| System | Function | Evidence |
|---|---|---|
| TRLCBench | Benchmark for verifier-guided temporal plan repair under genuine duration, concurrency, deadline, and resource constraints | Two PDDL domains, three failure classes, four repair strategies, TRLCScore, OPTIC/VAL validation, and a 64-case evaluation |
| COUNTERFACT-SWE | Predicts which tests a proposed code patch can affect before executing the complete suite | Detected 8/8 observed failures while selecting 134 of 2,652 tests |
| OpsPilot | Evaluates AI agents for operational diagnosis and remediation | Reproducible failure scenarios, agent baselines, execution traces, and outcome-based evaluation |
- vLLM #57528 — Offload streaming derender detokenization
- Moved incremental streaming detokenization off the FastAPI event loop.
- Added executor-boundary regression coverage and concurrent workload validation.
| Priority | System | Intended contribution | Status |
|---|---|---|---|
| 01 | SentinelKWS | Privacy-preserving, real-time keyword spotting with speaker-disjoint evaluation, open-set unknown/noise handling, calibrated abstention, and NeMo → ONNX → TensorRT deployment | Building |
| 02 | Hierarchical Laya | Remove Laya’s large-option ceiling through hierarchical label selection and calibrated probability composition, evaluated on Banking77 | Research prototype planned |
| 03 | KAIROS | SLA-aware inference scheduling across execution backends using latency, throughput, VRAM, energy, and correctness constraints | High-risk systems research |