RPCB: Relational Probe Calibration Benchmark (v0.1.0)

RPCB (Relational Probe Calibration Benchmark) v0.1.0 is an open-set suite for scoring memory-grounded answers to relational probes under gold evidence states Clear / Partial / Absent / Conflict. This anonymous deposit contains:- data/rpcb_v0.json (N=195 probes + scenario memories)- code/eval_rpcb.py (rule-based scorer; Python 3.11+ standard library only)- examples/responses/*_rpcb0.jsonl (five-config DeepSeek T=0 dumps)- results/expected/metrics_rpcb0.json (expected aggregate metrics)- PROTOCOL.md, LICENSE, MANIFEST.json Open-set labels (mutually exclusive): CR (correct recall), FC (false confidence), AU (appropriate uncertainty), OR (over-refusal). Aggregate metrics: CRR, FCR, AUR, ORR, and CalScore = (CRR + (100-FCR) + AUR)/3. Verify (no API keys): python code/eval_rpcb.py --all-configs --tag rpcb0Metrics should match results/expected/metrics_rpcb0.json. Licensing: data and example responses under CC BY 4.0; scoring code under Apache-2.0 (see LICENSE).Associated with an anonymous manuscript under double-blind review; a camera-ready citation will be added upon acceptance.