Kharisara/mer-jit-llm-fgcs: Submission-Aligned Reproducibility and Validation Release
Authors/Creators
Description
ReplayBench-PG v2.1.0 is the submission-aligned reproducibility release for the manuscript:
"ReplayBench-PG: Fault-Detecting Deterministic Action-Trace Replay for Policy-Gated AI Pipelines."
This release freezes the implementation, configurations, compact evidence, validation manifests, and manuscript-traceability artifacts used in the final Array submission.
Main additions
- Independent downstream execution receipts for cross-log authorization–execution validation
- Record-bound, configuration-bound, and combined trace digests
- Controlled tests for unlogged calls, false execution logs, duplicate calls, and correlation mismatches
- Expanded validator-selectivity and fault-sensitivity evaluation
- Finalized 528-execution mixed-repetition timing study
- Dedicated 88-execution
bc_liveruntime-decomposition study - Ray execution-substrate validation
- Canonical MetroPT-3 provenance and cross-domain validation
- Exact local execution-environment manifest
- Frozen local and container dependency locks
- Docker base-image and final-image digests
- Local-versus-cloud environment comparison for Windows and two Cloud Run regions
- Claim-to-artifact traceability for manuscript tables, figures, and headline results
- Deterministic Figure 2 generation manifest
- Fail-fast final artifact validation with Python compilation and automated tests
Validation status
The frozen evidence package passed:
- Comment 12 reproducibility and claim-traceability validation
- Comment 13 local-versus-cloud environment-comparison validation
- Primary benchmark validation
- Finalized timing-study validation
- Ray validation
- Controlled-fault validation
- MetroPT-3 provenance and secondary-workload validation
- Cross-region and local-to-cloud action-trace comparison
- Repository-wide Python compilation
- Complete automated test suite
Frozen technical baseline
Final evidence commit:
e241710c7a16a52cbb9452089ef9960f15bfb002
Scope
ReplayBench-PG validates ordered policy-action reconstruction and selected execution-integrity properties for policy-gated AI pipelines. It does not claim complete program replay, policy optimality, causal effectiveness, content safety, production autoscaling, maliciously authenticated logging, or general distributed-systems reliability.
Data availability
Source datasets are not redistributed. MELD and MetroPT-3 remain available from their original providers. The release includes preparation scripts, configuration files, compact outputs, provenance manifests, checksums, validation summaries, and manuscript-ready artifacts.
Files
Kharisara/mer-jit-llm-fgcs-v2.4.0.zip
Additional details
Related works
- Is supplement to
- Software: https://github.com/Kharisara/mer-jit-llm-fgcs/tree/v2.4.0 (URL)
Software
- Repository URL
- https://github.com/Kharisara/mer-jit-llm-fgcs