naishashetty/autonomous-ai-infrastructure-benchmark

Phase 5.3/5.4 Benchmark — Release Package

An executable benchmark of 16 tasks across 8 tracks (uncertainty, abstention, failure prediction, diagnosis, recovery, memory, generalization, end-to-end), scored over the Phase 5.2 canonical dataset.

What this is: a scoring harness that loads a fixed dataset, validates it, scans for leakage, executes every task that has enough evidence to score, and reports a per-task capability matrix — never a single overall number.

What this is NOT: a claim that the underlying system (the "Autonomous AI Infrastructure" project) reliably predicts failures, recovers from them, or generalizes across environments. Of the 16 tasks, 0 are VALIDATED, 6 are PARTIALLY_VALIDATED (real but limited evidence), 3 are UNDERPOWERED (real point estimates, insufficient sample for a headline claim), and 7 are NOT_EVALUABLE (the current dataset lacks the evidence these tasks require). See BENCHMARK_CARD.md for the full capability matrix and PHASE5_3_LIMITATIONS.md for every disclosed gap.

Contents

src/benchmark/        executable implementation (runner, tasks, metrics,
                       baselines, ablations, leakage checks, validation,
                       reporting, reproducibility metadata)
src/phase5/            failure_mapping.py (diagnosis->failure-class map)
experiments/results/   frozen specification + the exact dataset this
                       benchmark is defined against (see below)
scripts/               run_phase5_4_benchmark.py (the runner entry point),
                       generate_sha256_manifest.py
tests/                 test_phase54_benchmark.py (unit tests)

The experiments/results/... subtree preserves the exact relative path layout the implementation expects (src/benchmark/constants.py resolves paths relative to this package's own root) — do not flatten or rename these directories.

Quick start

pip install -r requirements.txt
python scripts/run_phase5_4_benchmark.py --out-dir /tmp/my_run

This loads the bundled dataset, validates it (fail-closed on any violation), scans for leakage, executes all 16 tasks (scoring where supported, NOT_EVALUABLE where not), runs the 5 ablations, builds the capability matrix, runs the whole thing twice to verify determinism, and writes PHASE5_4_BENCHMARK_RESULTS.json, BENCHMARK_RUN_MANIFEST.json, PHASE5_4_BENCHMARK_REPORT.md, PHASE5_4_VALIDATION_REPORT.md, PHASE5_4_LIMITATIONS.md, PHASE5_4_SYNTHESIS.md, and a SHA-256 manifest to --out-dir.

Reproducing the uncertainty-track results

The uncertainty track (UNC-ARITH, UNC-SENT, UNC-QA) is the only track with enough per-episode evidence to compute full ranking metrics (AUROC/AUPRC/Brier/ECE) end-to-end from this package alone — see BENCHMARK_CARD.md. Running the quick-start command reproduces these exactly; see REPRODUCIBILITY_GUIDE.md for the full protocol and expected values.

Tests

python -m pytest tests/ -q

Documentation

• BENCHMARK_CARD.md — what each task measures, evidence, sample size, status, limitation, one line each.
• REPRODUCIBILITY_GUIDE.md — full protocol, environment, seeds, expected outputs.
• CITATION.cff — how to cite this benchmark.
• experiments/results/phase5_benchmark_specification/20260826T055915Z/ — the frozen specification this implementation follows exactly (PHASE5_3_BENCHMARK_SPECIFICATION.md, PHASE5_3_TASK_CATALOG.json, PHASE5_3_METRIC_CATALOG.json, PHASE5_3_LIMITATIONS.md, PHASE5_3_LEAKAGE_POLICY.md).

License / provenance

License: MIT for this benchmark implementation (runner, task registry, metrics, baselines, ablations, validation/leakage/reproducibility tooling) — see LICENSE in this directory. Decided by the project owner on 2026-08-27 after an upstream license/provenance audit (LICENSE_PROVENANCE_AUDIT.md, Phase 5.6 release directory) found no restriction on redistribution: all code here is self-authored. The bundled dataset copy under this package's experiments/ directory is dataset content, not software, and is licensed separately under CC BY 4.0, not MIT — see DATASET_CARD.md's equivalent disclosures under the paired dataset package. The dataset itself is derived exclusively from this project's own Gen-3 evidence — no third-party dataset content is included.