Skip to content
AptifoldGet in touch

PRIVACY, MODEL & CREDITS

A closer look.

What happens in your browser

The interactive modules calculate locally. These pages include no account system, advertising scripts, application analytics, embedded payment widget, or AI API connection.

Some tools keep settings in the page address. The Reading Room, Kitchen Lab, Animal Behaviour, Sound Studio, Ocean Notebook, Plant Conservatory and The Missing Signal use the query string after the ? mark. Changing a control updates the address without making a separate page request. Opening or reloading that address sends its query settings to the web host; sharing it makes them available to the recipient. Use the supplied examples rather than private operational data.

Setup links for The Gravity Race, The Last One Back, Movement Foundry and Shadow Observatory put settings after the # mark. Ordinary requests to the web host do not include that fragment. Resilience Designer saves its editable setup in a file you choose; it does not upload those inputs. A planning comparison link can name a built-in example in its query string, which the host can see. No account is needed to try these modules.

Serving a website involves ordinary request information, such as an IP address. Hosting-provider logging and policies are separate from the local calculations described here.

Buy Me a Coffee, email, and reference links take you to separate services with their own policies. Support is optional and handled on Buy Me a Coffee. No payment is processed by this site.

Movement Foundry

A rotating disk and a vertically guided knife-edge point illustrate how a requested rise, hold and return can determine a shape. Dragging the wheel or the lift trace inspects the same angle. Playback shows one optional six-second turn; it does not calculate an operating speed.

The bench computes the minimum base radius in its declared family, considering the requested motion, minimum base and contact-angle limit. Adding lift gives the required outer radius. Reducing the allowed radius does not redesign the part. An impossible brief keeps the required shape visible; invalid inputs withdraw the design. A saved setup preserves inputs and inspected angle, with motion paused.

This is an ideal maintained-contact kinematic model, not a roller follower, force or spring design, safety assessment, or manufacturing plan. Its equations, numerical limits, model source and continuous bound remain available in How this works. CMU’s Introduction to Mechanisms explains the underlying inversion and pressure-angle principles.

The Factory in the Box

Follow prepared inputs through a fictional sterile fill-finish line. Filling a batch does not release it for use: the material and its records must both pass the release gate. The line does not make the active ingredient.

Supply compares three fictional plans with the same daily need and outside deliveries. Only the reserve plan starts with extra stock. Local production adds usable stock after the declared release interval; orders missed earlier remain missed. The day ledger distinguishes held, released, expired and timely supplied units. Batch, Supply and Capacity are separate teaching models with separate inputs, not one linked simulation.

The real shortage case provides context, not calibration. No model here specifies a medicine process, predicts clinical outcomes, establishes compliance or prices a factory. The Capacity model has different illustrative assumptions. Read the two fixed supply cases or inspect the assumptions in each room. The browser stores only the optional walkthrough motion preference, under factory-step-scroll.

Factory retains its bundled Three.js 0.186.1 notice, separately from the Gravity Race renderer.

One connected run

The separate connected Factory profile joins production, minimum waiting, finite review work, whole-batch release and usable stock in one 30-day ledger. A unit stands for one imagined filled container. It explicitly declares one production cycle/day; the original Capacity room does not define a daily rate. Its separate native rooms keep their earlier assumptions.

Keep plan A while editing B against the same daily need. Making more can increase held material without covering more need. Partial review work does not release a partial batch. Expiry occurs before review and daily use; missed need remains missed. The selected day, batch trace and paired stock/need plots read the same ledger. Exact JSON and CSV keep the inputs, units, event order and both runs.

Qualified inputs/utilities, complete records and successful reviews are assumed. A separately written one-work-unit scheduler checks the bookkeeping, not a real process. This fictional profile supplies no clinical, manufacturing, quality-approval, pricing or operating forecast.

Ocean Notebook: choose the years

A separate workshop compares the earlier and later calendar-year halves of a visitor-selected window in the same frozen 355-row Station ALOHA surface-carbon record. Missing measured values are not filled with calculated values. Each dot is a source cruise mean, with source notes retained.

One comparison gives each available cruise mean equal weight. The other computes separate earlier/later means for each calendar month available in both halves, then gives those shared months equal weight. This changes the included observations and their weights; absent months are excluded. It is a descriptive check of one sampling imbalance, not a trend estimator, seasonal model, significance test, causal attribution, forecast or organism-response calculation. The plotted vertical pH scale is fixed within a source column.

Local links and JSON preserve the window, source column, plotted weighting and inspected month/cruise. Reopening recalculates against the frozen source and ignores saved result/model claims. CSV retains each selected source row, null pH and its comparison weights. The 2016–2020 measured-at-25°C reversal is a selected example from this record, not a population performance claim. HOT source definitions, sampling exceptions and attribution ↗.

Resilience Designer: two corridors at once

The separate joint-closure workshop extends the fictional booked-noon receiving example. Advance bookings stay fixed; each dispatch reroutes after its closure case is known. The original menu checks 32 plans against 7 no/single or 22 no/single/pair cases. Both harbour exits can close together, and none of the original downstream investments creates another exit.

An optional, explicitly hypothetical contract adds two three-hour direct routes at cost 9 reservation units. Both routes share one closure group and its failure is included. This extended menu has 64 plans and 8 or 29 cases. Costs are authored reservation units, not supplier quotes or total operating cost. No failure probability or independence assumption is assigned.

Complete search minimizes cost subject to a stated minimum total in every case. Coverage and budget remain separate. A partial target of four crates can leave one depot empty; a displayed dispatch is one maximum-total allocation, not a fairness rule or the whole-family result. The unchanged native flow engine, world-row flow/cut witnesses and separately authored crate-path reference check all 2,560 original/extended plan-case pairs. Saved inputs recalculate; the complete comparison and inspected certificates can be downloaded locally.

This is finite capacity, storage and receiving-window screening. It does not model surprise mid-run failures, fleet or crew execution, measured rates or operating approval. The original dispatch board retains its separate assumptions and experiment format.

The Gravity Race: a wire that fits

The separate design workshop fixes the original 6 m span and 2 m finishing drop. It checks maximum depth below the start and available wire length over 864 quadratic curves plus the original straight and cycloid references. The editable-only scope explicitly excludes those references; the all-shapes scope never hides a faster fitting known baseline.

Depth uses exact integer comparisons for custom controls. Length and cycloid-depth decisions reserve 10⁻⁷ m; numerical boundary cases are not silently admitted. Estimated Earth-like times within 10⁻⁷ s form a numerical tie with a stable representative. Moon-like gravity rescales time without changing geometry or ranking. This finite, numerically checked catalogue is not a continuous optimum, formal floating-point certificate, rolling-bead simulation or real-track safety design.

Preview does not replace track A. Apply and session-only Undo are explicit. The front-view diagram uses native physical coordinates and atTime at one shared physical instant; half-speed playback is labelled, saved restoration starts paused and the original 3D race stays linked. A separate 50-digit Decimal/Gauss check verifies all 866 scalar rows and selected design cases. That expanded check is unblinded and does not replace the original independently committed native baseline. Saved JSON recalculates inputs; CSV keeps the full finite catalogue.

The Reading Room: compare searches

The separate ranked-list comparison uses one question and a fully labelled collection. The opening has twelve authored paragraphs, five useful labels and invented ranks. It does not search real papers or run an AI model. Shared misses require reference labels outside the returned lists; returned lists alone cannot establish recall.

Each option receives the same maximum paragraph-reading budget. The known ranked prefix gives exact counts under supplied labels and correct recognition. Combined alternates rank slots in the stated A-first or B-first order, skips duplicate IDs without another read and drains the longer list. It is one declared policy, not an optimized ranking. Useful unread and useful never returned remain distinct. The unchanged native unknown-order bounds bracket these exact prefix counts. A separate set/prefix reference enumerates all 16,384 three-paragraph label/list/order/budget cases and replays saved calculations.

Local CSV and JSON imports require one question, 1 to 200 unique IDs, explicit useful/other labels, bounded text and unique contiguous ranks. Imported labels are unverified, derived JSON scores are recalculated, and imported text stays out of URLs. Save JSON to retain the custom collection. The downloadable CSV is always the fixed authored template. No general tool-performance, scientific reliability, calibrated label quality, confidence or reading-time claim is made. Original aggregate counts, source bindings, paper rows and saved setups stay separate.

Animal Behaviour: trial designer

The separate trial designer compares explicitly authored rules: prefer yellow, prefer dark, always left, always right and optionally choose dark yellow when offered. A rule with no preference permits either side, with no probabilities supplied. Twelve ordered layouts of four distinct labelled cards define the finite palette; a board holds one to six trials.

A pair of rules is forced apart if one trial gives them disjoint permitted sides. Exact subset search finds the smallest unused-layout addition that separates every pair, or distinguishes insufficient board space from no guaranteed separation within the palette. A separately written Cartesian response-pattern reference checks all 4,096 layout subsets and all 26 eligible rule sets. User-supplied hypothetical choices filter permitted patterns without generating animal observations, confidence or sample-size advice.

Screen colours are uncalibrated and the cue labels are supplied. Original study records, training/conflict mechanisms and saved notebook setups remain separate. This is not an exhaustive account of behaviour, an error model or a real-animal conclusion. Links and labelled JSON/CSV downloads retain the authored plan and hypothetical choices locally.

Kitchen Lab: mixed sizes

The separate mixed-size bench keeps the oil volume in 100 reference spheres of radius 40 micrometres. Each sample puts an integer percentage of that oil into spheres of radius 5, 10 or 20 micrometres; the remainder stays in 40-micrometre spheres. This explicitly uses 100 times the original equal-drop lesson’s reference volume and keeps every permitted population a whole number of spheres.

Keep A and edit B. Number share, oil-volume share and surface share have different denominators. Sum each group’s count times its sphere volume or surface. Exact fractions, common-scale total areas and an area-equivalent radius expose those differences; an arithmetic average radius cannot generally substitute for the distribution. An equal-surface example deliberately has different populations. Links retain both samples and inspection after the # mark; JSON and CSV keep the calculated populations and units.

The original geometry, drop inspection, motion and saved setups remain separate and unchanged. Native equal-size endpoints and a separately written exact-rational geometry reference check this declared model. The pictures are representatives, not microscope data or packing. No fluid, recipe, texture, emulsifier-dose, stability or food-safety prediction is supplied.

The Missing Signal

A fictional distance-only search asks which locations satisfy every stated reading. The four exact cases distinguish two possibilities, a useful third reading, conflicting readings and a fragile single answer. The terrain does not constrain the search.

The linked receiver-placement investigation holds A and B exact, leaving two fixed possibilities for one beacon. Only C has a chosen absolute error bound. Its allowed report ranges may overlap, touch or separate; touching endpoints still belong to both ranges. Testing a hypothetical report can retain both, one or neither location. No hidden true location, probability or calibrated instrument is supplied.

The separate all-reading-errors investigation gives each of the three reports its own closed distance band. Exact circular geometry returns one witness or a finite exclusion for their joint intersection. A separate finite map admits whole cells and keeps unresolved boundary cells explicit; exhausting its work cannot establish infeasibility. A witness is not a unique answer, and the drawing is not a probability, calibrated measurement or floor plan.

Inspect the exact cases or choose the next receiver position. The in-module methods retain equations, exact certificates, sources and assumptions. The tagged-equipment story illustrates the decision; it does not qualify an installation.

The Gravity Race model

Ideal point particles start from rest and slide along fixed wires under constant gravity. The wires span 6 metres and finish 2 metres below the start. Friction, air resistance, rolling and the strength of a real track are left out.

Height lost determines speed; the shape changes how early the bead gains that speed. A steep early drop can offset a longer distance. The steep-start reference is the classical least-time cycloid for these endpoints. Your editable wire explores a limited curve family, rather than every possible track.

This is a mathematical exhibit, not a measured race, an engineering safety assessment, or a model of rolling balls. The experiment’s “Why it works · model & sources” control opens the assumptions, formulas and calculation details.

Using the experiment

Make a prediction if you like, then reveal the arrival times or play the race. Use the two wire sliders to change your route; their arrow keys make small adjustments. Choose “Shape in 2D” for a planar diagram, or inspect a moment with the time controls. Animation is optional, and a reduced-motion preference is respected.

A saved link restores settings, not a recording of earlier attempts. Changing gravity starts a new comparison. If a shared link cannot be restored, the experiment explains the recovery.

The Last One Back model

In the original room, all service calls start together. Each takes 40 or 800 milliseconds, and the report needs every reply. Its 200 millisecond deadline lies between those times. Three authored probability models preserve the same individual call behavior while changing how slow replies coincide.

The separate Ready enough? completion desk permits partial reports explicitly. It finds the greatest required reply count meeting a minimum reply coverage and a maximum lateness risk. Unknown dependence uses the sharp upper bound; three named laws are optional alternatives, not a substitute for that bound. Coverage counts interchangeable replies, not stock quantity, value or specific critical warehouses.

The same selected configuration compares the all-replies rule and an inspected rule. Every reply tied at first completion enters the report; its issued contents then stay fixed after later arrivals. A chosen configuration is not a sampled workload, and no probability is assigned to it under unknown dependence.

Exact integer fractions decide both limits before display rounding. The local JSON calculation includes inputs, units, every rule and a replayable selected query. Independent rational checks qualify the arithmetic. These are illustrative systems models, not service measurements, forecasts, reliability certification or recommendations to synchronize failures. Retries, streaming updates, staggered starts and changing load are outside their contract. The native methods explain the formulas, examples and primary sources.

The Resilience Designer model

A fictional finite network moves whole crates through declared routes and receiving windows. Reserving capacity happens before the selected corridor closure is known. Routing then uses that known closure. A place to wait consumes declared storage.

The minimum-cost search examines reservation combinations in increasing cost order. It stops when a plan meets every order in every declared case; all cheaper combinations have been ruled out. A separate comparison can enumerate the full menu. Its cost is an abstract reservation budget, not a quoted price or total operating cost. The weakest case is a declared scenario, not a failure probability. It does not model a surprise disruption during a trip, vehicle operations, simultaneous failures or a real company.

Calculations run in a local browser worker. The model explanation, dispatch ledger and editable inputs stay in the tool. Save or load an experiment to preserve its declared setup; a saved file does not establish a current result until the model checks it.

Movement Foundry: a wheel for a roller

The separate roller workshop asks a shaped wheel to lift an illustrative gate, pause and lower it. It compares adding a roller to a point-contact outline with reshaping the edge for that roller. Both comparisons use the full selected wheel radius, rather than the original tool’s minimum-size shape.

The native quintic lift law defines a roller-centre path. Its inward normal offset supplies the working edge. Finite rollers are qualified only for a strictly convex centre path with a regular inward offset: outward-rounded adaptive intervals cover both motion segments. A too-large roller can have a rejecting witness; a nonconvex path is outside this checker’s family, not proof that all cams fail. Exhausting the check remains unresolved. Clearance and the existing continuous contact-tilt bound are separate requirements.

An independent finite reconstruction rotates each polygon and intersects its roller-expanded boundary with the guide. The 2-degree plot and polygon replay are approximations, separate from continuous qualification. The zero-roller endpoint uses the existing point model; no lift reduces both shapes to a circle. Maintained contact is assumed. Forces, springs, friction, wear, operating speed, manufacturing tolerances and physical qualification are outside the model. This is a chosen-size construction, not a global minimum-size search.

Local setup files and fragment links preserve inputs and recalculate on open. A qualified conceptual SVG contains millimetre polygon vertices. Rejected or unresolved constructions retain diagnostic JSON without a new outline. Sources: CMU, chapter 6.5.3 and Ablamowicz and Liu, section 5. The project’s restricted-family derivation and implementation are separate from those references.

Shadow Observatory: test the next dip

A separate timing workshop compares two stated clocks against the same 993 QLP readings. Only early dates fit each clock’s starting phase, baseline and nonnegative dimming. The fixed fitted values then predict every later source timestamp. The supplied 3.28883-day reference is not discovered here.

Both clocks use identical rectangular dips, a shared assumed duration, an approximate uniform half-hour exposure average and the same 512-point phase grid. The drop is bounded between zero and 0.05 normalized flux. Later mismatch is unweighted root-mean-square error, not significance or a planet probability. Background is not an error estimate. The source measurements are unchanged; no new clipping, detrending, normalization or fabricated readings is applied.

One dip per P and two equally spaced identical dips per 2P are exactly the same brightness template. This declared family cannot distinguish those descriptions. The calculation is not an orbit model, period-discovery search, radius estimate or planet confirmation. Inspecting later scores while revising settings makes this exploratory rather than blind validation. Local setup files/fragment links retain supported inputs; reopening recalculates results. A CSV retains all source rows, both predictions, residuals and fitted settings.

The Shadow Observatory model and observations

Move an opaque planet across a star and follow the same orbit position on the light curves. The gold curve assumes an evenly bright stellar surface; the dashed comparison uses a linear model of dimmer edges. Turning the camera changes your viewpoint while the telescope and calculated brightness stay fixed. The Sun and Earth artwork illustrates an invented system; it is not an image or physical model of the observed planet.

The real-starlight room retains all 993 original readings of WASP-126 from the MIT TESS Science Office QLP data, sector 1. It replays a known planet using a supplied 3.28883-day period. Aligning repeated dips helps explain the signal; it does not independently discover or confirm a planet. The original replay adds no clipping, detrending, fitted parameters or measurement errors. The original background column is retained in the download and is not a flux uncertainty.

Keep this setup creates a snapshot in the address fragment. It saves declared inputs and the selected reading, not a history of your actions. Guided tours and running motion are not saved; restored motion starts paused. Copying the link is optional. Without JavaScript, these notes and the original three-column readings remain available.

Data: MIT TESS Science Office QLP, Huang et al. (2020a,b), via MAST; dataset DOI. The QLP data retain a CC BY 4.0 license. Illustrative Sun/Earth texture artwork from Solar System Scope / INOVE also retains CC BY 4.0. The textures are rendered with lighting, color adjustments and a cloud layer. Sources & math and Credits inside the observatory retain its equations, limitations and third-party notices. No institutional endorsement is implied.

The recorded research case

The gravitational-wave page keeps the two saved notebook-reproduction runs and all seven numerical comparisons. It also displays the recorded fixed-template filter output and a follow-up timing-sensitivity calculation. Its controls inspect saved curves and local linear predictions. Separate synthetic ringdown studies compare uncertainty regions and mass-and-spin posteriors. An optional calibration study repeats the inference across 256 systems drawn from the stated priors; its numerical checks, countercases and limits are available alongside the results. The noise model, sampling rate and amplitude priors stay explicit. These pages inspect recorded calculations; they do not run the Python analysis or perform a new detection. The downloadable follow-up record retains both complete 16,383-point curves, the declared finite-change checks, source credits and scientific limits. The original discovery and public data belong to their credited scientific sources.

Credits and sources

Aptifold’s AI agent researched the sources, implemented the models, designed the visuals, wrote the code and ran the tests. Aptifold’s founder set the goals and reviewed the results. The physics and probability principles are established science. Sources and equations are linked in each experiment.

The locally bundled Three.js r180 library retains its complete MIT notice. Instrument Sans retains its SIL Open Font License and author credits. Aptifold’s name and mark identify the project.

For a correction or an access problem, email aptifold@proton.me.