Changelog

All notable changes to Straud are recorded here.

Why this file matters more than it does for most tools. Straud's checks and thresholds change as its calibration improves, so the same data can produce a different verdict under a different version. Every report records the version that produced it (straud --version). If you are comparing a result to an older one, check here first — the difference may be Straud, not your strategy.

This project follows Semantic Versioning. For Straud, a major version means a check's behaviour changed in a way that can alter a verdict on unchanged data.

---

[Unreleased]

Changed

• Verdict and severity labels renamed for clarity. The same four verdicts and four severities, in the same order, under clearer names:

  was	now	why
  verdict Fatal flaw	verdict Fatal	shorter, and it reads as a verdict rather than as a finding
  verdict Fragile	verdict Conditional	Fragile sounded weaker than Unproven; it is in fact the stronger result, because a signal was found. Conditional says what it is: real, subject to assumptions holding
  severity Fatal	severity Critical	one word was doing two jobs. A Critical finding is what produces a Fatal verdict; naming them differently keeps the two ladders apart

  No check, threshold or exit code changed — a report on unchanged data reaches the same conclusion under the same conditions and returns the same exit code. If you match on verdict text, Fatal flaw → Fatal and Fragile → Conditional; in the Python API, Verdict.FATAL_FLAW → Verdict.FATAL, Verdict.FRAGILE → Verdict.CONDITIONAL, Severity.FATAL → Severity.CRITICAL.

Added

• Second independence axis: reports events, dates and entities, each entity's share of P&L, and whether the two axes disagree. Distinguishes a concentrated sample from an edge that lives in recurring entities.
• Design-time gates: power as a gate (--expected-effect) and counterparty identification (--counterparty).
• --policy strict|standard|lenient to choose how much severity stops a pipeline. Findings are identical under all three; only the exit code moves.
• straud --version, and a version stamp recorded in every report.
• Layer B and C calibration, plus calibration against real market returns from three markets and against 204 published equity anomalies.

Changed

• Checks are classified as direct or circumstantial. A circumstantial check no longer reaches Critical without independent corroboration.
• Untrusted values read from data files (column names, tickers) are neutralised and length-capped before appearing in MCP output, and flagged when an identifier does not look like an identifier.
• File reads under MCP are bounded; parquet is checked by metadata before it is opened.
• The ledger, its return sidecar and the seal registry are created readable only by their owner on macOS and Linux.
• Builds install from a pinned dependency set, and releases publish SHA256SUMS.txt.

Fixed

• coverage_gaps measured gap length by calendar-day adjacency, so on a business-day calendar a long hole was reported as at most five days.
• A corrupt holdout registry or trial-returns sidecar was discarded silently; both now warn, because a vanished seal is indistinguishable from never having sealed one.
• A binomial tail overflowed above roughly 1,000 entities.
• MCP relative paths resolved against the process directory rather than STRAUD_ROOT, so most relative paths were refused.

---

[0.1.0] — unreleased

First release.