Contributors
Roles are recorded per instrument, not as a standing title. What follows is evidence rather than credentials — the methodological decisions actually made, and the ones that had to be corrected. Judge it yourself.
Taufik Adi Nugraha
Data science · statistics · machine learning · advanced visualization · software engineering
| Instrument | CRediT roles |
|---|---|
| every instrument | Methodology · Formal analysis · Validation · Software · Data curation · Visualization |
What this contributor does not have. Domain expertise in any of these subjects. That is what a collaborating author brings, and why the interpretation sections are open.
Contact
- Email · taufik@openstudy.id
- ORCID · 0009-0006-0753-4496
- LinkedIn · to be added
If one of the open instruments is your subject, this is the person to write to.
Methodological decisions, with reasons
Each of these is a choice that could have gone the other way, and each is defensible or attackable on its merits. That is the standard a methods credit should be held to.
- nightlights
Re-ran the levels regression on annual composites 2012–2025 rather than accepting the published relationship. Rural R² of 0.825 in 2015 against 0.01 in the benchmark study pointed at the sensor product rather than at electrification — so the case reports the discontinuity instead of the correlation.
- rice
Calibrated the radar detector to the official survey with an interpretable OLS per kabupaten-month, carrying cropping intensity and observation density, rather than a single national scale factor. A scale factor would have made the headline agree and taught nothing; this keeps the residual structure visible — where detection over- and under-counts, and against what.
- fire-haze
Required the ignition model to clear both a per-cell day-of-year climatology and the CEMS Fire Weather Index, at every lead, with the climatology fitted on training folds only. A baseline fitted on the full record has seen the test set, which is a common and quiet way to publish skill that is not there.
- forest-watch
Computed the base rate before believing the headline. Three-quarters of alerted hectares fall inside a mill catchment — but sampling the alertable domain itself gives the share chance alone would produce, which is what makes the number informative or empty.
- provenance
Replaced a similarity threshold with a model relevance judgement after measuring that a question about a policy absent from the corpus scored 0.723 — above several answerable questions. No threshold can separate an adjacent topic from an answerable one, so the gate became a reading task and the threshold was demoted to a cost filter.
- provenance
Benchmarked five embedding models on the live corpus and found the entire multilingual-e5 family inverts — unanswerable questions outscoring answerable ones at every model size. The cause was range, not capacity. Also found that one hosted model silently returns a single vector for a batch, which had truncated an index to 28 of 1,383 rows.
Errors found and published
A page of successes is not evidence of judgement. These are cases where the first answer was wrong and the record says so.
- Published a nowcast that is 41% worse than carrying last year forward, with the 99.5% intercept share stated, instead of dropping the model.
- Replaced a gate tally that read 'one of three passed' with per-gate outcomes, which exposed two failing hard gates that the count had hidden.
- Rewrote a house-rules checker four times after it produced three hundred false findings — a field named 'name' is not personal data, a gate id in an HTML attribute is not visible to a reader, and an outcome recorded as a boolean is still an outcome.