The compiler that retires ETL estates — and proves it.

Deterministic by construction

The same export always produces the same SQL. A source adapter parses the tool-specific export once, at the boundary, into typed canonical facts — each carrying evidence, trust, and confidence. A planner freezes every decision into a typed MappingPlan. Fail-closed emitters render it: they never invent an identifier, a table name, or a strategy. A missing decision is a build error — never plausible-but-wrong SQL. Anything that cannot be safely automated becomes a typed, counted ACTION REQUIRED marker, never a silent drop.

$ python main.py --source-tool powerCenter --target teradata
    --profile teradata-production --folder estate/ --output out/
    --with-validation --with-workflows --scheduler airflow --explain-facts full
queued
running · parsing source export…
planning… lowering… emitting…
validating… writing explain traces…
succeeded · readiness: READY▌

• Preview command — the exact python main.py argv, reproducible outside the app.
• Queued execution — live-tailing Log, elapsed time, Cancel.
• Profiles — discovery · teradata-production · snowflake-cloud · dbt-modern · strict.

Readiness, not estimates

The verdict is a file, not a feeling.

Every run ends in a machine-readable readiness report: READY, REVIEW_REQUIRED, or BLOCKED. Every mapping gets a confidence score built on 36 deterministic inputs, classified HIGH / MEDIUM / LOW / BLOCKED. A strict profile turns thresholds — confidence floors, fallback-rate ceilings, blocker categories — into a CI pass/fail. Compare any run to a baseline and see exactly which mappings moved, band by band.

READY is a status, not a status meeting.

Evidence on every fact

A confidence score you cannot audit is a vibe. Every mapping opens into Evidence, Proofs, Translation, Trace, and Runtime: classified signals with explainers linking to the exact explain.json line; a proof bundle across Semantic, Ordering, Replay-safety, and Adapter facets; equivalence and write-contract verdicts; side-by-side diffs of source expression versus translated SQL; an ordered decision trace — rule id, decision, confidence — for every choice the compiler made.

{
  "evidence": {
    "source_product": "powerCenter",
    "source_object_ref": "operation:RNK_BY_REGION",
    "source_property": "Top/Bottom",
    "extraction_method": "parsed"
  },
  "trust": "declared",
  "confidence": "exact"
}

-- infa2td generated
-- Source : m_RANK_TEST.xml
-- Folder : estate
INSERT INTO DW.FACT_ORDERS_RANKED ( … )
SELECT …
FROM STG.STG_ORDERS_RANK
QUALIFY ROW_NUMBER() OVER (PARTITION BY REGION
  ORDER BY AMOUNT DESC NULLS LAST) <= 10;

{
  "status": "READY",
  "confidence_score": …,
  "counters": {
    "diagnostic_count": …,
    "blocker_count": …,
    "warning_count": …,
    "manual_action_count": …
  }
}

The decision basis for cutting over a mapping without parallel-running it.

Gates, not goals

# every number below is pinned by a CI gate or a checked-in manifest

sources: PARITY_COVERAGE.md · function_manifest.json · Makefile · pyproject.toml · CONFIDENCE_SCORING.md · MATURITY_MATRIX.md · GENUINE_CEILINGS.md

Invariants are build failures, not review requests.

Ten lenses over every run

1. WORKSPACE upload the source export
2. CONFIGURE Profile · Target · Scheduler
3. RUN queued, live-tailed, cancellable
4. TRIAGE readiness, evidence, diffs
5. SHIP download the .zip

One workbench puts the inputs, the generated code, and every analysis lens side by side.

• Readiness KPIs, band donut, failure Pareto with drill-down.
• Validation row-count reconciliation matrix per table.
• World View the whole estate as one graph, ringed by band.
• Parameters unresolved setup burden, ranked.
• Code view Monaco over every artifact, jump-to-line.
• Scheduler Map workflow task DAGs, linked to the script at line.
• OpenLineage column-level lineage: field ← source.field.
• Graph visualiser the engine's .dot graphs, rendered in-browser.
• JSON Model explain.json explored as a node graph.
• Explain Map how each mapping decomposed into execution units.

Workspaces shared as viewer or editor · Teams · manual actions tracked: Open / In progress / Resolved / Won't fix · Download .zip — the deployable deliverable.