CyberNativeOpen Research Lab Browser replay

Interactive lessonRetained evidence only

How much can one random seed hide?

Select retained runs and watch the summary change. This lesson replays fixed trial-a results in your browser. It does not train a model, execute research code, or earn a reproduction label.

← Back to Labs

Learning objectivesAbout 20 minutes

Read variation.
Limit the claim.

By the end, you should be able to explain why one successful seed is weak evidence, summarize a selected set of runs, and distinguish replaying retained numbers from independently reproducing an experiment.

  1. 01Select evidence

    Compare a single retained run with a larger retained sample.

  2. 02Read the spread

    Track the live mean, range, and number of selected seeds.

  3. 03Check a claim

    Choose the statement that stays inside this evidence boundary.

Evidence workbenchTrial A · 12 retained seeds

Change the sample.
Watch the story move.

The vertical scale starts at 90% so the small differences are legible. Every bar carries its exact rounded accuracy; color is never the only selection cue.

The interactive enhancement is unavailable. The complete retained set and its static summary remain readable below; selection and answer checking are disabled.

All retained trial-a runs selected
Selected runs
12of 12
Mean accuracy
93.06%selected runs
Accuracy range
91.67–95.00%min–max
Select retained trial-a runs. Each item is labeled with test accuracy and seed.

Test accuracy · truncated vertical scale: 90–95%

Read the retained values as a table
Exact retained trial-a inputs used by this page
SeedTest accuracy
110.933333333333
290.916666666667
470.933333333333
710.95
1010.933333333333
1310.916666666667
1670.933333333333
1990.933333333333
2390.933333333333
2810.916666666667
3370.933333333333
3970.933333333333

Claim checkStay inside the evidence

What can this page support?

Choose the strongest claim that the retained trial-a values actually support. “Strongest” does not mean “most exciting.”

Choose one claim
Reveal the worked solution
The middle claim is supported.

It names the observed set, reports only its measured range, and does not generalize beyond the retained evidence. The first claim is false because this browser lesson performs no training or independent execution. The third predicts unseen data that these 12 stored values cannot establish.

A careful next question is: would the range stay similar under a preregistered run in an independently controlled environment?

LimitationsDo not overclaim

Useful lesson.
Narrow evidence.

  • RetrospectiveThe canary protocol was not preregistered.
  • Single contextOne model, dataset split, metric, environment, and retained trial-a series.
  • Stored valuesChanging the selection recomputes descriptive statistics only; it does not create new observations.
  • No scientific reproductionThis page cannot earn “Results reproduced” or “Results replicated.”

ProvenanceExact retained bundle

Trace the lesson
to its evidence.

Read the research status and source boundary →
Evidence slice
Trial A · 12 fixed random seeds · test accuracy
Bundle SHA-256
71e4a800fee459cd3af667fdc4e20ecac88e956b1d2afd6fc123424f33d74065
Page behavior
Local arithmetic over values embedded in this document; no fetch, storage, upload, or execution service.
Earned label
Executed source bundle The lesson itself earns no reproduction label.