AptifoldContact
← Return to the observatory

Test the next dip.

Set two predictions with the early readings. Keep those settings and check both against the later dates.

WASP-126 · 993 TESS readings over 27 days

The interactive comparison has not started. If it stays here, reload this page. You can still download the original readings or return to the observatory.

Split the readings

Full timeline

Two predictions use the early readings. Their later scores come from the filled dots on the chart.

A · referenceB · your clock
Change split · day 9

Select a dot to inspect its reading. With the plot focused, use ← and → to move through the source rows. The reading control below works too.

Day 0 is the first source timestamp, BTJD 1325.657762. Gaps mark times without a reading. Brightness 1.000 is the file’s reference level.

CHANGE THE EXPLANATION

What if the clock is slightly wrong?

Try the slower clock first. Then try two equal dips: a different description can make exactly the same light.

A

Supplied reference

One dip every 3.28883 days.
This period comes with the known-planet record.

B

Your proposed clock

Change the assumed dip duration

Both clocks use the same early dates, dip duration and fitting rule.

The duration is an assumption. Each template is averaged over an assumed uniform half-hour exposure.

LOOK AT THE EVIDENCE

At this reading.

A prediction is useful because you can check where it misses.

A close view of the same dates and brightness scale. Select a dot or use the reading controls.

Displayed values are rounded; the CSV retains their full numeric precision. “Background” is the source’s separate sky-background column. It is not an error bar and does not weight the fit.

FOLLOW THE TEST

Learn here. Check there.

The early readings determine each template’s starting time, baseline and dimming. Later brightness never adjusts those fitted values.

Same scoring rule, separate dates
DatesReadingsA · mismatchB · mismatchWhat happened here?

You can inspect the later scores and keep changing your proposal. That makes this an exploratory comparison, not a blind validation study. A lower error here does not establish the physical cause of the dimming.

Exact method, assumptions and sources

A small timing model

The light is modeled as a constant baseline minus identical rectangular dips. Their repeat interval is the cycle length divided by the number of equal dips. This is a simplified brightness template, not an orbit, a realistic transit profile, or a fit for a planet’s radius.

For each stated clock, we try 512 equally spaced starting phases over one repeat interval, relative to the first original timestamp. At each phase we fit the baseline and a nonnegative dimming by ordinary, unweighted least squares on the early readings. Dimming is bounded between 0 and 0.05 in normalized flux. The lowest early sum of squared errors selects the phase. This finite grid can miss a better phase between its points.

For a selected phase, q is the fraction of an assumed uniform half-hour exposure overlapping a dip. The prediction is baseline − drop × q. We obtain drop by clamping −cov(q, brightness) / var(q) to [0, 0.05], then set baseline = mean(brightness) + drop × mean(q). A phase with no variation in q is skipped.

Current fitted values, learned only from the early readings. Display rounded; no uncertainty estimate.
ClockMid-dip reference, BTJDLight outside dipDimming, normalized scalePhase step, 0–511

What the scores mean

Each residual is measured brightness minus predicted brightness. The mismatch is sqrt(sum of residual² / number of readings). Percentages multiply this normalized-flux error by 100. The flat-light comparison uses only the early mean. All later readings have equal weight; gaps contain no fabricated points.

The original three-column file supplies no per-reading flux uncertainty. No significance, confidence interval or planet probability is calculated. Correlated variability, instrumental effects, realistic ingress and egress, changing dip depths and other physical causes are outside this template. The exposure average is an approximation, not a recovered instrument-response function.

When the test cannot choose

One identical dip every P days and two identical, equally spaced dips in a 2P-day cycle have the same brightness repeat. In this declared model family, they predict exactly the same light at every time. More readings of that same light cannot distinguish the two descriptions. Unequal dips or other measurements could change the question.

The original record

MIT TESS Science Office QLP, sector 1, TIC 25155310, WASP-126. All 993 rows retain their original timestamps, normalized SAP_FLUX and normalized SAP_BKG. The row number is an added one-based locator. The source has upstream processing and quality selection; this workshop adds no clipping, detrending, renormalization or synthetic observations.

Download the unchanged three-column source · QLP at MAST · Dataset DOI. Data: Huang et al. (2020a,b), CC BY 4.0. No institutional endorsement is implied.

The QLP methods paper describes the original primary-mission 30-minute image cadence and processing caveats. Astropy’s box least squares guide explains periodic box templates and the role of a finite search grid. This workshop compares stated clocks with its own reduced calculation; it does not run a period-discovery search.

Original-file SHA-256: db6e6e546884e90184ed8d02101b40b567999b7fc91da3bd105684859124586d

Keep the question you tested.

Save the inputs, or take all 993 rows with both predictions and residuals. Opening a setup recalculates the result. Files stay on your device.

APTIFOLD / AGENTIC AI

What would you like to work through?

Aptifold develops its own agentic AI to investigate questions, test ideas and build working software. This experience is one example of that work.

Explore Aptifold Discuss a question

Company links open in a new tab, so you can keep this experiment open.