Skip to content
AptifoldGet in touch

A research partner that builds what the work needs.

Aptifold is an AI partner for research, design and hard decisions. Email us about a project of your own. You set the limits and decide when to act.

Move a planet across its star. See the starlight dip in a model Aptifold built.

See what Aptifold has built

What can a dip in starlight tell you?

Illustrated planet crossing a star in the Shadow Observatory model
Illustrated transit · Artwork credits

Aptifold built a timing model and checked it on 603 readings it had set aside. Two different cycles predict the same dips. See why both fit the same starlight.

Compare the two predictions

Can both depots get their orders if two routes close?

Map of a two-closure example. Direct routes deliver four crates to each depot by noon.

Aptifold built a model of travel, storage and deadlines. In the worked example, a 17-unit booking guarantees four crates at each depot across the modeled closures.

Try the delivery planner

Can you build a sound from its parts?

Copper-coloured sculptural illustration of a sound wave

Aptifold made a playable instrument from a sound synthesis model. Mix tones, change how a note starts and fades, then hear and see what your changes did.

Play Sound Studio

What Aptifold takes on

What you can ask

Bring a research dispute, a design that won’t fit, a costly decision or a creative idea. Aptifold can work out what to check and build the model or tool the job needs. Each brief links to related work you can try or inspect.

Research audit

“Two teams used the same data and reached opposite conclusions. Rebuild both analyses and find which decisions changed the answer.”

Change one analysis choice at a time, from the filter to the rule for dropping readings. Give the team a calculation it can rerun and challenge.

Engineering design

“Replace this sliding contact with a roller. Keep the gate’s lift and timing. How large a roller can we fit in the same housing?”

Generate a new profile from the motion it has to produce. Check contact and clearance through a full turn and keep the outline for review.

Creative tools

“Make the same melody feel tender in one version and unnerving in another. Give me an instrument I can play, shape and share with another musician.”

Design how the mix of tones changes during a note. Build a playable prototype with editable sound settings and audio you can keep.

Experiment design

“Both explanations fit our measurements. We can afford one more experiment. What should we measure, and how precise does it have to be?”

Predict what each explanation would produce under a proposed test. Check whether the difference would survive the likely measurement error.

Business decisions

“Should we stock up or make locally if deliveries stop? Stock expires and a new line takes time to start. What would flip the decision?”

Model how much usable stock each plan delivers and when. Compare the cost of preparing now with the cost of running short, then vary the assumptions.

Software reliability

“Our report waits for the last warehouse. Can we send it sooner and still have every fact we promised the reader?”

Replay slow cases under different rules for how many replies to wait for. Check both the time saved and the information that would be missing.

Operations planning

“Two routes can fail at once. Find the cheapest plan that gets every depot its order by noon, or show which part of the promise breaks.”

Test bookings against the possible closures. Show the route or storage limit that prevents a cheaper plan from covering every case.

Games and teaching

“Build a mystery a player can solve by taking one more reading. Let a tempting choice fail and give the player a fair way to work out why.”

Generate clues from a consistent model. Let the player place a sensor and see exactly which locations their reading can rule out.

Everyday investigations

“Paper near one wall keeps getting damp, even though the room meter looks fine. Can two cheap sensors tell us what’s going on?”

Estimate humidity at the cold wall. Compare two spots, then swap the meters to see whether the higher reading stays at the wall or follows the device.

Browse all 13 engines and try changing the inputs in your browser.

Work alongside Aptifold

Start with a question in your own words, a folder of source material, a sketch, or a project already under way. Give Aptifold a goal to pursue independently, or work with it as you decide what to try and what to change.

Investigate a question
Aptifold can research the subject, trace claims to their evidence, and build a mathematical model. It can design and run a test to see which explanations hold up.
Develop an idea together
Try a prototype, shape a sound, bring new data, or change a constraint. Aptifold follows the change through the design, analysis and checks so you can see what it affects.
Take the work further
Get a recommendation with its tradeoffs, an experiment you can rerun, or a working tool people can try. Inspect the assumptions and checks, then carry the result into the next decision.

When the requirement changed

The requester changes the brief
Try two crates per depot instead of three. Keep the same deadline and closure risk.
Aptifold tests the new target
Reuse the same map, prices and closure cases. Check which bookings meet the smaller order.
What the test showed
Both depots still need a backup. The cheapest qualifying booking is still 17 budget units.

Is the cold cupboard too damp for paper?

The room’s humidity meter reads 50%. Water vapour is closer to forming droplets in cooler air, so the same amount of moisture can be a bigger problem near a cold wall. Aptifold used published data to estimate how much difference the wall temperature makes.

Example room air at 68°F
50% humidity
Estimate at a 50°F wall
About 95%

What you could do next. Put one humidity meter in the cupboard and one on an inside shelf. Compare both on cold nights, then swap them and repeat. If the higher reading moves with the meter, the meter may be off. If it stays with the cupboard, pay closer attention to that spot.

The timed practice run reached a tested tool in 6 minutes 26 seconds, with no follow-up questions for the requester.

How the estimate was made and checked

The calculation holds water vapour pressure constant and divides it by the lower saturation pressure at the wall. With the example’s ±1°F temperature and ±3 percentage point humidity errors, the estimate spans about 83% up to saturation. Actual airflow, leaks and local moisture can change the result.

An independent check caught an approximation error at a colder test temperature. Aptifold switched to a more precise equation and passed the same check at the same tolerance. The delivered calculator runs in a browser with editable inputs.

Paper storage guidance · Humidity equations · Independent vapour-pressure data

Can both depots count on a delivery?

Say you have 11 budget units to book transport to two depots. Can you guarantee three crates to each depot by noon if any two routes close? Aptifold tested every pair of closures in the model.

Guaranteed at 11 budget units
4 crates total
Guaranteed at 17 budget units
4 at each depot

At 11 units, a closure can leave one depot empty even if four crates get through to the other. In this example, even promising just two crates per depot costs 17 units. A different pair of closures can cut off each depot, so both need a backup.

How the planner tested each booking

Cap each depot’s order at the requested minimum and search for a booking that fills both orders in every case. A full sum then certifies that both depots get their share. A separate whole-crate routing calculation reproduces the minimum costs across the 29 expanded-menu cases. The options and prices are illustrative.

Read the mathematical model

Try the delivery planner

In four timed practice runs, Aptifold reached a checked result without the requester having to step in and correct the work. The cupboard run began without a supplied formula or method. The other three used earlier work and each included one planned change from the requester.

The operating system behind Aptifold

The owner sets the objective and grants authority for the work. Aptifold turns that objective into conditions a result must meet, chooses a method and builds the model, experiment or tool. It keeps the assumptions and evidence with the result so the owner can see why the decision follows and what could change it.

From brief to checked result

Follow the acceptance criterion, the work Aptifold can carry out, and the evidence the owner can inspect before acting.

Project examples

Starting brief

Will smaller orders make the guarantee cheaper?

The requester lowered the order at each depot from three crates to two. Find out whether that change lowers the minimum booking cost.

Alignment

Acceptance criterion

Deliver at least two crates to each depot by noon in every declared closure. Keep the routes, prices and closure set fixed.

Delegated authority

Permitted work

Model the network, search the booking options and check the delivery guarantee.

Owner decision · Set the delivery promise and approve any actual booking.

  1. Reasoning
    Test the promise at each depot

    A total can conceal an empty depot. Count deliveries at each depot only up to its own order, so the full capped sum requires both orders to be filled.

    3 to 2 crates per depot

  2. Build
    Recompute the bookings

    Represent travel and waiting as capacities over time. For each booking, calculate what can arrive by noon in each closure.

    64 plans × 29 cases

  3. Verify
    Match a dispatch to a bound

    A feasible dispatch shows what can arrive. A capacity cut bounds what can arrive. Matching values certify the modeled maximum in that case.

    Feasible dispatch and capacity cut

Verification result

Optimum within the model

The minimum stays at 17 budget units.

Different closures can cut off different depots. Both still need protection even with the smaller orders.

The 11-unit booking under two modeled closures
CriterionDelivered by noonResult
Four crates across both depots4 totalPass
Two crates at each depot0 north, 4 southFail

Run node aptifold-delivery-verification.mjs to recompute the results with the inputs in the file.

Independent replay
A separately written whole-crate routing calculation checks totals through the same modeled network.
Coverage
The declared set covers zero, one or two route closures, including loss of the backup service.
Scope of claim
The minimum depends on the listed routes, prices, capacities and closures. Confirm those inputs before booking.
Revise and recheck
Rerun the affected work

For another order change, keep the network, prices and closure cases. Recheck which bookings satisfy the new orders before recommending one.

Handoff for review
A checked booking comparison

See how the two requirements change the delivery guarantee and which cost stays fixed in the model.

Read the revised-order case
Technical details and limits

∀ c ∈ C, F2(b, c) = 4

F₂ is the maximum on-time flow for booking b and closure c when each depot can receive up to two crates. A flow of four certifies that both orders are filled. The search selects the cheapest qualifying booking in the declared menu.

Starting brief

Can these dips tell two explanations apart?

Start with archived brightness readings and two possible schedules for when a star dims.

Alignment

Acceptance criterion

Distinguish the clocks only if their fixed predictions differ in a way the later readings can resolve.

Delegated authority

Permitted work

Fit the two stated patterns using early public readings and check both against the later data.

Owner decision · Decide whether to pursue a different observation.

  1. Reasoning
    Compare two clocks

    Test one dip in a short cycle against two evenly spaced dips in a cycle twice as long.

    3.28883 and 6.57766 days

  2. Build
    Fit both predictions

    Use the early readings to estimate when the dips start, how deep they are and the star’s baseline brightness.

    390 early readings

  3. Verify
    Test the held-out readings

    Keep both fits fixed. Measure the gap between each prediction and the readings set aside for this check.

    603 later readings

Verification result

Inconclusive between clocks

The two predictions coincide.

The longer clock gives one dip at every turn of the shorter clock. These readings leave both explanations possible.

Independent input
The later readings are separate from the fit. The exported residuals show the errors at each observation.
Analytical check
Divide each period by its number of identical, evenly spaced dips. The repeat interval is the same.
Scope of claim
The result is about these two timing templates. A measurement they predict differently could resolve the ambiguity.
Revise and recheck
Look for a different measurement

A test that can choose between the clocks needs an observation on which their predictions differ.

Handoff for review
A timing workshop

Inspect the two predictions, all 993 readings and the errors at each reading.

Open the tested timing setup
Technical details and limits

Repeat interval = cycle length ÷ number of dips

The formula assumes identical, evenly spaced dips. The two clocks share the same repeat interval and fitted brightness template. The shorter period comes from the known-planet record. The comparison checks timing in this model.

Starting brief

Why are the uncertainty estimates too confident?

Simulate black-hole signals with known amplitudes and add noise. A well-calibrated 95% uncertainty region should contain the known values on about 95% of trials.

Alignment

Acceptance criterion

Compare coverage against the four known amplitudes. Pair the same signals, noise draws and filter across both methods.

Delegated authority

Permitted work

Use synthetic recordings to test how filtering affects the reported uncertainty.

Owner decision · Decide whether the model and evidence support use in an observational study.

  1. Reasoning
    Isolate the noise assumption

    A filter can change how errors move together. Compare the effect of that change with narrow interference and with white noise alone.

    Two noise settings

  2. Build
    Model the filtered noise

    Filter the recordings. Compare reusing the old noise model with recalculating how the filter changes relationships between errors.

    Same filter, two noise models

  3. Verify
    Check against known values

    Count how often the uncertainty region contains all four amplitude values used to make the signal.

    2,048 paired trials per setting

Verification result

Correction checked in simulation

84.5% → 94.8% coverage

The revised noise model raises coverage with narrow interference. With white noise alone, both methods land near 95%.

Known ground truth
The four signal coefficients are fixed before noise is added. Each repeat can be scored against those values.
Control case
Using white noise alone would hide the larger miss that appears with narrow interference.
Scope of claim
Coverage here applies to simulated signals with a known start time and the stated noise model.
Revise and recheck
Keep a test for the next filter

When the filter changes, recalculate its effect on the noise and repeat the coverage check against known answers.

Handoff for review
A comparison with a control

Inspect the effect in both noise settings and see where the correction changes the result.

Inspect the coverage results
Technical details and limits

Filtered noise covariance = HΣH⊤

Covariance (Σ) describes how errors in different samples move together. The revised model carries the filter (H) through this calculation as well as applying it to the signal. Coverage is the fraction of simulated trials in which the region contains the known values.

Starting brief

Can a note change character while it plays?

Give the same melody two distinct voices. To develop one of them, start with a note that begins bright and mellows as it fades.

Alignment

Acceptance criterion

Both versions start with the same ingredients, pitch and overall fade. Change how the higher tones fall away after the peak.

Delegated authority

Permitted work

Generate, compare and export synthetic notes from the selected settings.

Owner decision · Choose the sound that fits the music and decide how to use it.

  1. Reasoning
    Hold the pitch steady

    Compare two versions of the same note. Change only the fade applied to the higher tones so you can tell what changed the sound.

    220 Hz in both versions

  2. Build
    Give the higher tones their own fade

    Make a waveform by mixing the base tone with higher harmonics. Once the note peaks, give those higher tones a faster fade.

    120 ms to halve the higher tones

  3. Verify
    Compare with the fixed mix

    Compare the calculated mix at the same moments. Turn off the extra fade and check that the engine still gives the original samples.

    Sample-for-sample match at three rates

Verification result

Controlled signal comparison

The higher tones fade faster. The pitch stays fixed.

At each moment, 100% means the higher tones are as strong as in the fixed version. The whole note gets quieter in both versions.

At the peak
100%
120 ms later
50%
240 ms later
25%
Reference
With the extra fade off, the new generator matches the original samples at 8, 24 and 44.1 kHz.
Replay
Reopening a saved recipe reproduces the same digital samples and WAV.
Human judgment
Use the playback comparison to choose whether the sound suits the musical brief.
Revise and recheck
Tune the character by ear

If the end still sounds too bright, shorten the time it takes the higher tones to halve. Replay at the same pitch and keep the sound that fits.

Handoff for review
A sound you can play and share

Hear both versions. Adjust the fade, save the settings so you can reopen the note, or export a WAV another musician can use.

Try the changing note
Technical details and limits

After the peak, multiply the higher tones’ amplitudes by 2−Δt / H.

Δt is time since the peak and H is the chosen half-life. With the extra fade off, the original and new engines match sample for sample at 8, 24 and 44.1 kHz. Reopening a saved setup reproduces the same WAV.

Within a project

Resume from checked work

  1. Sources and assumptions
  2. Results and checks
  3. Next decision

Separate lines of work bring their evidence back to the same decision. After an interruption, the last checked result and open questions give Aptifold a place to continue.

Between projects

Keep a check worth reusing

  1. Failed check
  2. Verified correction
  3. Reusable test

Save the test with the conditions that make it relevant. A later project that changes a filter can check the uncertainty against known answers again.

The owner sets the outcome, approves changes to scope or permissions, and decides whether to act on the result.

Toward collaborative AGI. The goal is an intelligence that can enter unfamiliar work with people, make what the job needs, and get better at working with them across projects.

Read the delivery planner’s model and checks

The agent represents the network as locations at successive times. Travel and storage are arcs with capacities. A fixed reservation buys capacity before a closure is known. For each possible closure, the solver computes the maximum on-time flow using the routes still available.

In the workshop’s four-crate order at each depot, let F(b, c) be the maximum on-time delivery for booking b under closure case c. Each depot can receive up to four. The search minimizes booking cost subject to minc ∈ C F(b, c) ≥ 8, where C includes the backup service’s own closure. The revised order uses the same test with a cap of two at each depot and a required total of four.

A feasible dispatch shows what can get through. A capacity cut bounds what can get through. When the two agree, they certify the maximum for that case. A separately written crate-routing calculation also checks the delivery totals.

RESILIENCE DESIGNER

The network under two closures

Two depots need their full orders at noon. Choose which transport and storage to reserve before you know which routes will close.

Model map with the harbour bridge and ferry closed. Direct service delivers four crates to each depot by noon.

The original plan loses both orders when the bridge and ferry close together. The least-cost plan meets both orders in every case for 17 budget units.

Open the planning tool

Illustrative network. The map shows one of the 29 cases.

How Aptifold verifies results

Aptifold picks a check based on what could make the result wrong. A reviewable result shows the criterion, the inputs and assumptions, the expected and observed outcomes, and the conditions under which the conclusion holds.

Reproduce a reference
Rework a published result using its data and conventions. In the GW150914 reproduction, five of seven checks matched at first. Two peak times differed until the analysis used the notebook’s window convention. Inspect the two runs.
Check against an independent calculation
Use a second route to the answer. The delivery planner checks a feasible dispatch against a capacity bound. A separately written crate-routing calculation checks the totals using the same network assumptions. See the model and checks.
Use known answers and controls
Hold the inputs fixed and change the assumption under test. The ringdown comparison uses the same simulated signals and noise draws in each pair. White noise provides a control with a smaller effect. See the experiment and its checks.
Check that the result serves the brief
Try a case that could pass the easy test and fail the actual job. A delivery total can pass while one depot gets nothing. Check the end result against the owner’s requirements before they decide to act. Replay the counterexample.

When the data leave two explanations open

Aptifold used early readings of WASP-126 to fit two proposed dimming cycles. It fixed the fitted values before checking the later readings.

Early readings used to build the predictions
390
Later readings kept back to check them
603
The workshop shows the prediction for each reading. The runnable record includes the original readings and an independent fit that reproduces both comparisons.

The two predictions agreed with each other, so these brightness readings could support either explanation. To choose between them, you would need a new measurement that the two explanations predict differently.

Can you trust a 95% estimate?

A black-hole merger leaves a pattern of fading vibrations, a little like a bell after it’s struck. Researchers filter noisy recordings to measure those vibrations. A well-calibrated 95% estimate should include the true values about 95 times out of 100, when the same test is repeated with fresh noise.

Aptifold made simulated recordings with known answers. In one noise setting, reusing the old noise model captured the true values in 84.5% of repeats. Updating the model for the filtered noise raised that to 94.8%. With plain white noise, both methods came close to 95%.

See the chart, calculation and experimental controls

GRAVITATIONAL-WAVE SIMULATION

How often did a 95% region contain the known answer?

95% target

White noise with a narrow spectral line

Reuse the old noise model84.5%
Recalculate the noise model94.8%

White noise alone

Reuse the old noise model95.6%
Recalculate the noise model94.5%

Each method quotes a 95% region. We checked whether it held the true values in 2,048 simulations per noise setting. Whiskers show 95% Wilson sampling intervals.

Read the counts and intervals
White noise with a narrow spectral line
Reuse the old noise model. 1,730 of 2,048 repeats. Interval 82.8% to 86.0%.
Recalculate the noise model. 1,941 of 2,048 repeats. Interval 93.7% to 95.7%.
White noise alone
Reuse the old noise model. 1,957 of 2,048 repeats. Interval 94.6% to 96.4%.
Recalculate the noise model. 1,936 of 2,048 repeats. Interval 93.5% to 95.4%.

Reproduce a known result

The agent first used the public GW150914 notebook as a reference. Its implementation matched five of seven quantities, including the delay between the two detectors, while missing both absolute peak times. It traced the discrepancy to how the analysis window was centered. With the notebook’s convention, all seven checks passed. Inspect the two recorded runs.

Make the next question testable

The agent built a two-mode ringdown simulation with fixed, known signal amplitudes. It applied a finite-window filter and reduced the sampling rate from 4,096 to 1,024 Hz. For each noise setting it ran both methods on the same 2,048 noise draws and counted how often the reported 95% region contained all four true amplitude coefficients.

For a linear processing operation H, the original noise covariance Σ becomes HΣH⊤. Covariance captures how noise varies together across samples. The agent compared carrying this transformation into the inference with reusing the old covariance while applying the same processing to the signal in both methods.

Check what the result depends on

In the simulated noise with a narrow spectral line, reusing the old covariance put the known values inside the region in 84.5% of repeats. Updating it raised coverage to 94.8%. With white noise alone, both methods landed near the target. A check using white noise alone would have missed the problem.

These coverage results apply to simulated signals with a known start time and specified Gaussian noise. The separate GW150914 reproduction uses recorded detector data. The research page lets you inspect the models, vary the noise and sampling rate, and see which comparisons the conclusions rely on.

Open the ringdown experiment

How Aptifold learns from work

Aptifold keeps the evidence behind a correction and a test that can catch the same problem again. The lesson records when it applies, so a later investigation can use it when the conditions fit.

  1. Catch the miss

    After filtering, the old noise model missed known answers too often in one test.

  2. Keep a test to rerun

    Save the signal, the filter, the two noise models and the check against known answers.

  3. Try the lesson in different conditions

    Change the kind of noise and run the same comparison. See when accounting for filtering changes the answer.

Compare the results in both noise conditions
How often the 95% estimate included the true values
Simulated noiseOriginal noise modelUpdated for filtering
With narrow interference84.5%94.8%
White noise alone95.6%94.5%

The comparison used 2,048 paired trials per noise setting. Explore both settings in the research engine.

Updating the noise model made a large difference with narrow interference. Both methods came close to 95% with white noise alone.

The agent built the engines and improved its working methods through the projects on this site. Read how Aptifold carries a lesson into later work.

Paul Wishall founded Aptifold in 2026.

What would you like to investigate or build?

A few sentences are enough to start. Tell us your question, what you already know, and what you’d like to be able to do with the answer.

“Our two models give different estimates. Can Aptifold help us find out why and design a test to decide which one to use?”

Start an email
aptifold@proton.me

Open your email app or copy the address into webmail.

Before work starts, we’ll agree on what you’ll get, how it will be checked and what it will cost.

Explore project possibilities