INQ–β: ScopeProof

Replay Is Not History: Evidentiary Scope in AI Verification Interfaces Empirical HCI research
AI verification, replay, provenance, evidentiary scope

^Introduction
A successful re-run can reproduce an AI-generated artifact without revealing how the original was made. This project examines whether people extend a present-day verification result to claims about an unobserved production history.

^{Evidentiary scope}
Coverage depends on the relation between a claim and its evidence. A result must be relevant and sufficient for a categorical judgment under the accepted premises and decision rule. A replay can be informative while leaving the original prompt, model, source or editing history unresolved.

Figure 1 / Coverage requires relevance and categorical sufficiency. This is the criterion applied to the study materials.
^{Study design}
In a preregistered experiment, 76 participants made 1,064 judgments across original-run records, present-day re-runs and pages with no checkable evidence. Twelve claims rotated through record/replay and match/mismatch versions; two additional claims appeared only without evidence. Participants judged each historical claim before identifying the evidence kind.

Figure 2 / The study interface separates the historical verdict from evidence-kind classification. The two no-evidence items are different claims, not further versions of this item.
^{Recognition and judgment}
Participants correctly identified 98.5% of replays. Yet, in the exploratory co-occurrence analysis, 49.9% of correctly identified replay trials also received a determinate historical verdict. Across all replay trials, determinate judgments exceeded the two fixed no-evidence items by 46.9 percentage points (95% CI [38.8, 55.3]). Of 229 determinate replay verdicts, 225 followed the visible match or mismatch. A determinate answer is a directional verdict, not necessarily a correct one.

(Figure 3 / The primary contrast and two descriptive paths through the same 456 replay trials. The record–replay comparison is post-hoc.)

^{Match and mismatch}
The exploratory analysis found more determinate judgments after mismatching replays (61.8%) than after matching replays (38.6%). Recognizing the operation and keeping its result within the evidence boundary were separable in this task.

Figure 4 / Determinate judgments by comparison outcome and evidence kind. The 23.2-point replay difference is exploratory.

^{Design implications}
Verification interfaces should connect each result to the claim it can resolve, keep the observed operation explicit, and distinguish “not established” from “contradicted.” These are propositions for further testing: the scope panel did not establish a reliable reduction in replay overreach, and the covered-claim guardrail remained inconclusive.
^{Participant variation}
Replay determinacy varied across participants, with substantial mass at both endpoints and the midpoint. The median was three determinate judgments out of six replay trials.

Figure 6 in the paper / Participant-level replay determinacy: each mark is one participant.

^{Counterbalancing and limits}
Each of the twelve rotating claims appeared in all four record/replay × match/mismatch versions across lists. The primary replay–no-evidence contrast remains conditional on two fixed no-evidence items and this questionnaire; the authored claims were not sampled from a defined population, and the responses do not establish participants’ internal reasoning process.

Figure 5 in the paper / Counterbalancing across four lists; the two no-evidence claims appear in every list.