# Validation record

The v0.5.3 source completed **107/107 local MATLAB tests**, with 0 failures and 0 incomplete tests, starting on 2026-09-08 at 06:31:49 UTC. Environment: 25.2.0.3312555 (R2025b) Update 6, macOS Apple silicon. Individual test durations totaled 49.10 seconds. The suite did not depend on historical MAT/FIG data, paper PDFs or author checkpoints.

[Summary](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/matlab-tests.json) · [Per-test CSV](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/matlab-tests.csv) · [81 source hashes](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/matlab-source-hashes.json)

| Group | Passed |
|---|---:|
| Baseline and repository entries | 46 |
| Koopman L4DC 2025 | 8 |
| Mean-field LQG | 8 |
| Off-policy Q-learning TAC 2023 | 7 |
| Infinite-horizon HJB PINN | 5 |
| Safe PINN ICML 2025 | 13 |
| Robust Koopman 2026 | 8 |
| Bias-PI Automatica 2026 | 12 |

One PINN replay test executes two single Adam updates. Safe PINN tests include zero-update fixed-set evaluation and three controlled failures on the second attempted update: nonfinite loss, nonfinite gradient, and overflow in the real Adam second moment. Each confirms that the saved network, nonempty finite moments and one-row history equal the first completed step. The temporary fixtures use the actual loss and Adam functions; production has no fault-injection option.

A separate two-update run matched v0.5.2's network parameters and all other numerical result fields, excluding timing and network object identity. This short run has its own curriculum, distinct from the default 5,000-update experiment. See the [training-state checks](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/safe-pinn-training-state.json). Only the Safe PINN demo and its test changed among the 81 MATLAB files. Complete training and historical paper results were not rerun.

The existing zero-update regression still verifies `boat:Training` / `No heldout improvement.` after preserving `result.mat`, `metrics.json` and `training.csv`. Its original [v0.5.2 record](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/v0.5.2/safe-pinn-save-checks.json), [test record](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/v0.5.2/matlab-tests.json) and [source hashes](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/v0.5.2/matlab-source-hashes.json) are preserved byte for byte, alongside the [v0.5.1](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/v0.5.1/matlab-tests.json) and [v0.5.0](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/v0.5.0/matlab-tests.json) records.

## Numerical reproduction scope

| Method | Recorded result and unresolved scope |
|---|---|
| Baseline | Analytic LQR gain/value comparison and Bellman checks. Figures retain the recorded September 7 experiments. A fresh September 8 quickstart independently returns the same gain; see its [metrics](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/quickstart-baseline.json). |
| Koopman | Normalized pendulum implementation with independent identification/closed-loop checks; not all original tables. |
| Mean-field | 100/4000-path numerical records; finite-sample failures are retained. |
| TAC Q-learning | Matrix Bellman implementation, with an explicitly identified MIMO initialization variant. |
| Infinite-horizon PINN | Reduced LQR/pendulum neural training and held-out checks; quartic-cost mismatch identified and corrected variant distinguished. |
| Safe PINN | Same 64 candidate states: reduced model executes 64, with 19 collisions and 47 budget violations; author checkpoint executes 63, with 1 collision and 9 budget violations. Counts concern executed trajectories and may overlap. See the [two evaluation paths](https://github.com/tanjunkai2001/adp-matlab/blob/main/reproductions/safe_pinn_icml2025/README.md). |
| Robust Koopman | Held-out model-error bound fails on 25.55% of points; the nominal comparator has slightly lower cost. |
| Bias-PI | Local-multistart pendulum/arm variants. The arm stabilizes but costs 8.84% more than local LQR. Original single-trajectory failures and iteration mismatches are retained. |

See the [method cards](https://github.com/tanjunkai2001/adp-matlab/tree/v0.5.3/docs/fulltext) for source versions, equations, configurations and each result’s scope. Historical records are described in [ARTIFACTS](https://github.com/tanjunkai2001/adp-matlab/blob/main/docs/ARTIFACTS.md).

## Publication infrastructure

The [first GitHub Actions run](https://github.com/tanjunkai2001/adp-matlab/actions/runs/34177624955) passed both repository checks and **102/102 MATLAB tests** on Ubuntu 24.04 with MATLAB R2025b Update 6, Control System Toolbox and Deep Learning Toolbox. It tested commit `b5f104d86662cd9c80c7751f43ff46c5ef9e9a2a`. The [cloud run record](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/github-actions-first-run.json) and [per-test CSV](https://github.com/tanjunkai2001/adp-matlab/blob/v0.5.3/docs/validation/github-actions-tests.csv) preserve that result; subsequent runs appear in [Actions](https://github.com/tanjunkai2001/adp-matlab/actions/workflows/checks.yml). Full paper experiments on another machine remain a separate check.

The project page and README use the same registry-driven method table. Local document/build checks establish their consistency; they do not prove scientific completeness, original-paper equivalence or hardware safety.