Watch data enter SkyMind — and watch bad data get refused.
This is a replay of a real ingestion: Moldova's vital-statistics rates, loaded on 8 July 2026. Every number on this page is verbatim from that run — the mixed-in region aggregates that had to be excluded, the two indicators that were rejected, the five validation checks, the volatility flag, and the 3,064 facts that passed the gate and are live on the map right now.
Real load, replayedOne actual ingestion, stage by stage — not a marketing animation.
Any FAIL blocks the loadThe loader exits before a single row is written. An honest gap beats a filled-in lie.
Try to break itInject a synthetic value and watch the gate lock.
What you are watching — and what you are not
This page replays one documented load (NBS Moldova, StatBank table POP020300reg.px, 2003–2024) with its actual values and its actual battery output. The same checks run in code, before any insert, for every source we load — the battery is production behaviour, not choreography for this page.
The "inject synthetic data" mode adds a fabricated value to the replay in your browser only — nothing touches the database. The 135.5% figure it injects is not hypothetical: it is the exact pattern of fabricated legacy data we found and removed from our own Gulf tables in July 2026, publicly, with corrections on the affected posts. That cleanup is why this battery exists.
The protocol behind the replay
Nine documented stages and one gate. The replay below is a single real run through this shape.
Numbers match the stages in the written ingestion protocol. Stage 05 is the battery: it runs on staged rows, before a single row is written, and any FAIL ends the run. The protocol lists nine acceptance tests and not all of them apply to every source — the replay below runs the five that applied to this one. The revision-log write is automatic — database triggers capture every insert, update and delete. Stages 07–09 (product wiring, live-DB recounts, run report) are folded into one box here.
The rules the pipeline enforces
Every stage above is an enforcement point for a written protocol. The short version:
No fabrication, everNational values are never distributed across districts. History is never generated around an anchor. Gaps stay gaps.
Verbatim valuesNative source precision, no rounding anywhere in the write path — rounding manufactures fake ties.
Official geography onlyUnits are the agency's own statistical grid. Region aggregates never mix with district rows.
Polarity requiredA metric enters the product only if "higher" defensibly means better or worse. Marriage and divorce rates didn't qualify.
Provenance per figureSource, dataset, indicator, URL and load date on every metric — inspect any of it on the provenance page.
Refusal is a featureA failing check stops the load before the first row. The same discipline retired entire score tables in 2026 rather than keep unverifiable numbers.