CONSOLE.

LIVE BRIGHT DATA RUN · NOT A REPLAY

EVIDENCE COVERAGE

100%

every published claim verified

FALSE REPORTS BLOCKED

incl. one overfit repair

GENUINE RUNS

4 published · 4 blocked

SELF-HEAL ROUNDS

2 human-approved in production

COLLECTOR

ROBUST

c_mt4m8fix1gze0scg44

SENSOR

ONLINE

live runs enabled

01 · CREATE — CUSTOM COLLECTOR

bdata scraper create https://magpie-lab.netlify.app/lab/source

Bright Data AI generated the output schema and Browser Worker code, including the Load-more interaction. One stable collector ID across the entire story.

02 · AUTHORITATIVE RUNS — COLLECTION API

POST /dca/trigger → GET /dca/dataset → GET /dca/log

Every published number comes from a genuine j_* snapshot: baseline 10 rows, broken run 7 rows — same collector, redesigned page.

03 · HEAL ROUND 1 — REJECTED BY HUMAN

bdata scraper heal … → awaiting_approval → approve --reject

Verbose prompt, weak preview sample — no proof of carousel extraction. Human rejected; production untouched.

04 · HEAL V3 — APPROVED, VERIFIED, THEN CAUGHT

bdata scraper heal … → approve --auto-save → done

Preview contained the missing carousel vendors. Human approved. Verification restored 10 rows with an identical facts hash — and a later probe caught the same repair failing on the original layout. The trust gate blocked it.

05 · HEAL V4 — ROBUST ACROSS LAYOUTS

bdata scraper heal … "Handle both layouts…" → approve --auto-save

Human approved. Verified on three genuine runs: both layouts 10 rows hash-identical; changed facts correctly classified TRUSTED_SOURCE_CHANGE.

06 · LIVE MODE — FROM THE DASHBOARD

POST /api/runs → poll /api/runs/:id → deterministic verdict

Live collections can be triggered from the product; the trust engine computes the verdict the moment data lands.