Changelog
What's new
Every release, new check, and platform improvement — newest first.
v3.4
HonestyJuly 2026- Re-audited every benchmark claim end-to-end — see the new methodology page at /benchmark for the full write-up, including where SimAPI does NOT help (gradient-boosted trees) and where naive IQR/z-score filtering is competitive.
- Corrected the recall figure from the previous release: with randomized corruption placement and Mann-Kendall + sliding-window drift detection, recall is 71% (up from 55%), not the previously stated 89% — that number didn't hold up under a harder, randomized benchmark and we're not keeping it on the site.
- Exclusion precision holds at 99% — when SimAPI flags a trial, it's genuinely corrupted.
- Added a naive statistical baseline (IQR + z-score) to every benchmark run so the comparison isn't just "vs. no filtering."
- New AI orchestrator: a 5-phase pipeline (dataset profiling → physics checks → pattern recognition → targeted follow-up probes → synthesis) replaces the single-pass AI reviewer.
- New automatic-repair layer (deterministic, reversible): duplicate rows, missing/duplicate IDs, out-of-order timestamps, wrapped angles, short NaN gaps — with a preview before anything is applied. `simapi repair <file> [--apply]` in both CLIs.
- Multi-format ingestion: YAML, TOML, TXT, and Markdown join CSV/JSON/VTK/NumPy/OpenFOAM.
- `simapi doctor [--fix]` and `simapi explain` added to both CLIs — real environment diagnostics and per-issue explanations, not stubs.
- Fixed a real bug where the deployed site's "Full engine" health check queried the wrong path and could never report a connected Python backend.
v3.3
DetectionJuly 2026- Relationship-drift detector: catches sensor drift that breaks a physical ratio (Re/v, Ma/v, P/ρ) even when the raw column looks stationary.
- CUSUM change-point layer for gradual drift in low-noise channels.
- Validation engine expanded past 400 deterministic checks across 21 domains.
v3.2
PlatformJuly 2026- Pre-flight validation: judge a mesh + solver + physics setup before the run and predict which output checks will fail, with a why and a fix for each.
- Geometric mesh upload (STL) with real watertight / open-edge analysis.
- Historical run comparison — diff two runs for resolved / introduced / persisting issues.
v3.1
AIJuly 2026- AI second-pass reviewer is now data-aware — fed per-column statistics, correlations, and sample rows.
- Honest benchmark methodology: multi-seed with error bars and absolute before/after MAPE.
v3.0
LaunchJune 2026- Public validation API, in-browser dashboard, and CLI.
- Python & Node SDKs; Homebrew, curl, and PowerShell installers.
- 287+ deterministic physics checks across 21 simulation domains.