← Kept

Cartograph — Research → feature map

Date: 2026-08-05 · Merges the trust audit (RESEARCH-trust-audit.md, 17 items) with the thesis-research reports (research-cartograph/h1–h7), then maps every valid item to what is actually built (verified in cartograph/ code + docs/HOW-IT-WORKS-TODAY.md, not specs). Items the audit missed are marked NEW.

How to read a row: The result = what the research actually found, in plain language. In the app = where and how that finding is embodied today (✅ built · ◐ partial · ✗ not built). Prospective = the feature that would close the gap.

Table 1 — Valid research → existing or prospective features

Memory & cognition

Research The result In the app Prospective
Recognition over recall (Nielsen #6) People are far better at recognizing something they've seen than at reconstructing it from memory unaided. One of the most replicated findings in psychology. ✅ Returning to a decision shows the frozen page and the decision record card (answer / Because / Let go / Revisit if). You recognize the state at a glance instead of rebuilding it from a chat transcript. —
Chunking (Chase & Simon 1973) Working memory operates on grouped units ("chunks"), not raw items; a well-learned label is a pointer into a rich structure (chess masters recall board positions as patterns, not pieces). ✅ Every page line is a compressed handle standing for a larger discussion. The question line is directly editable, so the most important handle can be tuned until it resonates. Log how often the question gets edited — that rate was meant to be the "does the handle resonate?" metric and was never instrumented
Working memory ~4 chunks (Cowan 2001) When grouping tricks are prevented, people can hold only about four chunks in mind at once — not Miller's seven. ✅ Hard section caps: Settled 6, Open 5, At stake 3 (types.ts SECTION_CAP + prompt rules). A full section forces the model to propose a consolidation instead of adding an item. This cap is what keeps the page glanceable rather than accretive — the most load-bearing research input in the product. —
Event Segmentation Theory (Zacks) Memory cuts continuous experience into events at boundaries. Things inside one event blur together; boundaries are what make moments findable later. A 200-turn chat is one undifferentiated event — that's why nothing in it can be found again. ◐ Boundaries are made felt: lines animate with a spring when they land or settle, and the conclusion card enters with a staggered reveal — conclusion is the one big manufactured boundary. Within a long page, though, sittings run together with no visible seam. Visible episode boundaries inside long pages (a seam per sitting; the old "sealed episodes" idea, never built)
NEW Ovsiankina effect (Ghibellini & Meier 2025 meta-analysis) An interrupted task creates a pull to resume it: across 21 publications, people spontaneously went back and finished 67% of the time. This — not Zeigarnik's memory claim — is the solid half of the "open loops" literature. ◐ Open lines carry closing conditions (what would settle them), and after 3 exchanges with no movement a stall note makes the model offer to close a specific item or set the thread aside. But nothing uses the resumption pull when you come back to an unconcluded page. On re-entry, lead with the open line closest to closing ("one step from settled") — ride the documented pull to finish
NEW Attentional residue (Leroy 2009) An unfinished task keeps occupying part of your attention while you do the next thing; marking it finished releases that capacity. ✅ Closure is engineered as release: when the last open question closes, the app auto-concludes after a ~1s beat, and the concluded footer asks "Does this feel settled?" [Settled]/[Not yet] (CAR-61, live in prod) so the loop is closed in the user's head, not just in the data. —
NEW Goal-gradient effect (Kivetz et al. 2006) People accelerate as they approach a visible finish line (café loyalty cards: purchases speed up near the free coffee). ◐ The open-question count is visible and shrinks toward conclusion, but the app never points at the finish line. Emphasize proximity to closure when near ("1 open question left") — subtle, no gamification

Defaults, automation & attention

Research The result In the app Prospective
Default effects (Jachimowicz et al. 2019 meta-analysis) Whatever happens when you do nothing gets chosen far more often (average d ≈ 0.6–0.7 across ~58 studies). Direction is robust; the size varies a lot by domain. ✅ The core disposition mechanic: model-proposed page edits are accepted by doing nothing — the next send sweeps them in (autoAcceptSweep); rejecting is one tap. This is what makes the per-turn cost of structure zero. Accepts are logged (edit_accepted {mode:'auto'}). —
NEW Defaults decay in repeated choice (Roth, Waldman & Erev 2024) When the same default is offered over and over, people keep it only while it keeps proving good: 88% stick with a good default, 33% with a bad one. A one-shot default rides inertia; a repeated one is silently audited by experience. ◐ The risky ops are carved out of the default: consolidations and decisions need an explicit ✓ or they drop. The quality audit this finding demands exists only as raw logs. Track own acceptance rate over time as a relative quality dial (no imported threshold — see Table 2)
NEW Automation bias (Parasuraman & Manzey 2010) People over-trust automated suggestions and rubber-stamp them; training and expertise don't fix it. ◐ Same ✓ carve-outs keep judgment in the loop for the moves that matter. But a silent accept and a considered accept currently look identical in the data. Record accept provenance (silent-default vs actively touched) so real judgment can be told apart from rubber-stamping — flagged as day-one-critical in the thesis research
NEW Alert fatigue (clinical-alert literature) Interruptive alerts get overridden ~90% of the time regardless of content; the interruptive form is the problem, not the message. ✅ No proposal ever interrupts: edits land inline in the page, there are no modals, no confirmations mid-thought. —
NEW Suggestions cost attention; deliver at breakpoints (Quinn & Zhai 2016; Iqbal & Bailey 2010; Fitz 2019) Every suggestion shown costs attention even when rejected, and delivering at natural pauses (task breakpoints, batched) hurts far less than streaming continuously. ✅ Edits land batched at the between-sends breakpoint — never mid-typing — and the section caps bound how much is ever on screen at once. —

HCI lineage (documented field experience, not RCTs)

Research The result In the app Prospective
Formalization tax (Shipman & Marshall 1999) A decade of structured-thinking tools died the same death: forcing people to express thoughts in the system's formalism (types, fields, links) costs so much — broken flow, premature structure — that users rationally refuse. ✅ The founding constraint. The model performs all structuring; the user never fills a type field, never names a link, never sees structure labeled on screen. The tax's notational half is paid by the machine; only the thinking itself remains. —
Facilitator finding (Kunz & Rittel → Conklin's gIBIS) Issue-mapping works when a trained facilitator does the structuring in real time, and dies when users must classify their own statements (gIBIS users burned effort debating Issue-vs-Position instead of thinking). ✅ The founding move: the model is the facilitator. The two-call turn classifies and files everything; the user keeps exactly the two jobs a facilitated participant keeps — naming (editing the question) and judgment (✓ on decisions). —
Spatial hypertext (Marshall; VKB) People happily express rich structure when it stays implicit — position, grouping, emphasis — and nothing forces them to name it. ✅ Structure stays implicit throughout: sections, replies anchored to lines, the collapsed "Set aside" footer. Nothing ever asks "what kind of thing is this?" —
NEW Four capture subtasks; LLM strength/weakness split (EchoMind, CSCW 2025) Capturing live dialogue means four jobs: splitting, classifying, naming, linking. LLMs are excellent at extraction (94% coverage) but poor at reading where the conversation is heading (28% precision on direction shifts). ◐ The model absorbs the four mechanical jobs; the user keeps direction. But the app doesn't yet focus the user's scarce attention on the model's documented weak spot. Make At-stake lines (the "direction read") the ones most inviting user correction, rather than spending user attention ratifying Settled bookkeeping the model gets right anyway
NEW Grudin's inequality A capture system survives only if the person doing the capture work gets value immediately — nobody records for a hypothetical future reader. ✅ The page pays every single turn: it is literally what the model reasons over (page-context.ts sends the structured page, not the transcript), so better structure → better replies, now. —
NEW Hand-structuring aids comprehension (MeetMap, CSCW 2025) People valued AI-built maps for low effort, but building the map themselves was part of how they understood the material. Removing the work can remove some of the comprehension. ✗ Not addressed — and it cuts against the zero-cost default, so deliberately so. Low priority: an optional "arrange it yourself" gesture (e.g., reorder Settled) if comprehension complaints ever appear

Closure, re-reference & payoff

Research The result In the app Prospective
NEW Saved ≠ re-found (Bergman 2021; NASA lessons-learned; Rhodes JITIR) Deliberately saved things almost never get looked at again (16% of bookmarks; NASA's mandatory lessons database "rarely consulted"). What works is proactive surfacing at the moment of relevance — just-in-time retrieval tripled re-use. ✗ Concluded decisions are recorded but nothing ever surfaces them again. The archive is currently write-only — the exact failure the research documents. Biggest gap. Trigger-based recall: when a new question opens, show related past conclusions unprompted (a recognition surface, zero search required)
NEW Structured debriefs work (Tannenbaum & Cerasoli 2013, d=.67; Keiser & Arthur 2021, d=.79) Structured post-event reviews improve later performance ~20–25% — two large meta-analyses, the strongest experimental evidence in the whole "review" space. ✗ No review mechanism exists. A lightweight revisit ritual when a "Revisit if" condition fires — the only review payoff with solid evidence behind it
NEW Terminal ceremonies die (Darling et al., HBR 2005) Corporate after-action reviews that end with the review — no link to a next action — degrade into empty ceremony and get abandoned. ◐ The decision card records "Revisit if …" but it's inert text; nothing watches for the condition. Wire "Revisit if" to real triggers (time, or keyword match when a related new page opens) so closure feeds forward
NEW Hindsight bias (Fischhoff 1975; Baron & Hershey 1988) Once you know how things turned out, memory rewrites what you expected — you "knew it all along." Only a record made before the outcome protects against this. ◐ "Because" and "Let go" are frozen at conclusion — a partial ex-ante record. What's missing is the prediction itself. One expectation line at conclude ("What do you expect will happen?") — cheap, and it's the defensible payoff claim per the thesis research
NEW Wicked environments (Hogarth et al. 2015) Experience improves judgment only where feedback is quick and clear ("kind" environments). Life decisions are the opposite — slow, noisy feedback — so outcome review alone breeds confidence without competence. ✅ Applied as restraint: neither product nor marketing ever promises "better judgment" (the honest-scoreboard framing). The defensible payoffs are recall and hindsight-bias protection, nothing more. —
NEW Game forms safe, reward loops poison (Deterding 2019; 2023 meta-analysis) Game forms — endings, bounded state, visible progress — carry no documented backfire. Reward loops — points, streaks — backfire by making the activity feel mandatory. ✅ Conclusion is an ending, the page is bounded state; there are no points, streaks, or scores anywhere. —
NEW Status-quo nudge (Levitt 2016/2021) People stuck on a change-or-not decision who were nudged to choose change were measurably happier six months later — we systematically over-stay. ◐ The stall note nudges a stuck page toward close-or-set-aside, which is the same shape at smaller scale. —

LLM-era (young — single studies, directional only)

Research The result In the app Prospective
DeLLMa (ICLR 2025); Human-AI Deliberation (CHI 2025) Early studies: both models and humans reason better over explicitly maintained structured state than over raw free-form transcripts. ✅ Consistent-with (not proof): the model's context each turn is the structured page + exhibits, not the conversation verbatim. —
Graph of Thoughts (Besta et al.) Model-internal reasoning as a graph beats linear chains on some tasks. About model internals, not human interfaces. — background framing only —
Verbatim chunks beat distilled artifacts (arXiv 2601.00821, preprint, unread — LOW trust) Claims raw verbatim excerpts outperform distilled summaries as long-conversation context — if true, direct counter-evidence to page-as-context. ◐ Already hedged by accident of design: exhibits are held verbatim alongside the distilled page (exhibits.ts, degrading to abstracts past 40k tokens). Open action: read it before hardening anything about what the seam feeds the model

Table 2 — Invalidated / do-not-cite research

Nothing in the app breaks when these fall — that was the audit's main finding. This table is about the bibliography, marketing copy, and thresholds, not the build.

Claim What it said Status Action
Zeigarnik effect (1927) Interrupted tasks are remembered better than completed ones. Refuted — 2025 meta-analysis over 59 publications: recall ratio 0.99, i.e. no effect. Stop citing anywhere thesis- or marketing-facing. Closure is justified by Ovsiankina + attentional residue + dogfooded experience — all sturdier
Miller's 7±2 (1956) Working memory holds seven items. Folklore — applies to digit span only; Miller himself was half-joking. Cowan's ~4 is the correction. Cowan is the citation; the page caps were always argued from 4 anyway
70%/85% acceptance threshold Below ~70% suggestion acceptance, an assistant becomes a nuisance. Invented — no such threshold exists anywhere in the literature. Loved systems run 15–30% (Copilot 27%, Smart Compose ~15%). Never state a target; track own acceptance rate as a relative trend only
Corpus-as-moat Accumulated personal decision data becomes a defensible asset. Refuted — users churn from note tools despite big corpora; Cloverpop abandoned the consumer market; labs now auto-track memory (OpenAI Dreaming V3) and make it portable (Anthropic export). Moat = the workflow/ritual labs won't build; keep the corpus aggressively exportable (already the marketing stance)
Ritual-efficacy psychology (Brooks 2016; Tian 2018; Norton grief studies) Performing rituals reduces anxiety and improves performance. Retracted / unreproducible (Gino scandal); the one pre-registered independent test came up null. Never justify the Conclude ceremony with this literature — interaction design + dogfooding carry it
Decision-journal "19% improvement" Keeping a decision journal improves decisions ~19%. No primary source exists — traces only to product marketing blogs. Drop
Forecasting-training gain (Mellers et al. 2014) Brief training improves forecasting accuracy ~10%. Contested — 2025 reanalysis: effects eliminated or reversed. Don't build payoff claims on it
Organ-donor default magnitude (Johnson & Goldstein 2003) Opt-out consent swings donation ~15%→90%. Overstated — 2025 meta-analysis: actual deceased donations +7% (not significant), living donations −29%. Registered consent ≠ behavior. Keep the defaults direction (Jachimowicz meta); drop the 6× story and this citation
Plan-making defuses open loops (Masicampo & Baumeister 2011) Merely making a plan (not finishing) releases intrusive thoughts. Unverified — no replication either way, from a lab whose flagship program collapsed. Treat as hypothesis, never cite as evidence

Not evidence, by design (fine as vocabulary, zero evidentiary weight): Nelson's transclusion, Hayek 1945, DELITE (position paper, evaluation deferred), Ironsworn's oracle mechanic and roguelike morgue files (existence proofs that the interaction pattern can be fun, nothing more).

The short version