Five ways in

To find out whether this archive can support a question of their own, and to get the data without rebuilding it.

1. 01Dataset cardCoverage, variables, and the missing-data semantics.
2. 02Limitation clauses L-1 … L-10Read these before designing anything. Two of them rule out whole classes of study.
3. 03FindingsWhat has already been tested, so the work is not repeated.
4. 04Static archive28 digest-addressed payloads, fetchable directly.
5. 05Reproduce the datasetThe commands, and the digest to check the result against.

Before you start · No instrument response exists in this archive, so calibrated flux and physical severity targets are unconstructible. If a question needs keV, this dataset cannot answer it.

To understand what was done and why, without a background in solar physics.

1. 01OverviewWhat the mission measures and what this project built on top of it.
2. 02Research journeyThe investigation in order, including the hypothesis that failed.
3. 03FindingsThe conclusion, in one line, with the comparison that supports it.
4. 04ValidationWhat ROC, precision-recall and calibration curves actually show.
5. 05The majority classifier's cardThe shortest argument on this site for why accuracy is the wrong metric.

Before you start · This is a detection study on one instrument in one phase of the solar cycle. It is not an introduction to solar physics, and nothing here is a space-weather forecast.

To see how the thing is built, and whether the engineering claims hold up.

1. 01ArchitectureSix views: pipeline, data flow, gates, deployment, layout, journey.
2. 02The evidence gateHow a rendered figure is mechanically bound to an artifact.
3. 03Reproducibility metricsIncluding the two that report as partial or unverified.
4. 04The lockfile case studyA digest mismatch, traced instead of silenced.
5. 05Static archiveWhy this is not called an API.

Before you start · There is no backend. No server runtime, no database, no queue — so there are no throughput or latency numbers, and none are invented to fill the gap.

To decide whether the central claim is supported, as quickly as possible.

1. 01ValidationCurves, confusion, calibration and error analysis from 192,541 held-out predictions.
2. 02MethodThe evaluation protocol, frozen before any model was fitted.
3. 03LimitationsWhat the dataset cannot support, stated by the project itself.
4. 04Evidence traceabilityEvery headline claim, traced to artifact bytes and commit.
5. 05Engineering recordSix times the implementation contradicted the specification.

Before you start · Confidence intervals come from a day-block bootstrap over an effective sample of events, not minutes. Minute-level intervals would be far too narrow, and the reason is published.

To see the strongest thing here in five minutes.

1. 01The verdictA negative result, published at full weight.
2. 02The lockfile case studyThe integrity system catching a real problem, and explaining it.
3. 03Evidence traceabilityClaim → figure → metric → artifact → commit, for six claims.
4. 04The contradiction recordIncluding the amendment that was declined and the one still open.

Before you start · The opening scene is illustrative and watermarked as such. No image on this site is Aditya-L1 imagery, because the payloads produce spectra and light curves, not pictures.