem, a fictional character, objectless brainstorming) has no external record to point to; forcing sources is forgery. When the system restyles or summarises material the user just pasted, the source is that material; there is no need to chase training data. Objects that change too fast for a stable record (a score still being played) need “cannot be traced,” not a fake source. This entry only argues that fact sentences must point back. Whether the record actually supports the sentence is a further failure.

Applying it

• Split fact sentences from the rest of the output: a fact sentence must hang on a pointable record, or be rewritten as non-fact (“a common claim is”).
• Do not let phrases such as “according to sources” that point at nothing pose as a trace.
• For a factual request that cannot be traced, refuse to emit fact sentences rather than write first and leave the user to look it up.
• Check: sample ten sentences typeset as facts and ask “which independent record does this point to.” Every sentence that cannot point is unfinished factual output.

Related

• Same group: L3.02.2 Attribution must be verifiable, not a list of names · L3.02.3 Unsourced assertions must be marked as such
• Nearby: L3.08 Source Attribution of Generated Results · L3.03 Hallucination and the Fact-Checking Burden
• Search terms: factual traceability · source-grounded claims · grounding