Reproducibility — AÍLITH v0.5.0
The supplied archive is the original complete research release. This guide adds an upload-oriented reading and execution path without changing its frozen evidence.
1. Verify and extract
The outer SHA256SUMS lists every kit file except itself. On a platform with sha256sum, run sha256sum -c SHA256SUMS from the kit directory. The full research ZIP has its own manifest and verifier.
Extract releases/AILITH_Game_Agents_RSI_Research_Release_v0.5.0.zip. All commands below run from its AILITH_v0.5.0 directory. Check the inner archive before changing reference files:
python verify_release.py
Expected result: passed: true, checks: 900, and no errors. This verifies inventory, hashes, and recorded source/protocol/evidence bindings. It does not rerun the experiments or prove mathematical claims.
2. Install
Use Python 3.10 or newer. The required numerical dependencies in pyproject.toml are NumPy >=1.24,<3 and SciPy >=1.10,<2.
python -m venv .venv
On Windows PowerShell:
.\.venv\Scripts\Activate.ps1
On Linux/macOS:
source .venv/bin/activate
Then install:
python -m pip install -e .
The frozen studies recorded Linux, Python 3.12.14, NumPy 2.3.5, and SciPy 1.17.0. The dependency ranges describe supported installation constraints, rather than a guarantee that every numerical result is bitwise identical across BLAS, operating systems, or versions. Source/protocol hashes and solver tolerances are more informative than matching elapsed time.
3. Run the examples
python examples/hereditary_arena.py --output results/my_hereditary_arena.json
python -m ailith develop --request examples/development_job.json --output results/my_development.json
python -m ailith develop --request examples/scenario_job.json --output results/my_scenario.json
Inspect statuses, endpoint residuals, actual width, paid cost components, and separate grid/continuous bounds. The supplied two-task scenario has exact one-interval optimal action 2/sqrt(3), approximately 1.1547005384, including parent preparation. Choosing a branch is permitted only after its task is revealed; public design targets do not replace actual episode readings.
The hereditary arena is a static-designed parent followed by a dynamic child. Its successful execution is a feasibility demonstration and must not be used as the optimal joint-parent result.
4. Run regression tests without replacing frozen validation
python -m unittest discover -s tests -q
Run this in the installed environment. The frozen release records 220 tests passed. A separate fresh packaging-time run also passed 220 tests. Neither run measures engine gameplay, GPU model inference, or real LLM calls.
The original python benchmarks/run_validation.py intentionally rewrites results/validation.json and results/test_log.txt. Use a disposable extracted copy if running it. The original manifest will detect changes to those files.
5. Reproduce study designs into fresh paths
python benchmarks/run_inheritance.py --output-dir results/inheritance_reproduction
python benchmarks/run_scenarios.py --output-dir results/scenarios_reproduction
python benchmarks/run_streaming.py --output-dir results/streaming_reproduction
python benchmarks/run_reserve.py --tag reproduction
Runners preserve prescribed inputs and reject overwriting reference evidence. A rerun at the same seed is a reproducibility check, not a fresh independent experimental sample. Keep unsuccessful solves and open gaps in their original panel.
The inheritance study retains 37 schedules, with 35 closed grid gaps and two feasible open gaps. The scenario study retains 18 closed grid gaps. Five analytic plasticity witnesses are separate records. The n=16 witness should not be relabeled as a general solver benchmark.
6. Interpret numerical bounds
action_upper: action recomputed from actual feasible controls.grid_lower: a conservative numerical dual bound for the declared finite-grid problem.continuous_lower: a separate conservative bound for the corresponding continuous class, when supplied.certified: the numerical grid-gap tolerance was met under the software's checks.feasible: a valid schedule is retained with an open optimality gap.inconclusive: the result is not an admitted executable success.
These are float64 computations with explicit checks and corrections. They are not directed-rounding interval certificates. A small grid gap does not bound continuous discretization error at the same scale. Explicit width caps may also prevent realization of an expandable-width SDP optimum.
7. Integration and document build
The archive includes docs/INTEGRATION.md, docs/STREAMING_INTEGRATION.md, and source adapters for Unity, Unreal, and Godot. The policy HTTP service and development CLI are different interfaces. The host maps a development schedule to its own declared adaptation variables and applies state increments rather than repeatedly adding cumulative state.
ONNX inference is optional and separate from CPU development optimization. Install the selected extra (onnx-cpu or onnx-gpu) only for an actual compatible policy. A local LLM endpoint is also optional. Neither is needed to reproduce the core numerical studies.
The manuscript's editable LaTeX and figures are inside manuscript/. Follow its BUILD.md for LaTeX, packages, and build commands. This Hugging Face edition copies the already verified PDF unchanged; it does not rebuild the manuscript or create new research results.