{"path":"research/does-descending-change-the-descender.md","content":"# Does descending change the descender?\n\n*2026-08-27. Founder asked for depth on a prediction registered earlier the same day: that\ndecomposition might be the debiasing act and endorsement the entrenching one. Read properly, **the\nprediction as registered is wrong in three separate ways**, and what replaces it is better\nspecified, more interesting, and — this is the uncomfortable part — **not currently checkable in\nthis corpus**, contrary to what was published in three places this morning.*\n\n---\n\n## 0. The three corrections, up front\n\n1. **The literature does not have two states, it has three.** The effect appears, disappears, and\n   in one large study **reverses**. A product thesis cannot rest on it.\n2. **The moderator turns out to be what the claim bottoms out in** — protected values are shielded,\n   consequentialist issues respond — which is *exactly what the terminus enum already encodes*. So\n   the prediction is per-terminus-type, never global.\n3. **Decomposition is not uniformly \"mechanistic\".** Fernbach's moderating condition is *causal\n   mechanism*; his non-moderating condition is *enumerating reasons*. Deliberus decomposition spans\n   eight axes, of which `causal_step` is one — and **decomposing a value claim into premises is\n   structurally the reason-enumeration condition**, i.e. the one that does *not* moderate and may\n   entrench. The naive mapping inverts for a large share of the graph.\n\nAnd a fourth, about us rather than the literature: **the effect requires self-generation.** Reading\nsomeone else's explanation does not produce it. That has a direct and slightly unwelcome\nconsequence for the engagement gradient (§5).\n\n---\n\n## 1. What the original actually found\n\nFernbach, Rogers, Fox & Sloman (2013), *Psychological Science*: people asked to give a **mechanistic\nexplanation** of a policy — how it actually works, step by step — reported lower understanding\nafterwards *and* held more moderate attitudes, and donated less to relevant advocacy groups. People\nasked instead to **enumerate reasons** for their preference showed neither effect. Asking people to\n*justify* a position makes beliefs more extreme.\n\nTwo distinct things are bundled in that result, and the replication history pulls them apart:\n\n- **(a)** the illusion of explanatory depth collapses — you discover you cannot explain it;\n- **(b)** attitude moderates as a consequence.\n\n## 2. The replication history, which is worse than \"mixed\"\n\n| Study | N / design | Outcome |\n|---|---|---|\n| Fernbach et al. 2013 | original, 3 experiments | **works** — (a) and (b) |\n| **Crawford & Ruscio 2021**, *Psych Science* | **three preregistered close replications** | **(a) replicated, (b) did not.** Participants reported less understanding after mechanistic explanation; policy extremity and in-group favouritism **unaffected**. Only one effect survived a small-telescopes analysis |\n| *Cognition* 2022 (Sloman among authors) | timely issues, protected vs consequentialist | **works, conditionally** — meat-eating shifted drastically, abortion barely budged; protected values shield intuitions from counterargument |\n| **Walker, Zhang & Cheng 2024**, *Political Psychology* | **N = 5,139**, post-conflict Hong Kong | **reverses** — positional extremity was *higher* when citizens articulated their understanding |\n\n**The dissociation in Crawford & Ruscio is the single most important line here.** The mechanism\nworks: making someone explain does puncture their sense of understanding. What fails to follow is\nthe attitude change. So *\"decomposition humbles people\"* is on firmer ground than *\"decomposition\nmoderates people\"*, and only the second is a product claim.\n\n**And the reversal is not a null.** In a post-conflict, high-stakes identity context, articulating\nyour understanding made positions *more* extreme in a sample of five thousand. Whatever the\nmechanism is, context can flip its sign — which is precisely the situation Deliberus is designed\nfor.\n\n## 3. The moderator is the terminus type, and we already have that enum\n\nThe 2022 result is the useful one because its moderator is a distinction this project has already\nbuilt. Attitudes grounded in **consequentialist reasoning** respond to mechanistic explanation.\nAttitudes held as **protected values** do not: they are shielded from counterargument, and\nexplaining does not touch them.\n\nThat maps onto the terminus enum almost cell for cell:\n\n| Terminus type | Predicted response to descent |\n|---|---|\n| `empirical` — bottoms out in a checkable fact | **Most likely to moderate.** This is the consequentialist case |\n| `dissolved` | moderation is the *point*; the disagreement was verbal |\n| `fittingness` / `structural` / `axiom_choice` | weakly, or not at all |\n| `is_bedrock` / a declared sacred premise | **Shielded — no effect predicted, and possibly reversal** |\n\nSo the registered prediction should never have been global. **The sharp version: descent moderates\nin proportion to how close the branch bottoms out to a checkable fact, and does nothing where it\nbottoms out in a value posit.** That is testable, and it is a prediction about *our own ontology*\nrather than a borrowed one.\n\nIt also carries a design consequence that runs against a naive reading: **the sacred brake's\nre-scoped form is not merely ethically right, it is empirically consistent.** Analysis may go\nanywhere; but expecting a descent through someone's protected value to *change their mind* is\nexpecting the one thing the evidence says will not happen — and in the wrong context it may harden\nthem. Proportional attunement is the design that fits the evidence, not just the values.\n\n## 4. Decomposition is eight axes, and only one of them is Fernbach's condition\n\nThis is the correction that most changes the picture, and it comes from reading our own\n`decomposition-axes.md` beside the paper rather than from any new source.\n\nFernbach's moderating manipulation is **causal-mechanistic**: *how does this produce that, step by\nstep.* His non-moderating manipulation is **reason-enumeration**: *list why you hold this.*\n\nDeliberus decomposition spans eight axes — predicate–argument, presupposition, degree threshold,\nattribution, scope, frame, **causal step**, other. Only `causal_step` is Fernbach's condition.\nWorse:\n\n> **Decomposing a value claim into its supporting premises is structurally the reason-enumeration\n> condition.** *\"I hold X because A, B and C\"* is a list of reasons, which is the arm that produced\n> no moderation — and the neighbouring *justify your position* framing produced **more** extremity.\n\nSo the naive mapping — *decomposition good, endorsement bad* — **inverts for value claims**, which\nare a large share of what this graph decomposes. If the effect is real, the graph contains both a\nmoderating act and an entrenching act, and **they are currently the same button**.\n\nThe refinement worth carrying: an interface that asks *\"how does this actually work?\"* is doing\nsomething different from one that asks *\"what is this claim resting on?\"*, even though both produce\nDECOMPOSES_INTO edges. That is an argument for the causal axis being **prompted differently**, not\nmerely available.\n\n## 5. The effect needs self-generation — which the engagement gradient should hear\n\nThe IOED literature is consistent that the illusion is punctured by **generating** an explanation,\nnot by reading one; passively consuming explanatory material can even *raise* felt understanding\nand perceived trustworthiness without raising actual understanding.\n\nConsequence, stated plainly because it is slightly unwelcome: **a reader of a beautifully\ndecomposed graph gets none of this.** Whatever epistemic humility descent produces, it is produced\nin the person who descends. This does not touch the library test — *you never owe us a\ncontribution* is a vow about obligation, and reading has its own value — but it does mean\n**a specific benefit that might be claimed for the artifact belongs only to the contributor tier**,\nand claiming it for readers would be unsupported.\n\nIt also sharpens the co-present session design in the same direction the corpus already leans:\nparticipants using it live, from scratch, *are* the generation condition. A read-only tour is not.\n\n## 6. What this corpus can and cannot check — the audit, and a correction\n\n**Published this morning in three places: *\"this graph records descent and endorsement separately\nper user, so it can check.\"* That is half true, and the false half is the half that matters.**\n\nVerified against the code today:\n\n| Requirement | State |\n|---|---|\n| Descent attributable per user | ✅ `POST /claims/{id}/decompose` writes `attributed_to = user email` on every created sub-claim |\n| Endorsement attributable per user | ✅ one vote per user per claim |\n| Timestamps | ✅ `recorded_at` corpus-wide since 2026-08-22 |\n| **Attitude extremity** | ❌ **does not exist.** `VoteRequest.vote` is `'agree'` or `'disagree'` — binary |\n| **Pre/post measurement** | ❌ nothing anywhere |\n\n**A binary vote cannot show moderation.** Someone moving from strongly-agree to weakly-agree is\ninvisible: both are `agree`. The dependent variable in every study above is attitude *extremity on\na scale*, and this corpus has no scale. So the prediction is **not currently checkable here**, and\nsaying it was is the same class of error the corpus keeps catching — a claim about our own\ncapability, published without opening the file.\n\n**What would make it checkable, minimally**: an optional position slider (−3…+3) beside the binary\nvote, and re-asking it after a descent. That is one column, one component, and one prompt — but it\nis a **UX addition to the contribution surface**, so it is a founder decision rather than a\ndrive-by. Note it would also serve the workshop's pre/post attitude reading, which was filed\nseparately today, so the two wants are one instrument.\n\n## 7. The honest verdict\n\n**Do not build on this finding.** An effect that works, nulls under preregistered close\nreplication, and reverses at N=5,139 is not a foundation.\n\n**Do keep three things from it.**\n\n1. **The dissociation.** Descent plausibly produces *epistemic humility* (replicated) rather than\n   *moderation* (not replicated). Humility is a smaller claim and a better one — and it is closer\n   to what this project actually promises.\n2. **The moderator prediction**, which is ours rather than borrowed: descent should moderate in\n   proportion to how close the branch bottoms out to a checkable fact. If the graph ever measures\n   attitudes, that is the shape to look for, and a flat result across terminus types would be\n   evidence *against* the ontology's own cut.\n3. **The entrenchment warning**, unresolved and worth holding: if reason-enumeration entrenches,\n   then a value-claim decomposition may harden the decomposer. **That is the one way this project\n   could make polarization worse while every internal instrument reads green**, since nothing here\n   measures the contributor's own attitude. It belongs in the threat model as a watched hypothesis,\n   not as a finding.\n\n---\n\n## 8. The timescale objection, which is correct — and what happens under *sustained* exposure\n\n*Founder, twice: \"this feels like it's talking about the near-term or immediate psychological\neffect… what about when the same user surfs around Deliberus and is met with many different\nviewpoints and evidence pointing in different directions that diverge from their own?\" The\nobjection is right, and taking it seriously changes the picture more than the single-episode\nliterature does.*\n\n**Everything in §§1–5 measures one sitting.** The standard finding in persuasion research is rapid\ndecay — the half-life of an advertisement's effect is about three days, and regression toward the\nprior attitude is the most frequently observed longitudinal pattern. So a null on a twenty-minute\nlab manipulation says little about a **persistent artifact a person returns to over months**, which\nis the thing Deliberus actually is. Two of the four studies in §2 measured attitudes within a single\nsession.\n\nFour anchors bear on sustained exposure, and they disagree in an informative way.\n\n**The bad one, and it is precisely the founder's scenario.** Bail et al. (PNAS 2018) paid Twitter\nusers to follow a bot exposing them to opposing political views **for a month**. Republicans became\n**substantially more conservative**; Democrats moved slightly liberal, not significantly. Sustained\nexposure to the other side *increased* polarization. **But what the bot exposed them to was elected\nofficials and opinion leaders** — conclusions and identity signals, not reasoning. That distinction\nis Deliberus's entire bet, and it must be stated as a **bet** rather than a rebuttal: nobody has run\nBail's design with reasons in place of flags.\n\n**The good one, and it is the sharpest external statement of this project's own dialectic.** Kalla &\nBroockman, three preregistered field experiments, 230 canvassers, **6,869 voters**: face-to-face\nconversations **deploying arguments alone had no effects**, while otherwise identical conversations\nthat *also* included the **non-judgmental exchange of narratives durably reduced exclusionary\nattitudes for at least four months**. Replicated across immigration and transphobia; the earlier\ntransgender study held at three months.\n\nRead that twice. **The argument arm was null. The argument-plus-attunement arm was durable.** An\noutside field experiment, at scale, found that analysis alone does not move people and analysis\ncarried inside a personal exchange does. `analysis-and-attunement.md` argues this from first\nprinciples; this is the measurement.\n\n**The cost one — and it is weaker than first reported here, corrected 2026-08-27 on founder\nchallenge.** Mutz's *Hearing the Other Side* finds cross-cutting exposure raises tolerance **and\nlowers political participation** — *discourse\ngenerates conflict; conflict generates ambiguity; ambiguity generates disengagement*. So even the\n**successful** version of what Deliberus does has a documented cost, and the mechanism is\nuncomfortably close to something this project treats as a virtue: a first-class `undecided`, a\npermissive zone, an honest gray band. **⚠ But the general effect does not survive meta-analysis.** Matthes et al. (2019) found **no\noverall relationship — neither positive nor negative** — between cross-cutting exposure and\npolitical participation, concluding that the relationship is **not direct** and is contingent on\nmediators and moderators. Mutz's finding has been challenged from both sides, with some studies\nfinding positive effects. So this is **not a standing cost of doing what Deliberus does**; it is a\ncost of doing it a particular way, and the meta-analysis's own conclusion — *contingent on\nmediators* — hands the question back to design.\n\n**And this corpus already names the mediator, and already has the rule.** PACE: a prediction error\nyields curiosity only after an appraisal that the gap is **closable**, and yields\n**anxiety-related inhibition** when it is not. That is Mutz's mechanism at the individual level.\nThe standing design rule that answers it is already ratified — **no gap displayed without a move\noffered** — which means the demobilising branch is a *design failure mode* rather than an inherent\nproperty of mapping disagreement.\n\n**⚠ And the rule reaches only one of two gap-types** (founder, same day): a **visible gap** the graph\ncan display and rank, versus a **frame gap** that is missing *because it lies outside the frame the\ngraph was built inside* — which is never displayed, so the rule does not reach it. There the\ndiscomfort is **correct** and designing it away would silence a working alarm; the answer is an exit\nrather than reassurance. Design direction (articulation help, not a contentless flag):\n[what-human-judgment-is-for.md](what-human-judgment-is-for.md) § Act two, continued.\n\n**The distinction that carries it**: diffuse ambivalence is *\"both sides seem to have a point and I\ndo not know what to think\"*; a mapped disagreement is *\"I know exactly which premise would settle\nthis\"*. The second is not less certainty, it is **located** uncertainty — and locating it is the\nproduct. The instrument for this already ships: **hinge-ordered questioning with `worth_asking`**\ntells a reader which open question would move the conclusion most, which is precisely the\nconversion of ambiguity into a next step.\n\n**The founder's own position, verbatim (2026-08-27), which is what forced the correction above:**\n*\"This is all true when the ambiguity and the complexity becomes overwhelming, but I don't\nnecessarily think that has to be the case with Deliberus. In fact, I think it's one of the platforms\nthat should be able to cut through the noise and increase the volume of the signal if we do\neverything right.\"* The meta-analysis agrees with him on the evidence, and the *if we do everything\nright* is the load-bearing clause: the mediator is design, and the design rule already exists.\n\n**What survives as a real risk, narrowed.** The failure mode is specifically a surface that\ndisplays a gap with no move — and the corpus has a standing violation on exactly that surface: the\ncompleteness oracle's `unsupported_value_premises` reports what is unsupported and offers nothing.\nThat is where this cost would bite if it bites at all, and it is fixable rather than structural.\n\n**And the reassuring one.** Backfire is far weaker than folklore holds: Wood & Porter ran five\nexperiments, **>10,100 subjects across 52 issues**, and found no correction capable of triggering\nbackfire — with Nyhan & Reifler, whose original finding started the idea, later collaborating on a\nstudy that also found none. People do generally update toward evidence. **The exception is the\nsacred case** (§9).\n\n### The synthesis worth keeping\n\n**The difference between Bail's backfire and deep canvassing's durable change is neither dose nor\nduration.** Bail's was a month; a canvassing conversation is ten minutes. What differs is **what the\nexposure carries** — a group's flag, or a person's reasons with a face attached.\n\nThat places Deliberus between two anchors: **the thing that failed is exactly what it removes**\n(identity-marked conclusions), and **the thing that worked is exactly the pole it hosts but does not\nyet feed** (narrative and recognition). The core-dialectic indictment in\n[convictions-across-scales.md](convictions-across-scales.md) — that this project *hosts* both poles\nand couples neither — now has an experimental price attached.\n\n**Nobody has measured what the founder is asking about**, because sustained, self-directed,\nreasoning-level exposure had no medium. Bail's study is the closest and it is conclusion-level. So\nthe honest position is that the question is **open**, and this project is a plausible instrument for\nanswering it — which makes it a research contribution rather than only a product claim.\n\n## 9. What happens when overwhelming evidence meets a sacred value\n\n*The founder's second question, and the literature is unusually precise here.*\n\nThe general answer is that argument is the **wrong instrument**, and adding more of it makes things\nworse. Ginges, Atran & Medin (PNAS 2007) and the follow-on work found that **material incentives to\ncompromise on a sacred value backfire** — moral absolutists showed *increased outrage* when offered\nmore, because treating a sacred value as fungible with secular goods is itself the taboo — while\n**symbolic concessions worked**: Palestinian devoted actors became more willing to consider\nrecognising Israel's right to exist if offered an official apology for 1948; Israeli settlers became\nless disapproving of a land compromise if Palestinian groups symbolically recognised Israel.\n\n**More reasoning is a material incentive in argumentative clothing.** Piling evidence against a\nsacred value is offering a bigger payment for a thing that is not for sale, and the measured\nresponse to a bigger payment is more outrage.\n\nWhat moves is **recognition** — the other side demonstrating that it understands what the value\n*is* and treats it as real. That is not a complement to analysis on the hardest cases. **It is the\nonly channel with a measured effect.**\n\nThree consequences, and the first two the project has already half-arrived at:\n\n1. **The re-scoped sacred ruling is empirically correct, not merely humane.** Analysis may go\n   anywhere; what changes at the boundary is proportional attunement and heightened humility. The\n   evidence says an analysis-only descent through a protected value will produce outrage, and in the\n   wrong context, hardening.\n2. **Stake testimony is the fourth human act, and this is why it is not decoration.** *Here is what\n   this costs from where I stand* is the graph's symbolic-concession channel — the only recorded act\n   whose function is recognition rather than argument.\n3. **A demonstration of understanding is a buildable artifact.** The corpus already has one shape\n   for it: the **position fingerprint** (traverse the divergence rather than the whole tree). An\n   accurate, attributed rendering of *why the other side holds what it holds*, ratified by them, is\n   structurally a symbolic concession — and always-mint plus the consent verdict is most of the\n   machinery.\n\n**The honest limit.** Everything in this section is measured in *interpersonal and inter-group*\nsettings, where a concession comes from an identifiable other party who could have withheld it. A\ngraph is not a party and cannot apologise. Whether recognition works when it arrives as structure\nrather than from a person is **unmeasured**, and it is the load-bearing uncertainty in the\nco-present session design, where a human is present to supply exactly what the artifact cannot.\n\n## Sources\n\n- [Fernbach, Rogers, Fox & Sloman (2013), *Political Extremism Is Supported by an Illusion of Understanding*](https://journals.sagepub.com/doi/abs/10.1177/0956797612464058)\n- [Crawford & Ruscio (2021), *Asking People to Explain Complex Policies Does Not Increase Political Moderation: Three Preregistered Failures to Closely Replicate*](https://pubmed.ncbi.nlm.nih.gov/33667138/)\n- [*Is political extremism supported by an illusion of understanding?* (Cognition, 2022)](https://www.sciencedirect.com/science/article/abs/pii/S0010027722001342) — the protected-values moderator\n- [Walker, Zhang & Cheng (2024), *Puncturing the Knowledge Illusion Does Not Reduce Policy and Political Extremism*, Political Psychology 45:581–599](https://onlinelibrary.wiley.com/doi/abs/10.1111/pops.12938) — the reversal\n- [Bail et al. (2018), *Exposure to opposing views on social media can increase political polarization*, PNAS](https://www.pnas.org/doi/10.1073/pnas.1804840115)\n- [Kalla & Broockman (2020), *Reducing Exclusionary Attitudes Through Interpersonal Conversation: Evidence From Three Field Experiments*](https://isps.yale.edu/research/publications/isps20-01) · [Broockman & Kalla (2016), *Durably reducing transphobia*, Science](https://www.science.org/doi/10.1126/science.aad9713)\n- [Mutz, *Hearing the Other Side*](https://www.goodreads.com/book/show/129774.Hearing_the_Other_Side) — the tolerance/participation trade-off\n- [Matthes et al. (2019), *A Meta-Analysis of the Effects of Cross-Cutting Exposure on Political Participation*](https://www.tandfonline.com/doi/full/10.1080/10584609.2019.1619638) — **no overall relationship in either direction**; contingent on mediators\n- [Wood & Porter, *The Elusive Backfire Effect*](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2819073) — 5 experiments, >10,100 subjects, 52 issues, no backfire\n- [Ginges, Atran & Medin (2007), *Sacred bounds on rational resolution of violent political conflict*, PNAS](https://www.pnas.org/doi/10.1073/pnas.0701768104) · [Atran & Axelrod, *Reframing Sacred Values*](https://public.websites.umich.edu/~axe/research/Reframing_Sacred_Values.pdf)\n- [*How Quickly We Forget: The Duration of Persuasion Effects From Mass Communication*](https://www.tandfonline.com/doi/abs/10.1080/10584609.2013.828143)\n- [Rozenblit & Keil (2002) via the IOED literature](https://en.wikipedia.org/wiki/Illusion_of_explanatory_depth); [generation-versus-reading](https://www.tandfonline.com/doi/full/10.1080/21548455.2021.2018627)\n\n## See also\n\n[cognitive-bias-codex-and-human-contribution.md](cognitive-bias-codex-and-human-contribution.md) §4.4 (where the prediction was first registered) · [decomposition-axes.md](decomposition-axes.md) (the eight axes §4 rests on) · [engagement-gradient-priors.md](engagement-gradient-priors.md) (§5's consequence) · [parity-and-the-zone-interior.md](parity-and-the-zone-interior.md) (the sacred ruling §3 turns out to support) · [islands-of-coherence.md](islands-of-coherence.md) (the co-present session as the generation condition)\n"}