{"path":"research/curiosity-as-growth-fuel.md","content":"# Curiosity: The Missing Term in the Cost Model\n\n*Aug 12, 2026. The founder's framing: \"Curiosity is a fuel for Deliberus growth and all indirect growth factors either input/before or after/output from Deliberus.\" The literature agrees, and then says something stronger and less comfortable: curiosity is not only fuel. It is the variable that decides whether comprehension infrastructure converges people or polarizes them. Deliberus is comprehension infrastructure. So this is not a growth topic. It is a precondition on the wager.*\n\n---\n\n## 1. The answer is bigger than the question\n\nThe question was about growth: what brings people to Deliberus, and what the system sends back out. That question has good answers, and they are in §5 and §6.\n\nBut the strongest empirical finding in this literature is not about growth at all. Dan Kahan spent years measuring what happens when you give people more capacity to make sense of evidence on contested questions. The answer is consistently the wrong one:\n\n> \"Whatever measure of critical reasoning we used, we always observed this depressing pattern: The members of the public most able to make sense of scientific evidence are in fact the most polarized.\"\n\nThe most science-literate conservatives were *more* likely to reject anthropogenic climate change than their less-literate peers. High-numeracy partisans misread data that threatens their identity at a higher rate than low-numeracy partisans. Kahan calls it **motivated System 2 reasoning** (MS2R): comprehension capacity does not correct motivated reasoning, it serves it.\n\nDeliberus is comprehension capacity, industrialized. The founding conviction — *Opacity Is a Cost, Not a Mystery* — commits the project to lowering the cost of understanding why another mind holds its position. Kahan's result says that lowering that cost for a motivated reasoner does not produce understanding. It produces a cheaper weapon. The red team already named the mechanism from the other direction and called it **weaponized decomposition**; this is the same seam approached from psychology rather than from adversarial design.\n\nThen Kahan found the exception, and it is the reason this document exists. In the same research program, one disposition ran the other way:\n\n> \"As their science curiosity goes up, the polarizing effects of higher science comprehension dissipate, and people move the same direction on contentious policies like climate change and fracking.\"\n\nNot knowledge. Not numeracy. Not reasoning proficiency. All of those made it worse. **Curiosity made it better**, and it did so by changing what people chose to read: science-curious partisans preferred information that contradicted their political predispositions, at effect sizes of 20 to 44 percentage points. Kahan's own explanation is that the pleasure such people take in being surprised outcompetes the motive to avoid disconfirmation. His phrasing: *\"For them, surprising pieces of evidence are bright shiny objects — they can't help but grab at them.\"*\n\nSo the finding to carry forward is this. **Lowering the cost of understanding changes the rate at which people attempt it. It does not change their motive for attempting it. Curiosity is the motive term, and without it the cost reduction can run backwards.**\n\nThat is a proposed second term on the founding conviction, and §11 puts it to the founder rather than assuming it.\n\n## 2. What the corpus already held, and why that is the interesting part\n\nThe Kahan result is already in this repository. It has been for months:\n\n- `research/trust-economics-and-false-polarization.md` cites it with the cognitive-dualism mechanism.\n- `research/non-zero-sum-economics-and-civilizational-cooperation.md` states the implication outright: *\"a crucial design insight: Deliberus should cultivate curiosity, not just present information.\"*\n- `analysis-and-attunement.md` carries the load-bearing sentence: *\"The cooperation research treats curiosity, not mere analytical skill, as the route around motivated reasoning.\"*\n- `research/epistemic-gamification.md` lists curiosity among the epistemic virtues, unoperationalized.\n- `research/adoption-problem.md` names the gap: argument mapping historically required external motivation \"rather than intrinsically motivated (curiosity, social reward).\"\n\nFive docs, one true sentence each, and nothing built. Across the whole public corpus the word appears 27 times, and the dominant use is as a **label for a progressive-disclosure tier** — casual / curious / expert — which is a UI resolution setting, not a theory of motivation.\n\nThis is the same shape as the audit that produced the founding conviction earlier today: an idea the founder holds as a conviction, sitting in the corpus as a borrowed citation, stated once and never made load-bearing. The recurrence is worth noticing on its own. A project whose purpose is detecting where one idea hides under two names, or hides in plain sight as a passing reference, keeps finding its own strongest claims filed as footnotes.\n\n## 3. The mechanism: what curiosity actually is\n\nFour sources carry the weight.\n\n**Loewenstein (1994), the information-gap account.** Curiosity is \"a form of cognitively induced deprivation that arises from the perception of a gap in knowledge or understanding.\" A gap opens when *a question becomes activated but the answer is not known*. Two consequences matter for design. First, curiosity is a function of **attention**, not of missing information as such: the gap has to be *noticed*. Second, Loewenstein names three factors that intensify it, and each is a design variable rather than a personality trait:\n\n| Factor | Loewenstein's claim | Deliberus surface |\n|---|---|---|\n| **Importance** | Curiosity is stronger about questions that matter to the person | Topic selection, the user's own submitted text |\n| **Salience** | Curiosity rises with how *visible* the gap is made | Sorry markers, the completeness oracle, the open frontier |\n| **Epiphany** | Strongest of all for information that would close **many** activated gaps at once | Nothing computes this yet — see §9.1 |\n\n**Gruber & Ranganath (2019), the PACE framework.** Prediction errors — contextual ones in the hippocampus, informational ones from cognitive conflict in the anterior cingulate — are **appraised** by lateral prefrontal cortex, which asks whether the information could be valuable later. Curiosity is what follows a positive appraisal, and it recruits the dopaminergic circuit to strengthen hippocampal encoding, improving memory for the target *and for incidental material encountered while curious*.\n\nThe load-bearing passage for this project is the failure branch, in the authors' own words:\n\n> \"a prediction error could trigger curiosity and exploration if one feels that one has the capability to resolve the uncertainty (i.e., appraisal of higher coping potential) or it could **trigger anxiety and behavioral inhibition** if one believes that one has no ability to cope with the prediction error (i.e., appraisal of low coping potential).\"\n\nA gap plus the felt capability to close it produces curiosity. **A gap without that capability produces avoidance.** That is a mechanism, from neuroscience, for something dogfood run 1 arrived at from friction and named the *inversion principle*: the machine holds the user's hand through the descent, taps rather than essays. It also constrains honest reporting. Displaying an unsupported value premise to someone with no idea what to do about it is not neutral disclosure. It is a repellent.\n\nAnd the paper's own illustration is almost exactly the sorry-marker scenario, which is worth quoting because it locates the risk in *framing* and not only in affordances:\n\n> \"a student who discovers an information gap as he completes a homework assignment might become curious if he focused on the value of the necessary information on a future test. However, **he could become anxious and procrastinate if he instead interpreted the information gap as a sign of his incompetence** and inability to understand the topic.\"\n\nSame gap, two readings, opposite behaviour. \"Your claim rests on an unsupported value premise\" can land as *here is something worth working out* or as *here is proof you do not understand your own reasoning*, and which one it lands as is decided by wording, placement and tone — not only by whether a button is present. The existing UX directives already lean the right way (the opening experience should feel like help; markers are invitations). This says the lean is load-bearing rather than decorative.\n\n### The finding that changes this document's conclusion\n\nThe abstract does not contain it. The full text does, and it inverts the pessimistic reading of §1.\n\nKahan treats curiosity as a *trait* — a disposition you can measure, and therefore a population you must find. PACE reports that the trait largely reduces to the appraisal step:\n\n> \"individual differences in the strength of appraisal (i.e., how well a person can cope with the uncertainty) explains the individual differences in trait curiosity … 'curious people appear to be curious because they are more likely to **appraise their ability to understand as high**'\"\n\nand, from the experimental work it cites, \"participants with high coping potential show higher curiosity/interest in complex novel products and inventions, suggesting that **appraisal leads to curiosity only if the prediction errors seem manageable**.\"\n\nRead that against what Deliberus is for. The entire function of the system is to raise a reader's felt ability to understand why another mind holds its position — that is the founding conviction restated as a mechanism. If appraised ability-to-understand is what curiosity mostly *is*, then **Deliberus can manufacture a meaningful share of the disposition its own success depends on.** The loop the founder was pointing at is therefore tighter than §1 alone suggests: curiosity is not only the fuel and not only a precondition, it is partly an *output*.\n\nTwo guards keep this from becoming a comfortable story. It is a mechanism claim in a framework paper, not a demonstrated intervention effect, and PACE is explicit that its predictions await testing. And it does not repeal Kahan: a motivated reasoner who appraises their ability to understand as high is exactly the well-informed partisan whose comprehension serves their motive. Raising appraised capability raises *both* curiosity and confident motivated reasoning, so the honest statement is that Deliberus can shift the appraisal term and cannot, by that route alone, shift the motive.\n\n**Kidd & Hayden (2015) plus the learning-progress account.** Infants look away from events that are highly predictable *and* from events that are highly surprising, holding themselves at intermediate information rates. The modern reconciliation is that curiosity tracks **learning progress** — the rate at which prediction error is falling — rather than information quantity. Hence the finding that people \"engage with stimuli that offer easily interpretable information\" rather than the most informative stimuli available, and that curiosity wanes when information arrives too slowly. §8.1 draws the uncomfortable consequence.\n\n**Litman's I/D distinction.** Two kinds of epistemic curiosity with different affect and different consequences:\n\n| | I-type (interest) | D-type (deprivation) |\n|---|---|---|\n| Trigger | Little or no prior knowledge; novelty | Some prior knowledge; feeling close to an answer |\n| Feeling | \"Learning is fun\" — mild, positive | An itch to be scratched — intense, initially negative |\n| Reward | \"Oh, that's interesting\" | \"Ah-ha, now I understand\" |\n| Correlates | Neutral or negative with anxiety and anger | **Positive** with anxiety, depression, anger; low ambiguity tolerance; high need for closure |\n| Strength as a motive | Weaker | **Stronger** |\n\nD-type pulls harder and costs more. §8.2 argues this is an unmade design decision sitting under the residue map.\n\n## 4. Two dyad-level results worth having\n\n**Question-asking increases liking** (Huang, Yeomans, Brooks, Minson & Gino, 2017). Across four studies including speed-dating field data, people who ask more questions — especially **follow-up** questions — are better liked, and speed daters who ask more get more second dates. The mechanism is that questions solicit self-disclosure, which the answerer enjoys and attributes to the asker. People systematically fail to anticipate this effect.\n\nDeliberus mints follow-up questions about people's claims as its core mechanic. Critical questions are literally machine-generated follow-ups. But the bound matters and must be kept: the liking effect **did not transfer to third-party observers**, and the mechanism runs through the *answerer's own* disclosure. A CQ nobody asked me personally is not a compliment. So the finding is not \"CQs make people like Deliberus\"; it is narrower and more useful — *a CQ addressed to my claim, which I then answer, is the shape of interaction the literature says builds relationship.* That is the friends-round path, not the public feed.\n\n**Conversational receptiveness is measurable and improvable** (Yeomans, Minson, Collins, Chen & Gino, 2020). An interpretable algorithm detects, from language alone, whether someone is signalling willingness to engage with an opposing view. Receptive Wikipedia editors receive fewer personal attacks. Speakers misjudge their own receptiveness, which is what makes intervention possible, and a short \"receptiveness recipe\" measurably improves how counterparts rate them.\n\nPer reference-architecture-first: if Deliberus ever measures *how a user writes a disagreement* rather than only what the graph does to it, this instrument exists, is validated, and is interpretable. It is the dyad-rung cousin of the disagreement-preservation instrument, which currently measures the machine's flattening and not the human's posture.\n\n## 5. Input side: curiosity before Deliberus\n\n**Curiosity is the recruitment filter, and it cuts across the communities the outreach frontier has been sorting by.** The beachhead question has been *which community* — EA, rationalist, academic argumentation, AI safety. Kahan's data says the better filter is a **disposition** that appears in all of them and is absent in parts of all of them. A science-curious conservative and a science-curious liberal both chose the disconfirming article. Their less-curious co-partisans both refused. Community predicted the *direction* of bias; curiosity predicted whether bias operated at all.\n\n**Curiosity also predicts who can carry it outward.** Kahan's own recommendation, unprompted by us:\n\n> \"It makes a lot of sense for anyone who wants to promote public engagement with science to try to focus their efforts on culturally diverse, science-curious people because they'll be listened to when they return to their communities and interact with their peers.\"\n\nThat is an outreach strategy stated by the named demand skeptic's own research programme, and it matches the dogfooding-first posture already in force: a small number of genuinely curious close collaborators, chosen for disposition rather than domain credential, is not a compromise on scale. It is what the evidence says the mechanism requires.\n\n**The pull-not-push framing.** The podcast that prompted this — Victoria Griffin's *Writing Woods* episode on curiosity — makes a claim about practice rather than psychology: you cannot force curiosity, you can only create conditions for it, so \"stop pushing yourself toward the work and start creating opportunities for the work to pull you in.\" That is the same instruction the UX directives already give in different words — *the opening experience should feel like help* — and it is the motivational-layer statement of the inversion principle. It earns a place here because it names the design posture the mechanism in §3 requires: an appraisal that says \"I could close this\" is produced by affordances, not by exhortation.\n\n## 6. Output side: Deliberus is a gap generator\n\nThis is the half the founder's framing gets exactly right, and the literature sharpens it into something buildable.\n\nLoewenstein's account says curiosity requires a gap that has been *noticed*. Almost every instrument this project has shipped is, in that vocabulary, a **salience device for information gaps**:\n\n| Deliberus artifact | In Loewenstein's terms |\n|---|---|\n| Sorry marker | An activated question with a declared missing answer |\n| Unanswered critical question | A gap the system generated rather than the user |\n| `unsupported_value_premises` | A gap the oracle *found* and made visible |\n| Sorry frontier / completeness verdict | A map of where the gaps are |\n| Open typed residue | A gap declared irreducible under current moves |\n| `docs/frontier.md` | The project's own gaps, published |\n\nSo Deliberus does not merely consume curiosity. It manufactures the precondition for curiosity at industrial rate, and it does so *about other people's reasoning* — which is precisely the target the cooperation thesis needs and that ordinary information environments never supply. An echo chamber's defining property is that it presents no gaps. A decomposed argument graph is nothing but gaps, arranged.\n\nThe correct name for what the system emits, then, is not \"structure.\" It is **noticed absence**. And the design consequence follows from §3: every gap the system makes salient must arrive with an affordance for closing it, or PACE's appraisal branch converts the same gap into avoidance.\n\n## 7. The loop, and where each joint fails\n\nPut the two halves together and the founder's flywheel has four joints. Each one has a named failure mode, which is the useful part.\n\n```\ncurious person arrives ──→ descends into a claim ──→ gaps made salient\n        ↑                                                    │\n        └──────── shares / invites / returns ←───────────────┘\n```\n\n| Joint | What has to happen | Failure mode | Named by |\n|---|---|---|---|\n| Arrival | A curious person shows up | Recruiting by community rather than disposition brings motivated reasoners with better tools | Kahan (MS2R) |\n| Descent | The gap is appraised as closable | Honest gap display without affordances produces anxiety-inhibition | Gruber & Ranganath (PACE) |\n| Sustain | Learning progress stays perceptible | The deepest nodes have the slowest progress, so curiosity dies exactly where value is highest | learning-progress account (§8.1) |\n| Return | The closed gap creates new gaps | Satisfaction extinguishes the motive that produced it | Loewenstein (§8.3) |\n\nThe loop is real. It is also not self-sustaining at any joint by default, and three of the four failures are invisible from inside the system as currently instrumented.\n\n## 8. Three structural tensions\n\n### 8.1 The value gradient and the curiosity gradient run opposite\n\nCuriosity tracks learning progress and prefers intermediate uncertainty. Deliberus's most valuable locations — deep decomposition, hard termini, the residue frontier — are exactly where progress per unit effort is slowest and uncertainty is highest. The graph's epistemic value gradient therefore runs **against** its curiosity gradient.\n\nThis predicts something checkable: engagement should fall off with decomposition depth faster than difficulty alone explains, and dropout should cluster at the transition from \"answerable CQ\" to \"value premise with no evident next move.\" It also explains a puzzle the corpus already has and had no account for: dogfood run 2 found the hinge score *works but whispers at light engagement*. If depth is where curiosity thins, then the instruments that only pay off at depth will be the ones nobody reaches.\n\nThe mitigation is not to hide depth. It is to make progress *legible* at depth — which is what the completeness verdict and the terminus classification already do, and which argues for surfacing them earlier and louder rather than as expert-tier detail.\n\n### 8.2 The residue map is an I-type artifact recruited for by D-type means\n\nA sorry marker, an unanswered CQ, an unsupported value premise: each is a declared deprivation. That is D-type induction — the stronger motive, and the one that correlates with anxiety, low ambiguity tolerance, and high need for closure.\n\nNow consider what the descent delivers at the bottom. A **typed residue says the question does not resolve.** For I-type curiosity that is a satisfying result: an interesting structure, a well-stated open problem, the compression that the wager treats as the product. For D-type curiosity it is the itch left unscratched by design, delivered to the person least equipped to tolerate it.\n\nThe residue map is the project's falsification scoreboard and, on the wager's own accounting, its most important artifact. It is currently reached by a path that recruits the disposition least likely to value it. That mismatch is not a bug in either half. It is an unmade decision about which curiosity Deliberus is for, and it deserves to be made deliberately rather than by default.\n\nOne piece of evidence sharpens the fork rather than leaving it symmetrical. PACE reports that real-world encyclopedic information-seeking tracks the deprivation side specifically: engagement with Wikipedia articles \"correlated with trait curiosity; specifically, the particular aspect of trait curiosity that is related to closing an information gap (i.e., deprivation sensitivity).\" If the population that actually does volunteer knowledge work is deprivation-loaded, then Deliberus's likely contributors are the ones for whom a typed residue is least satisfying — and the mismatch is not a hypothetical about two equally-likely audiences but a prediction about the one that shows up.\n\n### 8.3 Success extinguishes the motive\n\nLoewenstein's least convenient observation is that curiosity is transient, superficial, and has a \"tendency to disappoint when satisfied.\" Answered questions *lose* their importance — the belief-resolution effect. The intense need to know a piece of gossip vanishes the moment you leave the tabloid rack.\n\nDeliberus's definition of success is closing gaps. So the mechanism that makes the platform valuable destroys the motivation that brought the user to it. Retention is structurally adversarial to purpose.\n\nExcept that this project has an unusual answer already built, and had not noticed it was an answer to this. **Self-similar decomposition means there is no permanent bottom** — every closed gap opens the next layer's gaps, because nothing is ever more than *currently* undecomposed. `depth.md`'s \"No Copout Axioms\" is an ontological commitment. It is *also* an inexhaustible supply of information gaps, which is exactly what Loewenstein's transience problem requires. The principle chosen for epistemic reasons turns out to solve the motivational problem, which is the kind of coincidence worth recording because it will not be the last one.\n\n## 9. What is buildable\n\nRanked by evidence strength over effort.\n\n### 9.1 Epiphany potential: the fan-out the hinge score does not compute\n\nLoewenstein's strongest intensifier is epiphany — one answer closing **many** activated questions. The obvious candidate in this codebase is the hinge score, and reading `deliberus/hinge.py` shows the candidate does not fit, which is the useful finding.\n\n`hinge_scores` clamps each descendant to 1.0 and 0.0 and records how far **the root** moves: `hinge = |σ_root(granted) − σ_root(denied)|`. That is single-conclusion sensitivity — \"how much does *this* conclusion depend on that node?\" Loewenstein's quantity is different: \"how many of the questions *I already have open* would this one answer move?\"\n\nSo the buildable thing is new, not a relabelling: **fan-out hinge.** For a node, compute its hinge against every ancestor root that depends on it, not just one. The corpus-level version ranks nodes by total epiphany potential. The personalized version — closer to Loewenstein's actual claim, since he specifies *questions she is already asking* — ranks against the roots this particular user has engaged with. Cost is bounded the same way the current endpoint is bounded, and it reuses `propagate` unchanged.\n\nPrediction, registered for scoring: high-fan-out nodes out-engage high-single-hinge nodes, and both out-engage random nodes matched for depth.\n\n### 9.2 Affordance audit of every gap the system displays\n\nFrom PACE's appraisal branch. Enumerate every surface where Deliberus shows a user a gap, and check whether the same surface offers a next move. The completeness oracle's `unsupported_value_premises` is the sharpest case: it names a deprivation with no affordance attached. `fully_exposed` at depth 0 is the inverse failure — an affordance-free flatterer. The rule to test: **no gap displayed without a move offered.**\n\n### 9.3 Curiosity-shaped feed calibration\n\nThe feed has eight modes and currently ranks by epistemic need. The Goldilocks result says the same ranking should be filtered by intermediate difficulty for *this* user. That is one clamp on an existing sort, and it is the difference between a work queue and something that pulls.\n\n### 9.4 A read of the friends-round through §4\n\nFor the guided path already planned (extraction page → justice-vs-equality claim → decomposition → fittingness leaf → disagree-and-say-why), the question-asking result says the highest-value moment is not the descent. It is the point where a collaborator *answers a CQ about their own claim* and someone engages with the answer. Instrument that moment.\n\n### 9.5 Receptiveness measurement, if the dyad rung is ever built\n\nDeferred deliberately, because it needs a user writing prose at another user, which does not exist yet. Recorded so it is not re-derived.\n\n## 10. What this does not say\n\nDiscipline, because it would be easy to over-claim from a literature this congenial.\n\n- **Kahan's result is correlational for the observational half.** The experimental half — curious partisans choosing disconfirming articles — is causal for *information choice*, not for belief change or for depolarization over time.\n- **Science curiosity is not epistemic curiosity in general.** Kahan's scale is domain-specific and was hard-won; whether \"curiosity about why someone disagrees with me\" behaves the same way is an open empirical question, and it is the one Deliberus actually needs answered.\n- **Curiosity does not touch interest-conflict.** The founding conviction's second guard applies unchanged: someone who understands you perfectly and wants your land was never a comprehension problem. Curiosity lowers no barrier there.\n- **The PACE mechanism is a framework, not a settled result.** Gruber and Ranganath present testable predictions, and the developmental work extending it is explicitly predictive.\n- **The obvious way to measure curiosity is a metric this literature has already discredited.** Engagement and preference are the cheap proxies, and a Google DeepMind study of real-money group deliberation (N=879, arXiv:2605.14097) found that LLM facilitation \"did not significantly improve group consensus in either study, yet participants consistently preferred facilitated discussion\" — while shifting allocations by up to 5.5 percentage points. So a curiosity instrument built on time-on-descent or self-reported interest would be measuring the variable that diverged from quality in the strongest available experiment. Whatever §11's third decision resolves to, the instrument has to be anchored to something other than how much people liked it.\n- **None of this measures Deliberus.** Every claim above is imported. The project's own curiosity data is currently zero rows.\n\n## 11. For the founder\n\nThree of these are decisions only he can make.\n\n1. **The second term on the founding conviction.** *Opacity Is a Cost, Not a Mystery* is ratified and deliberately unqualified. Kahan's MS2R result suggests it is incomplete rather than wrong: lowering the cost changes the rate of understanding attempts, not the motive behind them. The proposed completion — **opacity is a cost, and curiosity is the willingness to pay it** — is left here rather than written into `vision.md`, because the conviction is hours old, founder-ratified, and the project's most load-bearing public sentence. Ratify, reject, or reword. Note that the appraisal finding makes this less fatalistic than it first reads: if curiosity is largely the appraisal that one *can* understand, then lowering the cost of understanding partly generates the willingness to pay it, and the second term describes a loop rather than an external dependency.\n2. **Which curiosity Deliberus is for** (§8.2). I-type and D-type want different things from a residue. The recruitment path currently induces one and the terminal artifact rewards the other.\n3. **Whether curiosity becomes a measured quantity.** It could be an instrument like the others — gap-to-affordance ratio, descent-depth survival, fan-out engagement. Or it stays a design lens. The instrument version can embarrass us, which by the project's own accounting rule is the argument for it.\n\n---\n\n## Read depth, stated honestly\n\nThe same standard this session demanded of an outside model applies here. **Full text**: Gruber & Ranganath 2019 (PACE), read in full after the founder pointed at it — the first draft of this document had it at abstract depth, which cost it the appraisal-is-trait-curiosity finding above, the paper's own homework-assignment illustration, and the Wikipedia deprivation-sensitivity result. The claim already shipped into `ux-principles.md` and the frontier was checked against the primary text and stands. **Substantial extracts plus abstract**: Loewenstein 1994; Golman & Loewenstein (*Curiosity, Information Gaps, and the Utility of Knowledge*); Kahan et al. 2017; Kidd & Hayden 2015; Litman's I/D material; Huang et al. 2017; Yeomans et al. 2020; the *Trends in Cognitive Sciences* review on curiosity and optimal exploration; Parisi et al. 2026 (the N=879 facilitation study — **upgraded to full text 2026-08-14**; the §10 point stands and gains a caveat, since their consensus measure is agreement between numeric allocation vectors near a ceiling rather than anything about reasoning quality, so it is their *preference-diverges-from-outcome* finding that discredits the cheap proxy, not their consensus null. Full read: `synthesis-build-plan.md`). **Verified in code, not inferred**: `deliberus/hinge.py`, which corrected the §9.1 claim. **Not consulted**: the podcast episode audio; Berlyne 1954/1960 in the original.\n\nThe general lesson is the one this session keeps relearning. An abstract gives you a framework's claim; the full text gives you the sentence that changes what you would build. Three of the most useful findings here were in the body of one twelve-page review that the first pass had summarized from secondary sources.\n\n---\n\n**See also**: [analysis-and-attunement.md](../analysis-and-attunement.md) · [depth.md](../depth.md) · [convergence.md](../convergence.md) · [ux-principles.md](../ux-principles.md) · [trust-economics-and-false-polarization.md](trust-economics-and-false-polarization.md) · [convergence-wager-red-team.md](convergence-wager-red-team.md) · [dogfood-run-1-friction-log.md](dogfood-run-1-friction-log.md) · [adoption-problem.md](adoption-problem.md)\n\n## Sources\n\n- Loewenstein, G. (1994). \"The Psychology of Curiosity: A Review and Reinterpretation.\" *Psychological Bulletin* 116(1), 75–98. [doi:10.1037/0033-2909.116.1.75](https://doi.org/10.1037/0033-2909.116.1.75)\n- Golman, R. & Loewenstein, G. \"Curiosity, Information Gaps, and the Utility of Knowledge.\" [CMU](https://www.cmu.edu/dietrich/sds/docs/golman/golman_loewenstein_curiosity.pdf)\n- Kahan, D.M., Landrum, A., Carpenter, K., Helft, L. & Hall Jamieson, K. (2017). \"Science Curiosity and Political Information Processing.\" *Advances in Political Psychology* 38(S1). [doi:10.1111/pops.12396](https://onlinelibrary.wiley.com/doi/10.1111/pops.12396)\n- Gruber, M.J. & Ranganath, C. (2019). \"How Curiosity Enhances Hippocampus-Dependent Memory: The Prediction, Appraisal, Curiosity, and Exploration (PACE) Framework.\" *Trends in Cognitive Sciences* 23(12), 1014–1025. [doi:10.1016/j.tics.2019.10.003](https://doi.org/10.1016/j.tics.2019.10.003) · [PMC6891259](https://pmc.ncbi.nlm.nih.gov/articles/PMC6891259/)\n- Kidd, C. & Hayden, B.Y. (2015). \"The Psychology and Neuroscience of Curiosity.\" *Neuron* 88(3), 449–460. [PMC4635443](https://pmc.ncbi.nlm.nih.gov/articles/PMC4635443/)\n- Litman, J.A. (2005). \"Curiosity and the pleasures of learning: Wanting and liking new information.\" *Cognition and Emotion* 19(6). Plus Litman (2008), \"Interest and deprivation factors of epistemic curiosity,\" *Personality and Individual Differences*.\n- Huang, K., Yeomans, M., Brooks, A.W., Minson, J. & Gino, F. (2017). \"It Doesn't Hurt to Ask: Question-Asking Increases Liking.\" *Journal of Personality and Social Psychology* 113(3), 430–452. [HBS](https://www.hbs.edu/ris/Publication%20Files/Huang%20et%20al%202017_6945bc5e-3b3e-4c0a-addd-254c9e603c60.pdf)\n- Yeomans, M., Minson, J., Collins, H., Chen, F. & Gino, F. (2020). \"Conversational receptiveness: Improving engagement with opposing views.\" *Organizational Behavior and Human Decision Processes* 160, 131–148. [PDF](https://www.mikeyeomans.info/papers/receptiveness.pdf)\n- \"Curiosity and the dynamics of optimal exploration.\" *Trends in Cognitive Sciences* (2024). [ScienceDirect](https://www.sciencedirect.com/science/article/pii/S1364661324000287)\n- Parisi, A., Thain, N., Hallak, A., Tsai, V. & Qian, C. (2026). \"Real-Time Group Dynamics with LLM Facilitation: Evidence from a Charity Allocation Task.\" Google DeepMind. [arXiv:2605.14097](https://arxiv.org/abs/2605.14097) — cited in §10 for the finding that preference diverged from consensus improvement across N=879 with real stakes.\n- Griffin, V. *Writing Woods* podcast, episode on curiosity (curiosity vs. discipline; exploration vs. exploitation; curiosity vs. efficiency). The prompt for this document.\n"}