Audit Lab

Audit Lab is a deterministic, known-ground-truth correctness fixture. It creates invalid positive results on purpose, runs the corresponding safe controls, and writes a byte-stable evidence set.

microalpha audit-demo

No network, provider, credential, licensed row, or hidden holdout is accessed. The only inputs are schema microalpha.audit-lab.v1, seed 20260715, and NumPy-generated arrays bound by input_sha256.

The fixture uses transparent NumPy oracle constructions so every injected failure has known ground truth. The production Engine and ExecutionPlan path is guarded by test_tplus1_execution.py, while the shared max-statistic implementation is guarded by test_multiple_testing.py.

Four paired audits

1. Point-in-time availability

The unsafe feature is a revised value that contains the contemporaneous return but is not available until two sessions later. Using it at the decision time creates Sharpe +20.5276. The safe path enforces available_at <= decision_at, blocks all 756 unavailable rows, and produces Sharpe −0.1728 from the information actually available.

The same public require_point_in_time gate raises PointInTimeViolation with the exact violating row IDs and count instead of silently dropping late data.

2. Event-time execution

The unsafe oracle observes a return and fills on the same tick, producing Sharpe +21.8860. The safe vector fixture shifts the signal to the next observation; its Sharpe is +0.1739.

The separate production regression test uses a clock-guarded data handler: any attempt to read a future price before its event raises immediately. It also records cash and positions at each strategy callback, proving the future fill cannot mutate state early.

3. Cost reconciliation

A labeled planted control has gross Sharpe +0.5668. Commission, half-spread, impact, and borrow costs reduce it to −0.6753. Net returns reconcile exactly:

net = gross - commission - half_spread - impact - borrow
max absolute residual = 0.0

These costs are fixture parameters, not empirical venue calibration.

4. Model-selection control

Searching 128 noise strategies over the full sample finds an apparent winner at Sharpe +1.3802. A model selected only on the first 504 observations scores −1.2781 on the untouched final 252 observations.

The max-statistic test evaluates every candidate as a return differential to an explicit zero-return benchmark. It recenters differentials under the null and uses the same stationary-bootstrap timestamps for every candidate, preserving cross-model dependence. The noise family is not promoted (p=0.601); a labeled planted positive control is detected (p=0.001).

Canonical artifacts

File	Purpose
audit_results.json	Exact metrics and claim boundary
comparison.csv	Four-row machine-readable comparison
audit_lab.svg	Reviewer-facing result graphic
data_lineage.svg	Architecture and data-lineage graphic
receipt.json	Input and artifact SHA-256 hashes

The receipt itself has SHA-256 6e36c2397696d7e9eecbd058cbfc1ba522c8ffba7e5798224de86b20457b6575. Paths, clocks, hostnames, and Git metadata are excluded from canonical bytes.

Python API

from microalpha.audit_lab import run_audit_lab

result = run_audit_lab("evidence")
assert result["receipt_sha256"] == (
    "6e36c2397696d7e9eecbd058cbfc1ba522c8ffba7e5798224de86b20457b6575"
)

This equality verifies the fixture and generator version used by this release; it is not a universal hash across future schema versions.