Zillusion · verification
Engineered so it cannot lie to you.
The worst scraper failures aren’t crashes. A selector wired to the wrong field keeps returning plausible, wrong data — and a model reviewing its own scraper approves it. Zillusion treats verification as an adversarial, isolated step with a computed verdict.
Why “looks right” isn’t validation
Spot checks don’t scale, and self-review doesn’t work: an agent grading its own output has every incentive — and every blind spot — to call it good. Most tools either skip verification entirely or let the same model that built the extraction confirm it. Nobody re-checks the data against the live page.
Isolation: the validator can’t edit what it grades
In Zillusion, a separate read-only validator agent re-runs every scraper in a fresh session. It is physically unable to modify the workflow it is grading — it can only run it and score what comes back. Agent runs are container-isolated, and the builder and the judge never share a session.
The verdict is computed, not argued
PASS or FAIL comes from a scorecard, not from a model saying the result looks good. Four checks must hold:
- Re-runs clean. The workflow executes again, end to end, without errors.
- Reproducible. Running it again produces consistent output.
- Resample matches. A sample of rows is re-fetched from the live source and the values must match what the scraper extracted.
- Field semantics re-proven. Each field’s meaning is re-checked against the live page — the check that catches the “plausible but wrong column” failure.
Only after a gate-computed PASS is a workflow promoted into your robot library for reuse and scheduled monitoring. The scorecard and run artifacts stay visible in the workbench, so you can audit exactly what was checked.
Why this is the load-bearing feature
Everything downstream — warehousing, monitoring, reuse — is only as good as the moment you decided to trust the extraction. Our position: the interface can be copied; a verdict computed by an isolated gate is a property of the architecture. If you take one idea from Zillusion into your own stack, take this one — we wrote up the method as a tool-agnostic guide: how to verify scraped data.
Frequently asked questions
How does Zillusion verify scraped data is correct?
A separate read-only validator agent re-runs every scraper in a fresh, isolated session, and a gate computes PASS or FAIL from a scorecard of four checks: the workflow re-runs clean, the output is reproducible, a resample of rows re-fetched from the live source matches, and each field's meaning is re-proven against the live page.
Why must the validator be isolated from the builder?
A model that reviews its own work approves it. Isolation makes the review adversarial: the validator cannot edit the workflow it grades, so a passing verdict reflects what the scraper actually does, not what the building agent claims about it.
What does the gate catch that a normal test doesn't?
The silent failure mode: a selector wired to the wrong field that keeps returning plausible, wrong data. The resample check compares extracted values against the live source, and the field-semantics check re-proves what each column means against the live page.
What happens when validation fails?
A workflow without a gate-computed PASS is not promoted into your robot library. The scorecard shows which check failed, and the build stage iterates — the validator itself never patches the workflow it graded.