Dialogue testing becomes unreliable when every statement is treated as an objective database query. Characters conceal things, make mistakes and draw conclusions. A careful test tracks the relationship between information, belief and speech.
Distinguish truth, belief and performance
World truth describes what actually happened under the story's current rules. Belief describes what a character considers true. A spoken claim describes what they choose to communicate. These records can legitimately disagree. A frightened witness might repeat a rumour; an antagonist might tell a precise lie; an unreliable narrator might describe a remembered event incorrectly.
Start a review by classifying the line. Does it assert direct observation, second-hand information, an inference or a deliberate performance? The wording matters. “I saw you give Mara the key” has a stronger information requirement than “Mara probably has it.” Do not flatten both into the same possession assertion.
Trace the source of the information
For each sensitive fact, identify how the speaker could learn it on the selected path. Sources may include witnessing an action, hearing another character, reading a document or receiving a message through a game system. The source needs to occur before the line, not merely somewhere in the script.
A simple knowledge ledger records the fact, recipient, acquisition event and relevant condition. It can remain an editorial document rather than a new runtime subsystem. The purpose is to make assumptions inspectable. If the team intentionally relies on off-screen communication, record that convention and its limits.
| Fact | Character | Possible source | Path status |
|---|---|---|---|
| Mara owns the key | Mara | Receives the key | Observed |
| Mara owns the key | Gatekeeper | Messenger arrives | Not observed |
| Gate is locked | Player character | Inspects gate | Optional |
Ask AI for provenance, not mind-reading
A model can help extract claims from a dialogue export and propose which facts they depend on. Give it the scene order and relevant path conditions. Ask it to quote the information source for each attribution. If no source appears in the supplied material, the correct output is “source not found in this excerpt,” not “the character cannot know.”
That wording matters because the excerpt may be incomplete. A model's context window, retrieval setup or input filtering can omit the scene that resolves the concern. Keep the material supplied to the model with the review record, subject to the project's confidentiality rules. Otherwise another reviewer cannot tell whether the warning came from the story or the selection of evidence.
Preserve intentional uncertainty
Do not automatically replace surprising dialogue with exposition. An unsettlingly well-informed gatekeeper may be an intentional clue. The review should ask whether the effect is deliberate, supported by the narrative plan and consistent across comparable paths. A writer can confirm that intent without revealing the answer to the player.
Maintain an exception note for approved ambiguity, including the scope and the authoring rationale. Avoid a blanket instruction to ignore all lines from an unreliable character. Even an unreliable narrator has a designed pattern, and an accidental reference to a mutually exclusive event can still damage that pattern.
Check timing and address
Knowledge can be correct eventually but wrong now. A messenger may arrive after the disputed conversation rather than before it. A recalled event may belong to another companion's route. A pronoun or form of address can reveal that the wrong variation played even when the underlying factual statement is broad enough to fit.
Localisation adds another review layer. A translated line may change uncertainty into certainty, remove a conditional phrase or identify an object more specifically than the source. Check stable line identifiers and the approved source meaning with localisation reviewers; a general model's back-translation is not sufficient evidence on its own.
Report the gap without deciding the rewrite
A useful issue says that the gatekeeper claims direct observation on a path where the handover was private and the known messenger scene did not occur. It then asks whether an information event is missing, the line is misguarded or the apparent knowledge is intentional. That report gives the writer options without prescribing unnecessary new dialogue.
Include the acquisition events that were checked, not just a statement that no source exists. A negative claim is only as strong as the search behind it. Where the evidence remains partial, mark the finding as a continuity question and explain what additional context would resolve it.
After an editorial decision, retest the line selection and nearby variations. A rewrite may fix the wording but leave the wrong character response or journal entry in place. Character knowledge often travels through several presentation systems, and the player experiences their combined effect.