| App | Score | Shelf | Why | Passes |
|---|
| Mandrel | 70/81 | Trades & Field | Real uncertainty-propagation math (proper normal CDF, quadrature variance), a 53-material/42-core/17-brand catalog, Bayesian tare self-calibration, bulk CSV/Spoolman intake, and a batch planner all work end to end with 174 tests passing and zero console errors on both probes. | 0 |
| Recto | 70/81 | Education | Correct SM-2 implementation, 56 real passages across 4 tiers, genuine desktop two-pane breakpoint and mobile-first single column, search across passages/lines, and an Anki-TSV/JSON export path, all with clean console output and no boot errors. | 1 |
| Wordhoard | 69/81 | Education | Everything works and reads as finished except the wide layout, which parks a Today-sized review card and a thin still-to-come rail in the center of 1366px with a dead grey band filling the rest of the viewport. | 1 |
| Winnow | 69/81 | Engineering | Strong, deep build (69/81) but sits below the 72 keep bar, and the data layer is entirely a static curated catalog with no import/fetch path even though profile detection is exactly the kind of thing that could be pulled from the user's real machine. | 0 |
| Callsheet | 69/81 | Education | Strongly built (85/85 tests, clean console, algorithm matches spec exactly, real CSV export, well-composed dark/light-of-projector palette) but usefulness caps it: the audience is a narrow slice (prosopagnosia / weak face memory), not a general teacher-tutor-trainer daily tool. | 1 |
| Kotoba | 69/81 | Education | Post-improve build is strong across the board (48 phrases, 27 noun cards, drill mode, two-pane desktop, clean console/tests) but the vocabulary still has no way out of the device: custom words the user types in can't be exported or imported, so the one place this app touches real user data is still a dead end. | 1 |
| Parley | 68/81 | Hospitality & Leisure | A finished, well-crafted persuasion game with no found defects and real depth (20 scenarios, 340+ authored lines, earned reads, restraint gating), but the desktop pane still leaves real width unspent and the game has no non-content source of substance, which caps it short of a 72+ keep. | 0 |
| Invoke | 68/81 | Engineering | A well-built, correct composer (35 tools, real round-trip lexer, clean tests, no console errors) that still has room short of a 72+ keep: no way to export or share a preset off-device, and desktop's Library rail sits static/empty in the resting state instead of surfacing anything useful before the user saves. | 0 |
| Convect | 68/81 | Hospitality & Leisure | Strong, deep app (42-symptom catalog, real least-squares+Shepard field solver, per-rack drift tracking, clean console) but desktop leaves a real gap on the Symptom Catalog screen that a two-pane layout would fill, keeping the total under the keep bar. | 0 |
| Hush | 68/81 | Finance & Legal | Strong build (real 4-stage decay/clustering engine, 83 passing tests, clean console) held down by Today wasting the desktop viewport with a 900px centred column while Case/Patterns correctly split into two panes. | 0 |
| Deadfront | 68/81 | Trades & Field | A deep, polished, bug-clean trainer (52 faults across 5 categories, real SM-2 scheduling, a searchable code index, CSV export) held back only by one acknowledged desktop layout gap, so it lands just short of keep on both the numeric bar and an honest per-axis look. | 1 |
| Cadence | 68/81 | Operations | One concrete gap left after the last pass: buildCadenceTheme() ships a single light ThemeData with widgets reading hardcoded AppColors, so there is no dark theme at all, capping palette below the top band. | 1 |
| Plumb | 67/81 | Health & Care | Strong, deep, well-tested app (127/127 tests, empty console, refine pass killed real defects) but every axis lands at 6-8 rather than 8-9, and routine entry is manual-only with no import path, keeping the total below the keep bar. | 0 |
| Reprieve | 67/81 | Finance & Legal | A genuinely deep, correct solver app (56-chore catalog, real 0-1 knapsack, bisection break-even, ledger with a stacked-area chart) held back by two concrete gaps: household inputs have no lower-effort entry path than typing every field, and the desktop layout leaves the left pane's lower half empty at 1366x728. | 0 |
| Idlewild | 67/81 | Hospitality & Leisure | Strong build (73-entry catalog, real MMR-reranked scorer, tags/tune persist correctly, no console errors) but the catalog is entirely compiled-in with no freshness signal and the wide desktop pane wastes real vertical space, so it falls short of finished on two axes. | 0 |
| Amble | 67/81 | Hospitality & Leisure | A polished, bug-free breeding-and-gait-solver toy, but nothing a user breeds can ever leave the app (no export or share of genome/image), and only one of the shelf's five roles has any real reason to open it at work. | 0 |
| Onshift | 67/81 | Engineering | Genuine typed state machine (goal predicate, traps, 40 scenarios across 8 categories/4 tiers) with clean cold-boot onboarding, but the desktop pane leaves most of the 1366-wide viewport as empty dark surface at rest, which the minimal doctrine caps. | 0 |
| Thock | 67/81 | Creative & Media | Craft, depth and correctness are all strong (67/81) but nothing produced in-app (a favorite, a tuned switch, an A/B winner) is usable anywhere outside it, so it stays a diversion rather than something a sound engineer or producer returns to for work. | 0 |
| Demur | 67/81 | Creative & Media | Deep, correct, well-tested composer (242 lexicon entries/7 categories, 48 scenarios/7 categories, working habit decay and history rescoring) held back only by a desktop layout that doesn't use its width beyond the two-pane editor. | 0 |
| Redress | 67/81 | Finance & Legal | Polished, correct, deep build (42-operator engine, real persistence, clean console) but usefulness is capped by a shelf-role mismatch: a personal commuter-refund tool has no hook into a bookkeeper/founder/paralegal's actual work (expensing, client billing, case files), so total falls under the keep bar. | 0 |
| Outrigger | 66/81 | Trades & Field | The solver, catalog depth (46 profiles, 24 bends) and instrument are genuinely strong and bug-free, but the app has no export/share path and no memory of past runs, so there is nothing pulling a driver or trainer back after the first session's insight lands. | 0 |
| Calque | 66/81 | Education | Deep, correct, well-built app (139 idioms/19 languages/7 themes, a real weighted-decay engine, a genuinely useful tutor handout export) held back by one concrete real-estate doctrine violation on mobile. | 1 |
| Basalt | 66/81 | Engineering | Genuinely strong compiler tool (real lexer/parser/codegen, 44 snippets across 10 categories, error recovery, clean console) held back by mobile toolbar polish and a data layer that never lets the user round-trip an existing file, not by any core defect. | 0 |
| Pinch | 66/81 | Hospitality & Leisure | Polished, deep, bug-free build (45 dishes, 126 faults, real Elo model, clean console/tests) but usefulness caps it: a chef/bartender/host mid-shift has no reason to open a seasoning trainer at work, so the raw axis total lands short of keep despite strong craft. | 0 |
| Gloss | 66/81 | Finance & Legal | Strong, already-improved reference tool (70-term corpus, real search ranker, two-pane desktop, clean console/tests) but the content is one-way lookup with no way to hand a definition to the person you're explaining it to, capping usefulness for a founder mid-task. | 1 |
| Fanout | 65/81 | Trades & Field | Craft, depth and correctness are all strong (52-build catalog, real event-driven simulator, clean probe, deliberate palette), but the product is a hobbyist digital-electronics toy and none of this shelf's roles (builder, mechanic, landscaper, vet tech, surveyor, driver) would open it at work. | 0 |
| Rundown | 65/81 | Operations | Genuinely deep, well-built app (two real algorithms, 16 ops-tailored templates, paste-import, session planner, drag reorder, 87 tests, no console errors) but averages 7.2/9, below the 8/9 keep bar, and lacks a couple finishing touches an editor pass can add. | 2 |
| Runway | 65/81 | Finance & Legal | Real algorithm and a genuinely good build already, but entries can't be edited once logged and there's no export path off the data, both concrete and fixable. | 1 |
| Bailey | 64/81 | Engineering | Deep, real algorithm (65 node types, 6 starter grammars, 4 layout modes, 6 constraint kinds, seeded two-stage pipeline, 188 passing tests) undercut by a visible label-overlap bug on the very first screen and a shelf fit that skews hobbyist-fantasy rather than engineering. | 0 |
| Skein | 64/81 | Research & Policy | A genuinely deep, well-built clustering tool held back by a real desktop layout waste and a missing exit path for what it clusters (no search, no export/copy). | 0 |
| Sinew | 64/81 | Creative & Media | Deep, working app (206-entry lexicon, 6 flag categories, passive/adverb/nominalization/AI-tell detectors, multi-draft save+file import/export, 119 green tests, clean analyze) but mobile/desktop are unverified against the current build since the only probe screenshots on disk predate a 24-file improve-app pass that added the drafts rail and notes panel. | 1 |
| Sprocket | 64/81 | Hospitality & Leisure | Deep, correct, well-composed puzzle game that opens directly on the payoff, but the desktop layout leaves a large dead band around the board on early levels, which is what a real visitor sees first. | 0 |
| Swell | 64/81 | Operations | A substantial rewrite (real interval planning/logging with drift analysis layered onto the original tide/calibration app) that scores solidly on every axis, but the improve-app session that produced it ended without writing its own result file or a fresh verify pass, so the current code's correctness is unconfirmed rather than demonstrated. | 1 |
| Brigade | 63/81 | Hospitality & Leisure | Deep, well-built puzzle (60 dishes, 50 menus, a real branch-and-bound RCPSP solver, clean Hive persistence, 123 green tests) held back by concrete, fixable UI gaps: desktop wastes most of the viewport below the fold and mobile labels truncate mid-word. | 0 |
| Garland | 63/81 | Trades & Field | Craft, depth (36 presets/6 categories, layer+gradient engine, persistence) and correctness are all solid, but desktop wastes half the control pane and the app has no export path, both concretely fixable. | 0 |
| Nib | 63/81 | Engineering | Strong finished core loop (44-challenge catalog, real coverage+color match scorer, real SM2 scheduler, clean verify) held back mainly by three of four shell screens wasting the 1366 viewport as a centred capped column. | 0 |
| Snapchart | 63/81 | Research & Policy | A genuinely well-built parser tool (63/81) but single-series-only charts and a library with no search/rename cap features and usefulness for repeat work use, well short of the keep bar. | 1 |
| Superpose | 62/81 | Engineering | Solid, correct WFC solver and instrument-panel visual design but truncated weight-slider labels on the one control that lets an engineer actually steer the algorithm, and no export path for the thing an engineer would want to take with them (a working ruleset). | 0 |
| Assay | 62/81 | Engineering | Core loop, algorithm depth and content (154 questions, 14 topics, real Brier/Elo engine) are strong, but the desktop layout wastes most of a 1366px viewport on a centered 760px column with no second pane. | 0 |
| Braid | 62/81 | Education | A genuinely deep, correct app (real SM-2 math, 252 curated cards, clean persistence, no bugs found) held back by one real gap (no import path for a learner's own vocab lists, only single manual add plus raw paste) and wasted vertical space on the review card face. | 0 |
| Broach | 61/81 | Health & Care | Real signature algorithm, 111-entry catalog well past the 60-entry floor, and a refine pass already fixed three genuine data-flow bugs, but the desktop wide layout leaves most of the 1366 viewport empty and there is no fast-entry path (barcode/photo of the leaflet) for the manual-entry-heavy add flow. | 0 |
| Provenance | 61/81 | Finance & Legal | Signature tracing engine, persistence, sync and paste-import are all genuinely solid, but the app has no way to get the traced figure out to a lawyer or court, which is the actual point for its target user. | 0 |
| Caliper | 61/81 | Health & Care | A genuinely well-built, spec-faithful measurement tool (real 300-point norm tables, exact-age interpolation, two-pane desktop, JSON/CSV export/import, 110 tests) with one real correctness defect and no way in without manual entry, so it lands short of keep rather than being demoted. | 1 |
| Tincture | 61/81 | Hospitality & Leisure | A genuinely deep, well-crafted formulation tool held back from keep by a wasted desktop pane and a slightly narrow usefulness ceiling, not by anything broken. | 0 |
| Elbow | 61/81 | Engineering | Genuinely good drafting tool with a real routing algorithm and clean persistence, but desktop leaves a huge empty paper field with all controls compressed into a thin top bar, wasting the one axis where a wide canvas should be an advantage. | 0 |
| Coldframe | 61/81 | Trades & Field | Sound algorithm (DST-safe frost-anchored schedule engine, 52-crop curated database, succession cutoff math) and clean tests/analyze, but built for a solo home gardener while the shelf's role is a working professional, and mobile/desktop could not be visually confirmed. | 0 |
| Lumen | 60/81 | Hospitality & Leisure | Solid, bug-free harmony-idle sim with real algorithmic depth, but desktop wastes two thirds of the viewport on a centered column and the shelf already has 9 elected idle/companion-adjacent games, so it needs the desktop fix and a sharper hook to earn a slot. | 0 |
| Phial | 60/81 | Hospitality & Leisure | Solid, bug-free color-sort puzzle with a real solver-verified generator, but the tier ladder is hard-capped at 6 tiers x 6 levels (36 total, endless mode out of scope) and the desktop layout wastes most of the 1366px viewport on a tiny centered board. | 0 |
| Lull | 60/81 | Hospitality & Leisure | Genuinely polished, bug-free bedtime toy with a real algorithm, but desktop is a phone column stranded in a wide field and it has zero personalization path for a returning host role (no custom theme/word list, no export/share of a session), capping how often the Hospitality & Leisure roles come back to it. | 0 |
| Tend | 59/81 | Trades & Field | Real algorithmic depth (robust-median learned-interval engine, milestone-gated phenology, 45-crop catalog, 12-plant deep demo garden) and clean execution, held back by a desktop layout that wastes half the viewport and a landscaper-shelf fit that has no way to separate multiple gardens/clients. | 0 |
| Bloom | 59/81 | Hospitality & Leisure | Real signature algorithm and deep curated catalog (75 ingredients, 5 dishes, 14 principles) hold up well, but a concrete mobile defect hides the return-visit progress indicator and the app has no way to save or share a plated sequence once built. | 0 |
| Trafila | 59/81 | Hospitality & Leisure | A well-engineered, symmetric pairing engine with real curated content (45 shapes/22 sauces) and clean bug-free execution, but it caps out as a browse-only curiosity: no search across 67 items, no way to take a pairing result anywhere, and a desktop layout that just centers a phone-width column in a wide window. | 0 |
| Tally | 59/81 | Finance & Legal | Strong signature algorithms and honest depth (25 categories/5 groups, 3 correctly-implemented money algorithms with a fixed reentrancy bug), but every top-level screen wastes desktop width with a centered capped column instead of a real two-pane layout. | 0 |
| Decant | 59/81 | Hospitality & Leisure | Genuinely well-crafted apothecary puzzle with a real BFS solver, but desktop wastes most of the viewport on both screens and the fixed 24-level ladder is a one-and-done: neither is hard to fix. | 0 |
| Crucible | 59/81 | Engineering | Genuinely deep engine (Tarjan SCC cycles, memoized cost roll-up, inventory solver, tiering) with careful derived-state reconciliation and 44/44 passing tests, but no probe screenshots exist to confirm the visual/mobile/desktop execution and the blueprint has no export path, both fixable and both would move the score. | 0 |
| Rosetta | 59/81 | Engineering | Solid signature algorithm and clean architecture, but ships two hard-rule copy violations, an About-as-bottom-sheet violation, and a below-floor text contrast that are all cheap to fix and would move the score. | 0 |
| Cubit | 58/81 | Trades & Field | Strong signature algorithm and layout, but three user-facing strings violate CLAUDE.md's dash-punctuation copy rule and should be a quick, high-value fix. | 0 |
| Inkling | 58/81 | Hospitality & Leisure | Solid, bug-free daily-puzzle core (deterministic generator + uniqueness solver actually works and the phone UI is well composed), but the desktop layout wastes most of the viewport and the signature 'hidden picture' payoff only exists for 3 of 7 weekday sizes. | 0 |
| Fallow | 58/81 | Engineering | Real fatigue-with-recovery algorithm, a genuine git-log parser, and a working Hive snapshot compare hold up well, but the fixed 1080px content cap wastes desktop width even in the dual-pane success view, there is no dark theme at all, and file-only import adds friction to what should be a quick check. | 0 |
| Tessera | 57/81 | Engineering | Real, working WYSIWYG export pipeline and clean data layer, but a live JS crash on every boot capture and a hard-rule About-sheet violation are both concrete, fixable defects. | 0 |
| Minder | 57/81 | Trades & Field | Solid, well-engineered core (rrule expansion + now/next-up state machine, FM sweep already done, 31 passing tests) but wide viewport idles unused, task entry has no quick-start path, and the prep checklist ignores the species it already captures. | 0 |
| Pixloom | 57/81 | Creative & Media | Real flood-fill/onion-skin/quantize engine and clean persistence, but the palette can never hold an arbitrary color — 'Add color' just cycles 6 hardcoded presets, so no pixel artist can match a specific skin tone or hex value, which is disqualifying for the target user. | 0 |
| Mince | 56/81 | Hospitality & Leisure | Real, well-tested algorithmic core (48-unit table, exact Rational fractions, kitchen-fraction snapping, cross-recipe aggregation) but only 5 demo recipes and a desktop layout that leaves the right two-thirds of a 1366px viewport nearly empty. | 0 |
| Vernam | 56/81 | Engineering | Bug-free algorithmic core with real depth (12 scenarios, 53-word bank, degenerate-input-safe solver, 56 passing tests) but the desktop layout wastes width and the app gives an engineer no reason to reopen it after the first walkthrough. | 0 |
| Sear | 56/81 | Hospitality & Leisure | Solid, correct, well-tested reference app with real depth, but desktop wastes most of the 1366px viewport on a centred 820px reading column with no secondary content, and no screenshots were captured to verify the visual craft claims. | 0 |
| Payload | 56/81 | Engineering | Real, well-engineered SM-2 + exploit-simulation core (exact SM-2 formula, 5 independent Dart simulators) let down by every non-empty-state screen being a single unconstrained column that just stretches edge-to-edge at 1366px, and the probe never got past the boot screen so the real UI is unverified. | 0 |
| Longhand | 56/81 | Creative & Media | Real, well-engineered signature algorithm and clean code, but the settings screen carries a hard-rule copy violation (offline/privacy boilerplate + em-dashes across five screens) and desktop never uses its width beyond a centred 680px column. | 0 |
| Stint | 55/81 | Operations | Correct, well-tested core loop and a genuinely good decay algorithm, but the app stops at exactly the spec floor: one algorithm, 16 seed rows, no export/import, and a desktop layout that just widens the margins instead of using the width. | 0 |
| Delve | 55/81 | Operations | Solid push-down-stack core with real timestamp-derived math and clean tests, but the desktop layout wastes most of the 1366 viewport on two of three screens and a session can never be deleted once resolved. | 0 |
| Ember | 55/81 | Operations | Genuinely deep, correctly-implemented knapsack planner and freshness decay let down by a pervasive light-theme contrast bug and repeated em-dash copy violations that an editing pass can fix without touching the algorithm. | 0 |
| Proofmark | 55/81 | Creative & Media | Genuinely deep, well-built detector engine (36 real detectors, correct density-normalized scoring, real persistence) undercut by two concrete, fixable defects: the app's own UI copy is riddled with em-dashes (the exact tell its own connective-density detector flags), and the writing-surface card doesn't fill available height on either mobile or desktop, leaving large dead space below the fold. | 0 |
| Quanta | 55/81 | Hospitality & Leisure | Well-built, algorithmically genuine ECS/spatial-hash instrument with a real copy-standards violation (em dashes in user-facing strings, including the gallery one_liner) and a shelf mismatch that caps usefulness regardless of polish. | 0 |
| Markwright | 55/81 | Engineering | The core incremental-markdown engine and editor UX are genuinely well built, but the web build (the only shipped artifact) never actually overwrites the original file on save — it downloads a copy while the UI reports 'Saved' as if AC-06 held, which is exactly the in-place-editing gap this app was commissioned to fix. | 0 |
| Elementa | 55/81 | Hospitality & Leisure | Solid, algorithmically real recipe-graph toy (52 elements, 48 recipes, BFS depth, non-dead-ending hints) let down by a wasted desktop layout and two unexplained boot-time pageerrors on every load. | 0 |
| Swipewise | 54/81 | Finance & Legal | The optimizer engine and feature set are genuinely deep (argmax lookup, gap detection, redundancy, weighted coverage score, prospective preview, all unit-tested), but the wide-viewport shell wastes roughly 500px of a 1366px window behind a fixed 720px content column even in two-pane NavigationRail mode, and user-facing copy uses em dashes as sentence punctuation throughout, against CLAUDE.md's hard copy rule. | 0 |
| Sluice | 54/81 | Engineering | Correct, well-tested four-limiter engine and real persistence, but the idle scope renders as a near-empty black void that wastes ~40-45% of both the phone and desktop viewport, and three user-facing strings break the no-dash-punctuation copy rule. | 0 |
| Ligature Lab | 54/81 | Engineering | Trie/engine, persistence and validation are all genuinely solid, but the RulePanel (the second half of the product) defaults closed even at expanded desktop width, so the 1366 screenshot is an empty canvas in a sea of unused space instead of the docked two-pane layout the spec itself calls for. | 0 |
| Sillage | 53/81 | Creative & Media | Strong note-vector concierge algorithm and a real 48-entry curated catalog are undercut by a broken desktop library render (entirely empty body at 1366x728) plus dead whitespace on every card. | 0 |
| Mise | 52/81 | Hospitality & Leisure | Real algorithmic core (exact-rational parser/scaler/unit-ladder, well tested) is undercut by a desktop layout that never uses the extra width and by CLAUDE.md-banned em-dash copy shipped in five user-facing strings. | 0 |
| Larder | 52/81 | Hospitality & Leisure | Solid constraint solver and clean state model, but desktop never breaks into a wide layout despite the ui-spec promising it, and the 32-recipe bundled pool is thin once diet/allergen filters remove most of it. | 0 |
| Tracer | 50/81 | Engineering | Solid FSM/architecture and clean bugs axis, but the bundled vocab/output pool is too thin for repeat-visit depth, desktop leaves most of a 1366x728 viewport black, and user-facing copy violates the house no-em-dash rule in several strings. | 0 |
| Nocturne | 49/81 | Health & Care | Every cold boot crashes before the real app mounts (Icon/Text under the loading and error screens have no Directionality ancestor), but the underlying product (coupled debt/consistency/cycle math, direct-manipulation rail, two-pane desktop layout) is strong and the fix is small and precisely locatable. | 0 |
| Pip | 33/81 | Hospitality & Leisure | BuddyGame (FlameGame) never sets backgroundColor(), so Flame paints its default opaque black over the entire WindowsillScene beneath it in the Stack, so every screenshot is a black void with floating text and a green blob, none of the spec'd circadian palette, window, or sill art ever renders. | 0 |