CreditMap: Provenance Ledgers for Attribution in Human--AI Scientific Collaboration

Aayam Bansal, Ishaan Gangwani

Track: Track 1: Original Research/Position/Education/Attention Track

TL;DR: A machine-readable provenance ledger extending W3C PROV and CRediT with AI-specific roles, hash chaining, and a Python toolkit; evaluated via an audit-task benchmark (94% vs 0% for binary disclosure) and an exploratory reviewer perception study.

Abstract: We introduce CreditMap, a provenance ledger system for attributing credit in human–AI collaborative research. CreditMap extends the W3C PROV ontology with nine AI-specific contribution roles compatible with CRediT, a JSON-LD schema with append-only hash chaining for tamper detection, and a Python toolkit for automated provenance capture via LLM API interception. We evaluate CreditMap through three studies. Study 1 (structural expressiveness): across 45 instrumented sessions, CreditMap captured 6.7 unique roles per session vs. 4.7 for CRediT, with 3.0 attribution distinctions lost per session in CRediT projection. Study 2 (audit-task benchmark): on 250 ground-truth provenance queries, full CreditMap ledgers enabled 94% accuracy compared to 72% (role+timeline), 68% (role-only), 18% (CRediT), and 0% (binary disclosure), demonstrating graduated value of each schema component. Study 3 (reviewer perceptions): in an LLM-simulated reviewer experiment ($N=192$, four frontier models, linear mixed-effects models), CreditMap significantly improved perceived attribution fairness ($d=3.31$, $p<0.0001$) and trust ($d=0.55$, $p<0.0001$); effects on rigor, reproducibility, and overall recommendation were significant but smaller ($d=0.24$–$0.43$) and should be treated as exploratory pending human replication. Logging overhead is <0.1ms per event (<0.02% of typical API latency).

Keywords: provenance, attribution, human-AI collaboration, CRediT taxonomy, scientific authorship, AI governance, transparency, reproducibility, PROV ontology, credit assignment

Submission Number: 140