{"path":"research/site-declutter-directions.md","content":"# Site Declutter: Findings and Directions\n\n**Date**: 2026-08-19 · **Status**: Research + proposed directions, nothing decided. **Prompted by the founder**: *\"A curious visitor will likely not read everything on the site. Maybe the site is too cluttered with docs. It's too academic for most visitors. We need to rethink it.\"*\n\n## §0 The problem, stated precisely\n\ndeliberus.com currently serves roughly 159 public markdown documents (23 top-level plus 136 in `docs/research/`), and the landing page's primary affordance is an extraction input form that is paused for everyone but the owner. So the site's one call to action is disabled, and its deepest asset (the research corpus) is presented raw, indexed only by a README table written for AI sessions and collaborators. The audience is mixed: curious Facebook visitors who need the category explained from zero, academic friends who want the depth, and prospective workshop participants who need to see what a live session would feel like.\n\nTwo house constraints bound every direction below. First, the docs are not clutter to be removed: the corpus, the agent-readable surface, and the confession instruments ARE product, and the born-private/lossless rules forbid silent removal anyway. Second, the beginner-readable first layer is a hard directive (CLAUDE.md § UX/Product Directives), and the engagement-gradient priors say a reader who never contributes is a success, so the site's job for 90%+ of visitors is to make reading rewarding, not to convert anyone into a contributor.\n\nThe reframe this research supports: the problem is not that there are too many docs. It is that the site has **no first layer**. Everything is presented at the same depth, so the visitor must do the triage the site should have done.\n\n## §1 What the research says\n\n### 1.1 The visitor's clock (the numbers that discipline everything else)\n\n- Page-leaving behavior follows a Weibull distribution with negative aging: the first 10 seconds are a ruthless triage, the next 20 are still high-risk, and only after ~30 seconds does the leaving rate flatten. A page that communicates its value proposition within 10 seconds can earn 2+ minutes; one that does not gets nothing ([NN/g, \"How Long Do Users Stay on Web Pages?\"](https://www.nngroup.com/articles/how-long-do-users-stay-on-web-pages/), on Liu et al.'s 2-billion-dwell-time dataset).\n- 57% of viewing time is spent above the fold and 81% within the first three screenfuls ([NN/g, \"Scrolling and Attention,\" 2018](https://www.nngroup.com/articles/scrolling-and-attention/)).\n- Users read at most ~28% of the words on an average page, realistically ~20% ([NN/g, \"How Little Do Users Read?\"](https://www.nngroup.com/articles/how-little-do-users-read/)). Attention by paragraph position: 81% of users look at paragraph 1, 71% at paragraph 2, 63% at paragraph 3, and only 32% at paragraph 4 ([NN/g, \"Website Reading\"](https://www.nngroup.com/articles/website-reading/)).\n- Scanning is not laziness; it is efficient triage of an environment with more information than can be processed. Design that fights scanning loses ([NN/g F-pattern](https://www.nngroup.com/articles/f-shaped-pattern-reading-web-content-discovered/); [Atticus Li on progressive disclosure, 2026](https://atticusli.com/blog/posts/progressive-disclosure-less-information-increases-conversion/)).\n\nConsequence for Deliberus: the current landing asks a cold visitor to parse an input form whose main button is disabled, seven abstract concept keywords, and a four-block explainer, inside a 10-second window. The corpus never gets a chance because the first screen does not buy the next 30 seconds.\n\n### 1.2 Progressive disclosure is a content-strategy problem, not a widget problem\n\nThe 2025–26 consensus is stable and specific: reveal information in layers ordered by probability of relevance; three to four layers maximum; never hide what the visitor needs for the *current* decision; and treat the layering as an editorial question (\"what does the user need to know right now to make their next decision?\") rather than a matter of accordions and tabs ([Atticus Li](https://atticusli.com/blog/posts/progressive-disclosure-less-information-increases-conversion/); [Austin Lindsay](https://austinlindsay.design/insights/progressive-disclosure-web-design/)). A useful concrete ladder for complex products: the 10-second answer (hero), the 30-second answer (how it works), the 2-minute answer (capabilities), and the 15-minute answer (technical depth), with the observation that most complex products dump layers 3 and 4 into layer 1 ([Hyperhelios, \"How to Communicate Complex SaaS Products\"](https://www.hyperhelios.co/blog/how-to-communicate-complex-saas-products-on-your-website)). That last sentence is a fair description of deliberus.com today: the 15-minute layer (research docs) is one click from the hero and visually co-equal with everything else.\n\nOne sharp warning from Stripe's year-long homepage redesign: the team tested an accordion layer and found **people don't click** — visitors stare at a collapsed line and leave, and a collapsed title cannot carry a concept that needs explaining ([AINEXT interview with Stripe's design lead, 2026](https://ain3xt.com/en/posts/20260503-stripe-homepage-redesign-craft/)). So the first layer must *say* the essential thing in full view; disclosure widgets are for depth, not for the pitch.\n\n### 1.3 Story beats catalog, and the sections must depend on each other\n\nStripe's stated reason for rebuilding: every new product line had bolted another block onto the homepage until \"the story got buried in the details,\" and by the third screen a visitor no longer knew what kind of company it was. Patrick Collison's framing for the fix: \"A website is your manifesto\" — what you choose to present and not present tells the visitor who you are ([AINEXT](https://ain3xt.com/en/posts/20260503-stripe-homepage-redesign-craft/)). Deliberus's landing has accreted the same way (input form, think panel, preflight, keywords, explainer, extraction grid), session by session.\n\nThe structural test for a narrative page: each section must answer the question the previous section raised; if you can rearrange the sections without anyone noticing, you have a brochure, not an argument ([Sextant Labs, \"The Story Arc,\" 2026](https://sextantlabs.io/blog/story-arc-homepage-narrative); [roast.page, \"Your Landing Page Doesn't Need More Sections. It Needs a Story.\"](https://roast.page/blog/landing-page-story-structure)). The psychology under it is narrative transportation (Green & Brock 2000): a reader inside a story counterargues less and remembers more, whereas a feature list invites item-by-item skeptical evaluation — which, for a site whose visitors include professional philosophers, is not a bug to exploit but a reason to lead with a *demonstration* they can evaluate rather than claims they must trust ([Atticus Li on narrative transportation](https://atticusli.com/blog/posts/narrative-transportation-theory-story-driven-landing-pages/)).\n\nCTA discipline from the same literature: one primary action, appearing at most twice (optionally in the hero, definitively after the case is made); every additional competing action taxes the visitor ([Vantage Design, \"Landing Page Content Flow\"](https://vantage-design.com/resources/landing-page-content-flow/)). Deliberus's current primary action is a disabled extract button, which is the worst available configuration: a dead ask in the most expensive slot.\n\n### 1.4 The Linear lesson cuts against imitating Linear\n\nLinear can put a documentary about craft in its hero because its buyers already know what an issue tracker is; identity-signaling heroes work only on warm traffic that does not need the category explained, and demonstrably lose cold visitors who are still asking \"wait, what is this?\" ([The Teardown on Linear's docuseries hero, 2026](https://medium.com/the-teardown/cro-teardown-linear-put-a-documentary-where-the-product-screenshot-usually-goes-b30b74f89ff6)). Deliberus's Facebook visitor is maximally cold: there is no existing category (\"the missing layer\" is the whole point), so the hero must explain and show, not vibe. The transferable part of Linear is different: the Linear Method, a manifesto people share who have never opened the app, doing top-of-funnel work as a *single named artifact*. Deliberus's equivalent already exists in pieces (vision.md, the seven keyword docs, the Swedish introduction); what is missing is the one front door.\n\n### 1.5 Academic project sites fail in known, avoidable ways\n\nThe failure catalog for research-project websites reads like a checklist of risks for deliberus.com: homepage written like a grant proposal, insider terminology assumed, outputs scattered without context, and no plain-language answer to \"what is this and why does it matter\" within the first minute ([ResearchProspect, 2026](https://www.researchprospect.com/why-research-projects-need-better-websites/); [WIT Berry on Horizon Europe project sites](https://witberry.eu/2025/03/01/the-real-potential-of-your-horizon-europe-project-website/)). Two prescriptions recur across this literature. First: *less content means better content* — a good academic site carries only what is unique and necessary, because attention divides across everything shown ([Joshua Kim, Inside Higher Ed](https://www.insidehighered.com/blogs/technology-and-learning/5-arguable-assertions-about-academic-websites)). Second: every listed output needs a one-to-two-sentence plain-language annotation saying what it found and why it matters, because \"not every visitor knows why a working paper matters. The site should help\" ([research lab content guide](https://www.sitesgo.com/blog/research-lab-website-content-guide)). The README's research-doc table has titles and purposes but is written for collaborators and AI sessions, not for a curious reader deciding what to open.\n\n### 1.6 Docs as credibility: the asset is real, and it needs framing, not hiding\n\nThe \"docs are product\" position has strong precedent: Stripe and Twilio treat documentation as a product in its own right, and docs measurably drive adoption and trust ([Lucy Mitchell, \"Understanding Documentation as Product\"](https://lucymarmitchell.medium.com/understanding-documentation-as-product-cb1f730a0366)); Supabase's growth is routinely attributed partly to docs-as-marketing ([SaaS teardown table, 2026](https://socialanimal.dev/blog/saas-website-examples-2026-design-pattern-teardowns/)); and the build-in-public literature argues the mess itself is the trust signal, because \"the curated version teaches nobody anything\" ([Adrian Wedd, \"Why I Build in Public,\" 2026](https://adrianwedd.com/blog/why-i-build-in-public/)). So the corpus should stay public and prominent as an *archive with a frame around it*, not be trimmed to look like a startup site.\n\nThe best-developed pattern for exactly Deliberus's shape of corpus (one thinker, hundreds of interlinked research documents of wildly varying maturity) is the digital-garden lineage. Gwern.net gives every page a status tag (draft / in progress / finished), a confidence tag, and a 0–10 importance rating, and builds \"iceberg\" pages where a reader can stop at the title, the abstract, the section headers, or drill arbitrarily deep — what Gwern calls semantic zoom ([gwern.net/design](https://gwern.net/design)). Andy Matuschak's working notes open with an explicit disclaimer: \"They're mostly written for myself... If a note seems confusing or under-explained, it's probably because I didn't write it for you!\" ([notes.andymatuschak.org](https://notes.andymatuschak.org/About_these_notes)) — which converts a potential liability (unpolished density) into an honest, even charming, framing. Epistemic-status labels on research writing are by now a recognized garden convention ([history of the pattern](https://github.com/benchristel/benchristel.github.io/wiki/EpistemicStatus)). This maps directly onto Deliberus's own confession ethos: the site that says \"systems must be able to confess\" can say, on its own archive, \"these are working documents at varying maturity; here is each one's status.\"\n\nFor organizing the corpus itself, Diátaxis's core insight applies even though Deliberus's docs are not software documentation: organize by *reader need*, not by accretion order; landing pages should read as overviews that introduce, never as bare lists; and lists longer than about seven items need grouping ([diataxis.fr](https://diataxis.fr/); [complex hierarchies page](https://www.diataxis.fr/complex-hierarchies/)). Deliberus's implicit reader needs are roughly: \"explain this to me\" (the concept keyword family), \"show me it working\" (dogfood runs, extraction pages), \"where does the thinking stand\" (frontier, open questions), and \"the full record\" (everything else).\n\n### 1.7 Layout fashion: the bento verdict\n\nBento grids earn their place only when the content is 6–9 genuinely parallel facts with real weight differences, placed *below* a message that has already landed; equal-weight tile walls above the fold are a menu with no starting point and convert worse ([BuildMVPFast, \"When Bento Grids Help vs Hurt,\" 2026](https://www.buildmvpfast.com/blog/bento-grid-helps-vs-hurts-conversion-2026); [Landdding category breakdown](https://landdding.com/blog/bento-grid-design-by-website-category-where-the-pattern-wins)). The honest test: \"if you can rewrite the section as a single paragraph with a strong topic sentence, bento is the wrong format\" ([digitalheroes bento guide](https://digitalheroes.co.in/styles/bento-grid/)). Deliberus's seven concept keywords are close to genuinely parallel facts and could survive this test *below* the story; nothing else on the site should be tiled. Fashion adopted without this filter is exactly the \"used without thought, without purpose\" failure the trend's own defenders name.\n\nThe current strongest show-don't-tell idiom is the live artifact in the hero: Dub.co embeds a working link shortener, Neon a database playground, Axiom a live-looking dashboard ([2026 teardowns](https://socialanimal.dev/blog/saas-website-examples-2026-design-pattern-teardowns/)). Deliberus's equivalent is not the paused input form; it is the graph itself, which is live, real, and browsable right now.\n\n## §2 Directions\n\nFive directions, ordered roughly by leverage. They compose; they are not alternatives, though each stands alone. Effort is scaled small / medium / large relative to recent single-session builds (the synthesis-ledger build was a medium).\n\n### Direction 1 — One story, one ask: rebuild the landing as a narrative with a working CTA\n\n**What changes.** The landing page stops being a stack of accreted blocks and becomes a short story-arc page whose sections depend on each other, beginner-readable at every step: (1) the felt problem, in a bright-12-year-old register (\"Two people disagree. Somewhere under the argument is the real reason why. Finding it is expensive.\") — this is `Opacity Is a Cost, Not a Mystery` translated out of its own vocabulary; (2) what Deliberus does about it, in one sentence plus one image of a real descent; (3) proof: a glimpse of the live graph with real numbers (25 sources, 4,700+ claims, the residue map's honest fraction); (4) the single primary action. While extraction is paused, that action cannot be \"analyze something\" — the input form leaves the hero and the ask becomes **\"Follow one disagreement down\"** (Direction 2's showcase). The seven concept keywords survive as the layer *below* the story (this is the one place the parallel-tile idiom is legitimate per §1.7); the four-block explainer is absorbed into the narrative sections rather than duplicated after them; the extraction grid moves to a browse page. The input form returns to prominence only when extraction unpauses.\n\n**What it preserves.** All current content survives (absorbed or relocated, never deleted, per lossless demotion). The medium still embodies the dialectic: the story opens warm (the felt problem), grounds analytical (the mechanism, the numbers), and closes with an invitation, which is the same arc the current 4-block text attempts, now with structural dependency between sections instead of adjacency. The owner's live form remains behind `is_owner`.\n\n**Effort.** Medium. Copy is the hard part (and the founder's voice review is load-bearing); layout is a reorganization of existing components.\n\n**Tradeoff.** A narrative page is opinionated: returning visitors and the academic friends will scroll past a story they already know to reach navigation, so the nav header must keep direct doors (Browse, Docs, About). And per §1.1, the first sentence now carries everything; if it lands wrong for the philosopher audience it lands wrong prominently. Mitigation: the story's claims should be demonstrations (\"here is the crux neither author wrote down\") rather than assertions, which is also the register the confession principle demands.\n\n### Direction 2 — Show the graph, don't describe it: a curated showcase descent\n\n**What changes.** Build one guided walkthrough of a real disagreement already in the graph — the minimum-wage pair or the death-penalty pair are the strongest candidates, being legible, non-meta topics per the Pipeline Testing Principle applied to audiences — presented as a short scrollable path: the two sources in opposition, one claim from each, the attack edge between them, the unwritten crux (run 6's finding is genuinely arresting for a lay reader: \"the word 'occupation' appears 54 times in one text and zero in the other\"), one critical question, and the badge with a plain-language gloss of what it does and does not certify. Every claim shown is a link into the live graph (per the every-claim-is-a-link directive), so the walkthrough is a ramp, not a cul-de-sac: the reader exits into the real product at whatever depth they choose. This is the live-artifact hero idiom (§1.7) executed with the material Deliberus actually has, and it works during the extraction pause because it is read-only.\n\n**What it preserves.** Nothing is mocked or simplified in the data layer; the showcase reads the production graph through existing endpoints. The engagement gradient is honored: the reader tier gets a complete, rewarding experience with zero contribution asked, and each further rung (tap a badge, open a question) is one step away, satisfying \"no gap displayed without a move offered.\"\n\n**Effort.** Medium. One new route plus curation of which nodes to feature; the hardest part is the plain-language glosses, which must survive the stance-overstatement lesson of Session 19 (say exactly what was measured, no more).\n\n**Tradeoff.** The whole project gets judged by one example, so the example must be chosen for robustness, and the 115-of-116-claims-at-0.500 badge reality (the strength layer's honest NO DATA / neutral state) will be visible in it. That is a feature under the confession principle but requires the gloss to explain why a neutral badge is honesty rather than emptiness — done badly, the showcase demonstrates an unfinished product; done well, it demonstrates the only site on the internet that refuses to fake a number.\n\n### Direction 3 — Split the surfaces: a small front, and the archive as a framed research garden\n\n**What changes.** The docs surface splits into two explicitly named layers. The **front** is small: the seven concept keywords, the introduction (English sibling of `/introduktion`), the frontier page, and about — call it under ten destinations, each already written pedagogically. The **archive** is everything else, reached through one new page (`/research` or a rebuilt docs index) that replaces the raw file listing with a curated library: docs grouped by the questions they answer (\"Can values converge?\", \"How does extraction work?\", \"What broke and what we learned\", \"Funding and positioning\") rather than by filename, each with a one-to-two-sentence plain-language annotation per §1.5, and each carrying a small status header rendered from frontmatter: date, status (working note / stable / superseded), and one line of \"why it matters.\" A short Matuschak-style preamble frames the whole archive honestly: these are the project's working documents, written to think with, published because a system that demands legibility of arguments owes legibility of itself; they vary in maturity and say so. That preamble converts the corpus from apparent clutter into the credibility instrument it actually is, and the status metadata is the confession principle applied to the docs surface — which currently has no instrument of its own (the dogfood-run-4 lesson generalized).\n\n**What it preserves.** Every document stays served at its existing URL (the agent surface, `llms.txt`, and inbound links unbroken). Nothing is trimmed or rewritten wholesale; the change is one new index page, a frame, and a metadata pass. The README's table remains the collaborator/AI index; the new page is the human visitor's index, which resolves the current dual-audience overload of that table.\n\n**Effort.** Medium-large. The index page and preamble are small; the honest cost is annotating ~136 research docs (batchable, and draft annotations can be generated then founder-reviewed) and keeping status fields true, which needs a line in the docs-aren't-lying maintenance pass so the metadata does not itself become a lie.\n\n**Tradeoff.** Curation is editorial power: grouping docs by question quietly ranks them, and a stale annotation is worse than none. The grouping also partially duplicates the README table's role, so the two indexes need a single source (generate one from the other, per source-from-the-SSoT) or they will drift.\n\n### Direction 4 — Start-here paths sized to the engagement gradient\n\n**What changes.** One \"Start here\" page (linked from the story's end and the nav) offering three explicitly sized doors, matching the gradient's tiers rather than an imagined uniform reader: **two minutes** (the introduction, or the showcase descent), **fifteen minutes** (a hand-picked path of three or four docs: introduction → one keyword doc of the reader's choice → one dogfood-run story → the frontier), and **the rabbit hole** (the framed archive of Direction 3, plus the agent surface for the technically curious). This is the 10-sec/30-sec/2-min/15-min disclosure ladder (§1.2) made navigational, and it is the site-level version of \"every tier's action reachable one rung from the tier below.\" The academic friends get handed the fifteen-minute door directly; Facebook visitors self-select into two minutes without shame.\n\n**What it preserves.** No content moves; this is pure wayfinding on top of Directions 1–3. It also gives outreach a stable link target: one URL that works in a Facebook comment, a funder email, and a workshop invitation alike.\n\n**Effort.** Small. One page, mostly copy, once Directions 1–3 exist to point at (a degraded version can ship immediately by pointing at current pages).\n\n**Tradeoff.** A start-here page is one more surface that can rot; its links must be checked in the coherence-audit habit. And explicit time labels set expectations the linked content must meet — the \"two minutes\" door must actually be readable in two minutes, which will pressure-test the introduction's length.\n\n### Direction 5 — Put the honesty instruments on stage: \"How this site keeps itself honest\"\n\n**What changes.** One visitor-facing page that presents the confession layer as the product feature it is: the residue map with its current fraction, the disagreement-preservation instrument, the stance-conflict detector, the completeness oracle, and the synthesis ledger — each in two or three beginner-readable sentences (\"Deliberus checks whether its own summaries dropped a side of a conflict, and publishes the answer\") with a live number where one exists and a link to the instrument's endpoint or doc for depth. Framed by the register the missing-layer doc already prescribes (\"keeps itself honest,\" never \"measures its own distortions\"). A one-tile pointer to this page sits in the landing story's proof section. This is transparency-as-feature executed deliberately (§1.6): where every AI product claims trustworthiness, a page of running self-checks with live, sometimes unflattering numbers is close to unfakeable differentiation, and it is the strongest single answer to the \"too academic\" worry because it *demonstrates* rigor in one screen instead of asking the visitor to infer it from 136 papers.\n\n**What it preserves.** The instruments themselves are untouched; this is presentation only. It completes work already started (the llms.txt Instruments section did this for agents in August; humans currently have no equivalent), and it gives the archive's densest material a legible ambassador.\n\n**Effort.** Small-medium. The instruments and endpoints exist; the work is copy plus a page that calls a handful of read-only endpoints.\n\n**Tradeoff.** Live numbers can read badly out of context (a residue-map fraction of 0.5 over 2 classified termini is honest but tiny), so each instrument needs its scope stated plainly, and the page must degrade gracefully when an endpoint is down — a page about honesty that silently shows stale numbers would be the exact failure it advertises against. The numbers shown need the same overstatement discipline Session 19 enforced on the docs.\n\n## §3 Sequencing note (not a plan)\n\nIf the directions are adopted, dependency order matters more than priority order: Direction 2 (the showcase) is what Direction 1's story points at, so they land together or the story ends in a dead ask again; Directions 3 and 4 are independent of both and of each other; Direction 5 is independent of everything and smallest. The one shared prerequisite is copy in the founder's register — every direction above is mostly an editorial act wearing a small amount of code, which matches Stripe's conclusion that the redesign question was never visual (\"the problem wasn't the visuals, it was the story\").\n\n## Cross-references\n\n`docs/ux-principles.md` (P6 progressive disclosure, P16 value-before-identity, P20 structuring gradient) · `docs/research/engagement-gradient-priors.md` (the tier table these directions are sized against) · `docs/the-missing-layer.md` (the register rule Direction 5 inherits) · `docs/research/dogfood-run-6-israel-palestine-cross-domain.md` and `dogfood-run-2-orthogonal-experiments.md` (showcase material) · CLAUDE.md § UX/Product Directives (beginner-readable first layer) · § Docs-visibility (nothing here moves content across the public/private boundary; all five directions operate on already-public material).\n\n---\n\n## Founder revision to the showcase direction (2026-08-19, same day — supersedes the real-debate hero)\n\nThe founder on the curated real-debate showcase: *\"I might not agree that a real debate showcase is the most valuable thing to show new visitors, I'd personally rather stay on the general level and instead show synthetic quirky examples and how they relate/belong in the app/graph, how they can be untangled by Deliberus, what's unique about Deliberus and why we've been missing it, what Deliberus could conceivably unlock (based on vision docs + strong fractal priors) but presented in a language that beginners can grok.\"*\n\nThis is not just a taste call — it is **Kahan-compliant in a way the real-debate hero is not**. The corpus's own curiosity research holds that comprehension capacity *worsens* polarization on identity-charged questions: lead a cold visitor with Israel/Palestine (or capital punishment, or minimum wage) and identity-protective cognition may fire *before* the mechanism lands — the visitor picks a side of the example instead of seeing the machinery. A **synthetic quirky example is identity-safe**: nobody has a tribe about whether the chopsticks fit the silverware drawer, so the full attention budget goes to the untangling itself. The register already exists and is founder-endorsed (the penguin filed as \"large flightless duck\", the potato cellar, the dieter's scale — the gloss style from the taxonomy and adversary explanations). And synthetic examples are a *sanctioned instrument* here, not a compromise: Ontology Stress-Testing by Synthesis is a named method in CLAUDE.md.\n\n**Revised shape of the showcase**: the hero teaches with synthetic, quirky, beginner-grokable examples — a small argument visibly untangling into claims, the unsaid premise surfacing, the \"why we've been missing this\" beat, and the unlock framed from the vision docs and the cross-scale priors (shared-code, assembly-ceiling) in plain words. The **real-debate proof moves one click deeper**: \"does this work on real heat?\" → the run-6 occupation-crux walk, for the visitor who wants evidence it survives contact with reality. Cold traffic gets safety; skeptics get proof; neither is asked to do both at once.\n\n**Spec work begins next** (the founder's Socratic /spec flow): story-first landing + synthetic showcase as one unit, then the honesty page. Supporting design assets to load when implementing: `docs/design-guide.md`, `docs/ux-principles.md`, the design directives in CLAUDE.md.\n\n**Spec decision 1 (founder, 2026-08-19): the hero leads with VISION** — \"they caught the vision\" wins over felt-problem and mechanism-first. Consistent with the corpus's standing warning that the civilizational scope is the project's real register and trimming it is the amputation to refuse. Design consequence carried into the spec: vision-first survives the hype pattern-match ONLY if the grand claim is concrete-grand (a specific missing thing, not an abstract better-world) and the very next beat grounds it in the quirky synthetic untangling — grandness lands as \"and it's real,\" never as manifesto.\n\n**Spec principle (founder, same exchange): \"Everything else follows from the vision and the convictions it's based on.\"** The landing page's architecture mirrors the conviction stack itself: vision at the top as the generative root, and every section below visibly FOLLOWING FROM it (the quirky untangling as the vision made concrete, the honesty instruments as the vision applied to itself, the docs garden as the vision's paper trail). Not three parallel beats — one derivation, descending.\n\n**Spec decision 2 (founder): the hero register is CHECKING-VS-PERSUASION** — the newest ratified vision language (\"machines out-persuade us; Deliberus is checking built to keep pace\"), urgency and why-now built into the doorway. The flagged risk is on the record: it is the darker doorway (threat before hope), so the spec requires the pivot to invitation inside the same screen — the threat sentence never stands alone; the \"built to keep pace\" answer and the warm untangling beat follow before the first scroll. \"Weaving our minds together\" stays as the brand tagline; the two coexist (tagline = identity, hero = argument).\n\n**Spec decision 3 (founder): the untangling example is THE THERMOSTAT WAR** — the hidden-premise teacher: a years-looping domestic fight whose crux (\"whose comfort counts as the default?\") was never on the table, which is the corpus's own sharpest extraction finding (the crux of a real debate is typically implicit) in identity-safe miniature. Supporting roles retained: pineapple-on-pizza carries the residue-map beat (some disagreements honorably end in taste — Deliberus tells you what KIND you're in) later in the page or in the two-minute door; the dog decision carries the claim-kinds beat (value vs prediction vs promise-as-evidence) where claim types get taught.\n\n**Spec decision 4 (founder): the CTA is THREE DOORS** — the wayfinding chooser over any single ask: honest about the engagement gradient (the gradient doc's own logic: one CTA flatters one tier and lies to the rest). The walk-the-map idea survives INSIDE it: door one (two minutes) IS \"walk one map\" — the thermostat extraction seeded into the live graph; door two (fifteen minutes) the introduction; door three (the rabbit hole) the research garden. Outreach gets one stable link; each tier gets its own next step.\n\n**Spec decision 5 (founder): the thermostat war is SIGNED OFF** (\"that's actually funny. I think it's great\") with the three hardenings: roommates-not-couple (dodges the gender-trope), the unwritten premise rendered as a statable claim, and the one-line scale-up answer to \"you don't need software for this.\" Examples two and three (pineapple, dog) under review next — the founder's bulletproofing pass, one at a time.\n\n## The \"vibes\" ruling + the questions-as-entities finding (2026-08-19, Socratic round continued)\n\n**Spec decision 6 — \"vibes\" as a product-voice word, decomposed into five aspects and ruled on:**\n- **(a) The double-audience intent — YAY**: \"vibe\" is simultaneously self-irony to a neurotypical reader and the LITERAL correct name to a neurodivergent one for affective ground that is statable but not further decomposable. Founder framing, verbatim: *\"It gives a kind of self-ironic vibe (lol) that is also earnest if you ask me as a neurodivergent/aspie.\"* The label is more honest than a formal term, not less.\n- **(b) Display-layer alias ONLY — YAY**: \"vibe\" renders precise underlying types (taste-floor, fittingness residue, affective value premise) on the surface; the ontology keeps its exact vocabulary. Never an enum value.\n- **(c) The consent line — REJECTED by the founder in favor of teasing cheek**: *\"I'm open to being a bit teasingly cheeky. But ofc this could potentially repel some users. I however believe those users probably wouldn't even go sit down in front of Deliberus in the first place.\"* Analysis verdict: right-ish, with the \"-ish\" located at the TABLE not the DOOR — the repelled-user is a moment, not a personality type; the same humor-tolerant person is maximally exposed at the instant their just-surfaced load-bearing premise is displayed to a disagreeing counterpart, and there the tease reads as the system taking sides (PACE: gap-plus-mockery = threat, not curiosity; the conviction's guard: comprehension must contain attunement). **Proposed timing carve (awaiting founder's word)**: cheeky by default everywhere — landing, UI, self-labels, cooled contexts, and especially the machine's OWN uncertain classifications (self-irony proper) — with warm-neutral register in exactly one situation: the live moment of a person's freshly surfaced premise before a disagreeing counterpart. Consistency note: a STABLE cheeky register is itself the neurodivergent-safety mitigation (unpredictable teasing is what reads unsafe; \"this system always talks like this\" does not).\n- **(d) Sequencing — YAY**: landing copy first (the pineapple rung: \"ends in a vibe — and that's a legitimate floor\"); in-product self-label affordance later as its own decision.\n- **(e) The self-ironic streak in product voice — YAY**: the first deliberate humor in the register; rigor that can smile at itself is the counter to \"too academic.\"\n\n**Questions as graph entities — prior-art finding**: already BUILT narrowly — 712 live Question nodes, all machine-minted critical-question instances attached to relationship edges with both polarity claims pre-generated (Session 7 design). Never built and never discarded: PERSON-POSED questions as general graph citizens. The extension is natural (same both-polarities pattern), and the thermostat's unwritten premise could be rendered either as a minted claim or as a Question node with premise/negation polarities using only existing machinery. Filed as an ontology-extension idea in TODO.\n\n## SHIPPED 2026-08-19 (unattended build, founder walking) — draft copy LIVE pending his voice pass\n\nThe derivation landing is deployed: checking-vs-persuasion hero, three-rung conviction ladder (thermostat stepped walk with the crux panel glowing; pineapple with the \"vibe\" word in its earnest sense; dog claim-kinds), the confession line, the two beats (missing-layer with the twenty-platforms line relocated from the old about-section; the unlock beat with the assembly-ceiling idea in plain words), three doors, and the input form retitled \"Analyze your own thinking\" below the doors with the paused note. **The thermostat map is seeded in the live graph** (`scripts/seed_thermostat_example.py`, re-runnable; 11 claims + 10 edges, every attribution synthetic-labeled; entry claim `claim_thermo_t9` verified serving with edges + the agent-readable twin). **Editorial calls made under the autonomy grant, flagged for founder review**: (1) ALL COPY IS CLAUDE'S DRAFT — live without the founder-voice pass, on his explicit \"go with your editorial choices, I can always review and edit later\"; (2) door two = vision.md (the open pick); (3) the previous-extractions grid KEPT at page bottom rather than relocated (lower risk, real proof-of-life — a deviation from the spec's \"move it\", reversible in one edit); (4) badges on the synthetic claims read honest gray NO DATA (nothing examined — correct, and consistent with the filed badge-boundary founder decision).\n\n**Spec decision 7 (founder ratified Claude's read of his instinct, 2026-08-19): the dog swaps out for GRANDMA'S CAR KEYS; pizza stays.** The founder's severity worry (\"the dog one I'm not sure if it's too silly. Same for pizza but less so\") resolved via the severity-ascent reading: the ladder now deliberately ASCENDS in seriousness — thermostat (light domestic) → pineapple (silly, carrying the comedy-into-honesty vibe beat, which needs the lightness to land) → grandma's keys (tender-serious: a fact, a value, and a promise tangled in one sentence, \"everyone is right and everyone is losing\"). Grandma's keys teaches claim-kinds BETTER than the dog (a checked fact + a non-vanishing value + a commitment-as-evidence, ending on \"the actual trade-off instead of who loves her most\") at real stakes with zero party valence. Pineapple also gained its false-floor hardening from the review: legitimate once it really IS the floor (\"never actually tried it\" is a checkable claim wearing a vibe's clothing) — the CQ mechanism taught in one parenthesis. Middle-band candidates parked for future surfaces: the housing-association camera vote (crime register in within-community form), screen time.\n"}