Teaching module: a reconstruction you can check

A 60–90 minute graduate lab session. Students reconstruct one hologram, then check that someone else's result is the same run. No accounts, no installs beyond Python.

Learning objectives

By the end, students can:

Name the four settings a hologram reconstruction needs and their units.
Run a reconstruction from a browser and from Python.
Read a quality report and say what a fail band does and does not mean.
Rerun a reconstruction and decide, from its IDs and its arrays, what two results do and do not have in common.
State what a provenance record does not establish.

Materials

The sample hologram: https://api.apertureapi.com/static/samples/inline-sample-512.png
The reproducibility bundle: https://api.apertureapi.com/static/samples/reproducibility-bundle.zip (the sample, its settings, run.py and a reference output from the live service)
Python 3 with requests and numpy

Lesson plan

Time	Activity
0–10 min	Introduction: what a hologram records; why "how was this made?" matters in a paper
10–25 min	Browser run: upload the sample at https://api.apertureapi.com/ with wavelength `0.635` µm, pixel pitch `1.85` µm, mode `inline`, focus range `0.5`–`3.0` mm
25–45 min	Python run: `run.py` from the bundle; open `reconstruction_data.npz` and plot the phase
45–70 min	Provenance check: compare your `provenance.json` and `evidence.json` against the bundle's reference output (see `RECOMPUTE.md`)
70–90 min	Worksheet and discussion

The public path allows 5 runs per IP address per hour. A class sharing one network address should run the browser step in pairs, or let the instructor run it once and share the ZIP.

Worksheet

Which setting is in micrometres, and what goes wrong if you enter 635 instead of 0.635?
Where did autofocus land (pdhm_metadata.json)? How close is that to the 1.5 mm the sample was made at?
What is overall in quality_report.json? Which metric set it?
Is your provenance_id the same as the reference's? What three things does it commit to?
Is your evidence_report_id the same? Compare build_id too. Then load reconstruction_data.npz from both ZIPs and compare the phase arrays with numpy.array_equal. What does a matching provenance ID tell you, and what does it not?
Find the section of evidence.json that lists what is not-yet-established. Name two items.
Write the methods sentence you would put in a paper for this run.

Instructor notes

Matching IDs show the same input, settings and software were used. They say nothing about whether the result is correct.
Expect provenance_id to match and evidence_report_id and the arrays to differ. On 2026-10-12 two runs of this sample on the same build gave the same provenance ID but different output arrays and quality metrics. The provenance ID commits to the input, settings and software, not to the outputs. This is the lesson's best discussion point: identity of a run is not reproducibility of its result.
A fail band on the in-line sample is expected and is a good discussion point: the service reports image problems instead of hiding them.
The sample is simulated. Its 1.5 mm distance is known by construction, which is why it works for checking autofocus.

What not to conclude

Not that the reconstruction is accurate. Accuracy and repeatability are not established for any physical setup, and the service says so in every report.
Not that matching provenance IDs mean matching results. They show the same input, settings and software were used, nothing more.
Not that results on the simulated sample carry over to a real instrument.

Next step

Instructors who want the session on their own lab's frames can write to system@apertureapi.com. See also Reproducibility bundle and Methods supplement.