Skip to content

Portfolio demo · deterministic CI gate

What happens when a JOIN doubles revenue?

The build stays green. Tests keep passing. But net_revenue silently reports +66.7% because a new customer_tags join fanned orders out (N:1 → N:N). MetricGuard blocks the PR with row-level evidence.

Method →

MetricGuard CI · PASSED

All contracted checks passed

base 32ce1e0 → head demo/baseline-healthy

Deterministic engine v0.1.0 · dbt + DuckDB comparison · no LLM required for this verdict.

net_revenue

mart_revenue · grain: order_id

ok

base

$1,200.00

head

$1,200.00

absolute delta
0.00
relative delta
0.00%
tolerances
±100 · ±2%

Within contracted tolerances.

Findings

No findings — grain unique, metric within tolerances.

Blast radius (lineage from dbt manifest)

Downstream assets reachable from changed models, derived by graph traversal of the compiled manifest.

none

Suggested next check

No blocking findings. Next check: extend contracts/metrics.yml to cover another mart.

Demo runs on synthetic data generated by the Python engine against the dbt/DuckDB fixture project (12 orders · 6 customers). Numbers come from committed engine snapshots, not hand-written values. Reproduce locally:uv run metricguard --base <ref>. Limitations documented in docs/METRICGUARD_METHOD.md.