Executive summary
LineageSutra AI 0.2.0 answers one question before a schema change reaches production: if I change this column, what breaks, and where is the proof?
The problem
Analytics and data-platform teams rename and drop columns every week. Today the "impact analysis" is a grep, a meeting, or a catalog that is out of date. Mistakes surface as broken dashboards, failed jobs and silent wrong numbers.
The product
An offline command-line tool and 25 portable AI-agent skills. It reads SQL files (or a dbt manifest), builds a column-level lineage graph, and returns:
- Confirmed static impacts, each with a proof path and
file:line. - Possible impacts (filters, joins, group-bys, table-level dependents), clearly labelled as not proven.
- Warnings for everything it could not parse. It never guesses.
- A review checklist, regression-test ideas and a sealed Markdown / HTML / JSON report.
Why it is different
| Typical tool | LineageSutra AI |
|---|---|
| Needs a warehouse connection or SaaS account | Reads local files; no network, no credentials, no rows |
| Shows a graph, hides uncertainty | Separates confirmed from possible and lists unparsed SQL |
| Claims accuracy | Publishes a gold corpus, its labels, and its limits |
| Opaque output | Deterministic, SHA-256-sealed, reproducible reports |
Proof, today
- 157 automated tests: 153 pass, 0 fail, 4 skipped (recorded 2026-11-16T18:03:17.735Z on win32/x64, Node v24.15.0).
- Gold corpus: 54/54 cases exact, 148 expected edges, 0 false edges, 0 missed edges. Author-labelled regression corpus, not an independent benchmark.
- Self-audit: 28 checks passed, 0 failed, 7 explicitly not run.
- 2,000-view scan median undefined ms; impact over a 20,000-dataset graph median undefined ms (this machine, synthetic data).
Honest limits
A documented SQL subset, not a certified dialect parser. Static text analysis only. No live-database verification. No independent security review. The SHA-256 seal is tamper evidence, not a signature.
Where it goes next
SQLGlot bridge for dialect depth, OpenLineage import, git-aware schema diffs and signed evidence. See the roadmap.