Skip to content
AptifoldGet in touch

How Aptifold investigates, builds and learns

Give Aptifold a difficult question. It can work out what evidence to look for, build the model or tool the investigation calls for, and run checks that show what you can conclude.

From a question to a working result

Aptifold’s agentic operating system organizes the work around AI models. It keeps track of what the project is trying to accomplish, chooses what to investigate next, and uses tools to collect evidence, run calculations and build software. It can write a mathematical model or a program when the work calls for one.

  1. Frame the problemRecord the goal, constraints and what would change the decision.
  2. Build a way to test itFind sources and make the model, simulation or analysis.
  3. Challenge the resultRun the test, examine failures and revise what the evidence calls into question.
  4. Hand over the workDeliver the result with inputs, assumptions and checks someone can inspect.
Evidence can send the work back to the model or even to the original question. The project owner decides which results to act on.

You can start with a goal in ordinary language and let Aptifold carry the investigation. You can also work alongside it. A scientist can propose a competing explanation, an engineer can tighten a design limit and a business team can change what counts as an acceptable outcome. Aptifold can find which results depend on that change and redo the relevant work.

In a delivery-planning investigation, Aptifold modeled which routes and storage slots would be available at different times. It searched for bookings that could serve both depots, checked them against pairs of closures and used a separate crate-routing calculation to check the answer. The planner lets a visitor change the booking and inspect where the crates go.

Follow the delivery decision and open the planner.

Choose a test that could change the answer

Suppose a star dims every three days. One explanation has one dip in a three-day cycle. Another has two identical, evenly spaced dips in a six-day cycle. Both predict the same dips. A useful next experiment would measure something the two explanations predict differently.

Two cycles, the same predicted dips
CycleDay 0Day 3Day 6Day 9
3 daysDipDipDipDip
6 daysDip 1Dip 2Dip 1Dip 2
Illustrative timing. With the same dip shape and noise model, these hypotheses make the same brightness predictions at every time in the model.

Shadow Observatory makes this problem inspectable with 993 archived brightness readings of WASP-126. By default, the workshop uses 390 early readings to fit each candidate and 603 later readings to check the fixed predictions. Its export includes the observations and both predictions, so you can examine where they agree and where either misses a measurement.

Open the timing workshop and compare the predictions.

The reference clock had a later RMS error of 0.179% of normalized brightness. The nearby 3.31-day clock scored 0.221%. Two equal dips per 6.57766-day cycle gave the same predicted light as the reference.

Download the timing record and use node aptifold-timing-verification.mjs to repeat both fits and check the later scores.

How to put a value on another measurement

When the candidate models, measurement noise and downstream decision can be specified, Aptifold can build a numerical model for choosing a test. Expected information gain measures how much a possible observation would change beliefs about the candidate explanations. A decision model can instead score the measurement by how much it improves the action you would take.

Net value = current risk − expected risk after the measurement − measurement cost

Here risk is the expected loss of the best available action given the evidence. The second term averages over the possible readings, each followed by an updated decision. The calculation depends on the proposed models, likelihoods, losses and costs. It earns its place when those inputs make the comparison meaningful. Modern Bayesian Experimental Design.

Check what the result supports

In the gravitational-wave study, Aptifold reproduced a public analysis of GW150914. Five of seven checks matched on the first run. Two absolute peak times differed. It traced the discrepancy to the choice of analysis window and repeated the calculation with the notebook’s convention. All seven checks then matched.

That gave the investigation a checked starting point for a new question. After filtering a noisy signal and keeping fewer samples, how reliable is a quoted 95% uncertainty region? Aptifold made simulated recordings with known signal strengths so it could count how often each analysis included the known values.

How often the 95% region included all four true coefficients
Simulated noiseReuse the old noise modelUpdate for filtering
Narrow interference84.5%94.8%
White noise alone95.6%94.5%
Both methods used the same recordings within each setting. There were 2,048 paired repeats per setting. Updating the noise model made the larger difference when the simulated noise included narrow interference.

The technical difference is where the filter enters the model. If H is the linear processing operator and Σ is the covariance of the original noise, the processed noise has covariance HΣH⊤. This covariance describes how the noise moves together across samples. The comparison isolates the effect of updating it under the stated simulation conditions.

Compare the recorded reproduction runs. Explore the uncertainty results and download the evidence.

A different check answers each question. A reproduction checks whether the same inputs and method lead to the reported result. A controlled simulation checks behavior against a known answer. A reader trying an unfamiliar case checks whether the finished tool helps someone understand or decide. These checks can expose different failures.

What to inspect before relying on a check

Start with the decision and the criterion. A four-crate total can pass while one depot is empty. The delivery plan needs a test of the minimum at each depot under every declared closure. A counterexample that passes the total makes that distinction visible.

Ask how the result was checked and where the checking methods share assumptions. A feasible dispatch and a matching capacity bound establish the modeled maximum. A separate routing calculation can catch an implementation error. Both calculations can share a mistaken estimate of real-world capacity.

Download the delivery verification record. Run node aptifold-delivery-verification.mjs to independently recompute all three acceptance rules from the embedded inputs.

A usable handoff shows the inputs, criterion, observation or calculation, outcome, and open decisions. Aptifold can run these checks within the authority the owner grants. The owner approves a changed requirement, spending, publication, or use of a consequential result. Follow the decision and replay its counterexample.

Keep a lesson someone can put to the test

Aptifold can revise its working methods as projects reveal errors and better approaches. A useful record connects the lesson to the evidence and the conditions that made it relevant. The system can bring that record forward when a new question calls for it.

Keep the cause and the check
Record the earlier result, the assumption or step that caused the problem, and the source or test that exposed it. A rerunnable check gives the next investigation a concrete starting point.
Record when it applies
A lesson about filtering belongs with the suspected noise relationship and the processing that changes it. The same comparison under white noise shows how the size of the effect can depend on the conditions.
Change the working method
Update an instruction, calculation or test and keep the earlier version available for comparison. Later evidence can reopen a conclusion or replace the lesson.
Try it on a fresh case
Check whether the new method helps where the earlier problem recurs and behaves appropriately on a similar case where the lesson is irrelevant. Observe whether the agent can use it on continuation without a reminder.

These changes take place in the working instructions, project knowledge and software around the underlying AI models. If a good rule fails to affect the result, check whether it was available in context and whether the agent used it. Those observations help decide whether to revise the rule, improve how it is found, or correct its application.

Assess the change on later work. Compare the old and revised methods on cases kept for evaluation, and count the preparation and corrections a person had to supply. Repeated feedback from an evaluation set can influence the changes being tested. The Reusable Holdout.

Bring a result you want to reach

Aptifold can take an open question through research and testing, or work with you as a project develops. The result could be an analysis you can reproduce, an experiment you can run, a decision tool, or an interactive engine. The engines on this site let you inspect and change the inputs to already-built models.

The original articles