Chosen
The direction a person actively pursues from the current answer and context.
WP-02 · Next-Question Ranking
A Prudent Question
Teaching machines what to question next
Bounded choice, expected information gain, and the Curiosity Graph
August 2026
Abstract
Language models are trained to answer assigned problems. WHY studies the decision before the next problem is assigned: what deserves investigation now? A bounded slate turns ordinary exploration into comparative evidence. One option is pursued; two visible alternatives are left behind. WHY Pro is designed to add explicit expert judgments about what should be investigated next. WP-02 asks whether those signals can train a ranking layer that learns faster than likelihood alone, transfers across people and subjects, and improves what users discover. The deployed product is the baseline. The expert corpus and learned selector remain a falsifiable research program. Read the complete whitepaper.
The Instrument
1.1 One Answer, Three Questions
For each question, WHY returns one compact answer and exactly three possible directions. The interface deliberately bounds the choice set: the user can select one question or leave.
Experts teach machines how to solve problems. WHY studies how to teach machines what to question next.
Not selected
The two alternatives presented beside the choice. Exposure is observable; conscious rejection is not assumed.
Outside option
A rendered slate that receives no selection. It must be measured separately from a visitor who leaves before seeing one.
Consequence
Whether the selected path loaded, continued, deepened, corrected an answer, or brought the person back.
1.2 Why the Alternatives Matter
A selection without its choice set cannot explain what won. The unselected questions are not treated as independent votes or proof that they were read. They are the contemporaneous alternatives needed to evaluate the choice comparatively.
Consumer paths reveal where unprompted curiosity goes. A separate, explicit WHY Pro cohort is intended to evaluate where disciplined investigation should go. The research question is whether the difference between those signals improves training and evaluation—not whether popularity can substitute for expertise.
observation = (context, slate, order, selection | outside option, consequence)
The Method
2.1 Generate, Validate, Select
Layer 4
Learn
Update the ranking model from exposure, selection, continuation, depth, return, and correction.
Layer 3
Select the Slate
Score every feasible three-question board and keep the one with the greatest expected learning value.
Layer 2
Validate
Require answer anchoring, relevance, novelty, factual support, and meaningful distance between questions.
Layer 1
Generate Candidates
Ask a foundation model for more possible questions than the three the interface can show.
2.2 The Selection Objective
Choosing the three most likely questions can maximize immediate selection while teaching the system very little. Maximizing choice entropy also fails: it cannot distinguish useful uncertainty from genuine indifference and may reward three confusing near-equals.
WP-02 therefore proposes expected information gain: select a slate whose surviving preference hypotheses would rank differently, subject to hard quality constraints.
S* = argmax I(parameters ; outcome | S, context, history) subject to: every question clears a relevance floor the questions are not redundant at least one direction meaningfully diverges a deterministic fallback always returns a usable board
With twelve valid candidates, WHY can evaluate all C(12,3) = 220 possible boards. The combinatorics are small enough to exhaust rather than approximate.
2.3 Curiosity Graph and Rejection Signal
The Curiosity Graph is the complete network of canonical questions and paths. The rejection signal is the comparative evidence inside it: which presented alternatives repeatedly lose to which selected directions, in what context and position.
Keeping the names separate matters. The graph represents the system. Rejection describes one potentially distinctive signal—not a claim that every unselected question was consciously read and refused.
What Exists Today
3.1 The Live Baseline
Why.com already provides the instrument: one concise answer, three validated questions, stable directional roles, path-aware context, local curiosity memory, and outcome telemetry. The current production system selects three questions directly; it does not yet maintain the posterior or population-level Curiosity Graph proposed by WP-02.
Live
Candidate generation, factual and semantic validation, three-question presentation, path continuity, and local selected-versus-unselected outcomes.
Instrument next
Qualified slate exposure, outside-option measurement, randomized research cohorts, policy version, and reliable selection consequences.
Research build
A larger candidate pool, posterior preference model, exhaustive slate selector, and calibrated uncertainty estimates.
Graph build
Canonical question identities, exposure-corrected population edges, personal residuals, and privacy-preserving aggregation.
3.2 One Platform, Separate Evidence
Why.com is the consumer instrument for authentic curiosity paths. WHY Pro is the proposed opt-in expert instrument for deliberate next-question judgment. WHY AI is the generation, validation, and ranking layer. WP-02 tests whether those sources improve learned selection. WHY Desktop is the private personal layer; its memory is not population training data by default.
What Would Prove It
4.1 Pre-Declared Tests
- Information: Does the proposed selector reduce posterior uncertainty more per turn than a likelihood-ranked baseline?
- Learning: Does it predict held-out selections with fewer observations?
- Transfer: Does the learned signal improve prediction for unseen people and unseen subjects?
- Outcome: Do better-ranked questions improve discovery, correction, completion, or accepted work—not merely clicks?
- Persistence: At matched sample size, do unselected alternatives predict distal choices better than selected questions alone?
4.2 The Result That Would Matter
The strongest result would not be that WHY predicts the next click. It would be that comparative evidence from bounded question slates improves what people discover later—and that the effect transfers beyond the user or subject that produced it.
The moat test
If the learned signal does not transfer across people or subjects, the Curiosity Graph is a personalization feature rather than a population-level asset.
Limits and Obligations
5.1 Measurement Before Mythology
A displayed question is not necessarily a read question. An unselected question is not necessarily a conscious rejection. A visitor who leaves before a slate renders is not an outside-option observation. The research must record these states separately.
5.2 Generation Bounds Discovery
The system can only learn among possibilities its generator imagines. A divergence requirement helps, but cannot reveal an entire direction that was never generated. Generator diversity and reserved exploration must therefore be measured rather than assumed.
5.3 Ranking Can Narrow the World
A globally popular question can receive more exposure, win more comparisons, and become still more visible. Position randomization, logged propensities, bounded population influence, and graph-entropy monitoring are required to prevent a rich-get-richer loop from becoming a curiosity monoculture.
5.4 Rejection Is Sensitive
Patterns of avoidance may reveal more than selections. Personal decline histories should remain private by default; aggregation requires clear consent, minimum cohort protections, and no raw personal path attached to a population edge.
Conclusion
Experts teach machines how to solve assigned problems. WHY is building the questioning layer: a system for learning which problem deserves attention next.
WP-02 does not claim that the learned system is complete. It defines the instrument, the proposed selector, the evidence required, and the observations that would prove the thesis wrong.
AI learned how to answer. WHY is testing how to teach it what to question next.