The Best Query Is the One the User Never Knew to Ask

TL;DR for operators

A retrieval system can answer only the questions a user thinks to ask. That becomes a design constraint when the user’s main problem is not missing information but missing awareness that information is missing.

VeriForge addresses this by giving AI initiative over discovery while leaving the writer in control of synthesis. In a study of 12 writers, blind-spot alerting averaged 6.33 with VeriForge versus 3.00 with a strengthened baseline, and 11 of 12 participants rated VeriForge higher. Creativity support, exploration behavior, and expert-rated perceived domain competence also improved after statistical correction.¹

The comparison is especially relevant for product teams because both systems used the same graph-based retrieval backend, Semantic Frame, and source-grounded responses. The principal difference was the interaction architecture around that retrieval.

For knowledge-work software, the practical implication is not that every assistant should become more intrusive. It is that product teams should decide which stage of the work deserves machine initiative. Discovery of overlooked evidence may be a better candidate for proactive automation than synthesis of the user’s final judgment.

The caveat is equally operational. Source visibility did not reliably produce source checking: only 1 of 12 participants detected a deliberately mismatched source passage and Knowledge Card. Provenance can support verification, but its presence can also become a shortcut for trust.

Users cannot search for what they do not know is missing

Most AI research interfaces assume a recognizable sequence: the user encounters a problem, formulates a question, retrieves information, and applies it.

The sequence breaks when the second step never happens.

A fiction writer describing unfamiliar swordsmanship, medicine, engineering, regulation, or organizational practice may know enough to continue writing while still missing a constraint that would materially change the result. The user does not experience this as an unanswered question. No query is formulated because no knowledge gap has been recognized.

Sun and colleagues call these latent knowledge gaps and build VeriForge around them.¹ The underlying product problem extends beyond fiction: reactive assistance is structurally weak wherever users must first know enough about a domain to ask the right question.

VeriForge changes who initiates that discovery. While the writer drafts, the system proactively highlights passages that may conceal domain gaps. Those highlights can produce source-anchored Knowledge Cards, which can then be arranged and connected in a spatial Knowledge Canvas. A graph-based RAG system supports retrieval, while a Semantic Frame feeds information from the prose and canvas back into subsequent retrieval.

The writer still decides what belongs in the story.

That allocation is the paper’s central design choice: AI searches for potentially missing knowledge; the human performs narrative synthesis.

The result is about interaction design, not a stronger retrieval engine

It would be easy to attribute VeriForge’s gains to having richer retrieval. The experiment was designed to make that interpretation harder.

The strengthened baseline already included the text editor, conversational querying, spatial canvas, graph-based retrieval, Semantic Frame, and source-grounded answers. VeriForge added the mechanisms that changed initiative and sensemaking: proactive highlights, highlight-derived Knowledge Cards, tighter coupling between prose and cards, editor-derived canvas nodes, and AI-reasoned connections.

This makes the blind-spot result particularly informative. Participants rated VeriForge at 6.33 versus 3.00 for blind-spot alerting, with a large corrected effect ($r=.885$, $q=.016$). Eleven of twelve participants preferred VeriForge on this measure.

Behavior moved as well. Participants issued an average of 6.92 queries with VeriForge versus 4.00 with the baseline, while canvas node and edge counts also increased significantly after correction.

Those numbers do not establish which individual interface component caused the effect; the study does not contain component-level ablations. They do show that holding retrieval infrastructure constant did not hold user behavior constant.

For product teams evaluating RAG systems, that distinction is consequential. Retrieval quality determines what evidence can be returned. Interaction architecture partly determines whether the user reaches the evidence they never thought to request.

Proactive discovery did not require handing over the prose

More system initiative often raises a predictable concern: if the AI does more, does the user become less responsible for the work?

VeriForge separates two forms of “doing more.”

The system takes more initiative in identifying potentially relevant knowledge. It does not take over the final act of expression. In the study, perceived authorial autonomy was identical across both conditions at 6.17.

At the same time, Creativity Support Index scores increased from 69.31 to 81.47 ($q=.017$, $r=.804$), with Exploration and Expressiveness remaining significant after correction. Blind expert assessment also favored VeriForge on all three output dimensions, although only perceived authorial domain competence survived correction: 4.58 versus 3.58 ($q=.015$, $r=.818$).

There was also no significant difference in word count—285.83 versus 258.25—so the additional research and sensemaking activity did not produce evidence of reduced short-session writing output.

The paper therefore supports a more granular automation choice than “assist versus automate.” A system can become more proactive upstream while remaining restrained downstream.

Cognaptus inference

For research, consulting, analysis, design, and other expert-adjacent tools, this suggests a potentially valuable allocation of decision rights:

The evidence does not show that this structure is optimal across those professions. It shows that the allocation worked better than a reactive alternative in the paper’s short unfamiliar-domain writing setting. The broader design principle is an inference worth testing, not an established cross-industry result.

Provenance can become an authority signal instead of a verification tool

VeriForge’s source-linked Knowledge Cards were intended to make retrieved facts traceable. The study then tested whether traceability translated into scrutiny.

Researchers deliberately created a mismatch between a card’s description and its visible source passage. Only 1 of 12 participants noticed.

This is more instructive than a generic warning about hallucination. The failure occurred despite provenance being visible. Users could inspect the evidence; most did not.

The interface may therefore create two effects at once. Provenance makes verification possible, while the appearance of sourcing can itself reassure the user enough that verification feels unnecessary.

For deployments where factual errors carry material cost, source links should not be treated as the end of the governance design. Product teams may need risk-adaptive epistemic friction: interactions that make evidence inspection more likely when a decision is consequential, contested, surprising, or weakly supported.

That could mean requiring source comparison before accepting a high-risk claim, exposing the exact supporting passage more prominently, or distinguishing “source attached” from “claim verified.” These are product implications from the observed failure mode, not interventions tested by the paper.

The evidence stops at cold-start knowledge work

The study provides medium-strength evidence for its immediate setting, but the operating boundary is narrow.

The summative experiment involved 12 writers completing 30-minute story-continuation tasks in two unfamiliar historical martial-arts domains. Participants were young on average, with mean age 21.92 and mean writing experience of 3.18 years. The comparison tests VeriForge as a bundled interaction design, so it cannot tell us whether highlighting, Knowledge Cards, canvas connections, or another component produced most of the gains.

Some attractive findings also deserve different evidentiary weight. One-week recall was higher after VeriForge—8.04 versus 5.75 recalled domain-specific terms, with retention rates of .465 versus .321—but those analyses were exploratory and not FDR-corrected. Overall sensemaking and workload moved favorably but did not survive correction. Natural integration and specificity in expert ratings were also positive trends rather than corrected significant effects.

Long-form professional writing, highly experienced domain experts, and high-stakes settings such as medicine or law remain untested. A professional already possessing deep domain knowledge may have fewer blind spots for the system to discover, while a months-long workflow may make a spatial knowledge canvas either more valuable or more burdensome.

Better assistance may depend on who gets to notice first

VeriForge makes a narrow but useful contribution to the design of AI-assisted knowledge work. It shows that better assistance does not have to begin with a better model, a larger context window, or more autonomous generation.

Sometimes the missing capability sits in the allocation of initiative.

When users can state their needs precisely, reactive retrieval may be sufficient. When they are working outside their expertise, the system may create more value by noticing what deserves investigation before the user knows to ask. The human can still retain the decision that matters most: what the retrieved knowledge means and how it should shape the final work.

The design challenge is to make that proactive discovery genuinely informative without allowing the interface itself to become an undeserved source of authority.

Cognaptus: Automate the Present, Incubate the Future.