GoldenAnalysis

Measure and report across the Golden Suite. A read-only, cross-cutting analysis / metrics / reporting engine: it consumes any stage's typed artifacts (or a raw DataFrame) and emits a unified, exportable AnalysisReport.

GoldenAnalysis ships the generic frame path plus suite adapters (GoldenMatch / GoldenCheck / GoldenFlow / GoldenPipe), cross-run trend + regression detection, an edge-safe TypeScript port (optional WASM), and an optional Rust accelerator for the heavy aggregation primitives — all documented below. See docs/superpowers/specs/2026-06-08-goldenanalysis-cross-cutting-analysis-engine-design.md for the design rationale.

Install

pip install goldenanalysis

Zero suite dependencies for the generic path — it works on any polars DataFrame even with no other Golden package installed.

Quickstart

import polars as pl
import goldenanalysis as ga

df = pl.read_parquet("customers.parquet")

report = ga.analyze(df, analyzers=["frame.summary"])
print(report.to_markdown())

report.to_json("report.json")
report.to_parquet("report.parquet")   # long-form metric frame + table sidecars

CLI:

goldenanalysis report customers.parquet --analyzers frame.summary --format markdown
goldenanalysis report report.json --format markdown      # re-render a saved report

trend and regressions operate over a saved run history (see Cross-run below).

Over the suite

With the relevant extra installed (pip install goldenanalysis[match,check,flow,pipe]):

# A GoldenMatch dedupe result -> match.rates + cluster.distribution
report = ga.analyze_match(dedupe_result)

# A whole-pipeline manifest -> every analyzer whose artifacts are present
report = ga.analyze_pipeline(pipe_result)

match.rates emits match.recall_estimate when GoldenMatch ran dedupe_df(..., certify=True) (it attaches an unsupervised RecallEstimate), and match.recall_safe_bound when you pass an audit-calibrated certificate (analyze_match(result, certificate=...)) — the safe bound needs a labelled sample, so it can't be computed automatically. Both degrade silently when absent.

Cross-run — trend + regression detection

Store reports over time, then trend a metric or detect regressions without ground truth:

hist = ga.ReportHistory(backend="jsonl", path=".golden/analysis.jsonl")  # or backend="sqlite"
hist.append(report)                                  # keyed by (dataset, run_id)

hist.trend("cluster.singleton_ratio", "customers")   # -> TrendSeries

policy = ga.RegressionPolicy(default_pct=10.0, per_metric={"match.recall_safe_bound": 2.0})
regs = hist.detect_regressions("customers", baseline="rolling_median", policy=policy)
print(report.to_markdown(regs))                      # callout + Δ-vs-baseline column

The Baseline is a strategy (rolling_median default — immune to one noisy night — plus previous / last_known_good), and RegressionPolicy thresholds are per-metric and respect each metric's direction (a higher_better metric only flags on a drop). CLI:

goldenanalysis trend --metric cluster.singleton_ratio --dataset customers --history .golden/analysis.jsonl
goldenanalysis regressions --dataset customers --history .golden/analysis.jsonl \
  --policy "match.recall_safe_bound=2" --fail-on-regression   # exit 1 on a flagged regression (CI gate)

GoldenCheck vs GoldenAnalysis

They are easy to confuse and are deliberately distinct:

The hard line: GoldenAnalysis depends on other packages' types; never the reverse. It sits beside the pipeline as a reporting step, consuming GoldenCheck / GoldenFlow / GoldenMatch / GoldenPipe / InferMap outputs — it does not replace GoldenCheck's ingest-time profiling, and GoldenCheck does not import GoldenAnalysis.

Native accelerator (optional, goldenanalysis[native])

An optional Rust accelerator for the heavy aggregation primitives, gated exactly like goldenmatch[native] / goldencheck[native]:

pip install goldenanalysis[native]   # pulls the separate goldenanalysis-native wheel

The pure-Python path stays the default and the byte-identical reference. The compiled kernel (analysis-core pyo3-free + analysis-native abi3 wheel) mirrors core/aggregate.py's histogram / quantile value-for-value, reading input as a Float64 Arrow array (zero-copy). The loader gate (core/_native_loader.py, GOLDENANALYSIS_NATIVE=auto|0|1) uses a primitive only once it's in _GATED_ON — which holds histogram and quantile: both proved byte-identical (tests/core/test_native_parity.py) and measured 5.8–9.9x faster than the pure Python loop on Linux x86_64 at 1M–10M rows, including the list→Arrow conversion the dispatch pays (benchmarks/aggregate_benchmark.py + bench-analysis-native.yml). A new primitive joins only after the same two gates clear — "it's Rust" is never enough (the goldencheck composite-key kernel was 2.5x slower until the gate caught it). With goldenanalysis[native] installed, the auto default uses the native path automatically; GOLDENANALYSIS_NATIVE=0 forces pure. In-tree dev build: uv run python scripts/build_analysis_native.py.

License

MIT.