📡 RADAR II

Code for the paper "Toward Autonomous Radio Follow-up of Multi-messenger Transients with RADAR: Language-Model Benchmarking, Accelerated Inference, and Agentic Jansky VLA Scheduling" (Hategan, O'Dwyer, Corsi, Huerta, Li, Foster et al.).

RADAR (Radio Afterglow Detection and AI-driven Response) is a federated, privacy-enhancing framework for radio follow-up of gravitational-wave events, introduced in Patel et al. (2025), ApJS 280, 71. This paper extends it in three directions:

1. Accelerated federated inference: network round-trips through the Octopus fabric, not afterglowpy evaluation, dominated the original 3–4 hr fit (~40× slower than local emcee). Server-side thread pools and site-side subprocess pools restore ≈120 samples s⁻¹, matching the local parallel baseline. The JAX surrogate FIESTA is ~5× faster per call but loses to afterglowpy under parallel sampling due to GIL contention.
2. LLM benchmarking of the GCN parser: GPT-5.5, Claude Opus 4.7, and Gemini 3.5 Flash against the human-curated GW170817 radio catalog, using the original prompt set. Best event-match F1 = 0.893 ± 0.002, a 16% gain over GPT-4.1.
3. Agentic VLA scheduling: an LLM agent converts natural-language observing requests into scheduling blocks (SBs) that upload cleanly to the NRAO OPT; a 21-criterion auto-grader scores them against manually written SBs (123/126 criteria passed across six SBs).

Layout

Path	Paper section	Contents
gcn-ai-parser/	§2	extract_gcn.py runs the four-prompt extraction chain over GCN Circulars; evaluate.py computes P/R/F1 and GCN-level recall (mean ± std over 3 runs) for Table 1
sb-auto-grader/	§4	VLA SB grader — rules in rubric.yaml, engine in src/score_sb.py, full pipeline in src/llm_judge.py. See its README
radio-module/	§3	Updated radio module for RADAR adding parallelism to model evaluation.
radio-plots/	§3	Code to generate the plots for the radio section.

# parse + score GCN circulars
cd gcn-ai-parser && python extract_gcn.py --model GPT-5.5 --output result/gpt55_run1.json && python evaluate.py

# grade example SBs, fully offline
cd sb-auto-grader && python src/llm_judge.py examples/*_AItest.txt --ground-auto --no-llm