Sidi Chang

Research profile

Research summary

Sidi Chang is an early-career researcher whose work centers on how AI systems are evaluated and how evaluation claims should be reported. The Evolve catalog holds three 2026 arXiv preprints: a first-authored forensic audit of LLM-judge calibration reporting in biomedical machine learning (Four Ledgers, Not One Score), and two last-authored protocol papers — Refuse, Decompose, Refresh, a preregistered claim-safe protocol for closed-loop AI evaluation, and ClaimReceipt, a receipt specification for verifying evidence sufficiency and coverage in agent evaluations. Together they form one coherent methodological program: decompose evaluation claims, make them recomputable, and be explicit about what a single number does and does not measure. OpenAlex records an h-index of 1 and seven works in the recent three-year window, all attributed to 2026, indicating a concentrated recent burst; one earlier record dates to 2014. Collaboration is thin — one distinct co-author, with Peiying Zhu the recurring collaborator — and the collaboration dimension scores 14. Reproducibility is the strongest axis (78, averaged over two scored papers), although per-paper readiness in the catalog ranges from high to low. Topic assignments are broad, spanning AI ethics and social impacts, software testing, system reliability, cloud computing, and speech, while the catalog's fields split evenly between AI evaluation methodology and NLP/ML evaluation methodology. This remains an early-stage profile built on sparse bibliometric data.

Research keywords

Highlights

• First-authored audit of what LLM-judge review ratios actually measure

  In Four Ledgers, Not One Score, Chang (first author) shows that planted-card review ratios of 0.710/0.218 represent yield and composition rather than judge precision or recall, and proposes four-ledger reporting for LLM-judge calibration in biomedical ML. Evolve calls it a rigorous single-workflow forensic audit, with prevalence unknown and medium reproducibility readiness.

• Preregistered claim-safe protocol for closed-loop AI evaluation

  Refuse, Decompose, Refresh introduces a refuse/decompose/refresh protocol designed to keep closed-loop AI evaluation claims safe to state. Evolve rates it a rigorous, honest simulation study with high reproducibility readiness, while noting that validation rests on a single synthetic system.

• ClaimReceipt: making agent-evaluation claims recomputable

  As last author, Chang's ClaimReceipt specifies audit receipts so that evidence sufficiency and experiment coverage in agent evaluations become recomputable and checkable. Evolve notes careful preregistration but low reproducibility readiness, with validation that is self-referential and no code released.

Timeline

1. 2014Earliest record in the OpenAlex year-by-year counts: one work indexed for 2014, twelve years before the current output burst.
2. 2026Seven works are recorded for 2026 — the entire recent three-year total — including the three arXiv preprints reviewed in the Evolve catalog.
3. 2026First-authored Four Ledgers, Not One Score proposes responsible four-ledger communication of LLM-judge calibration in biomedical ML (reproducibility readiness: medium).
4. 2026Last-authored protocol papers appear: Refuse, Decompose, Refresh (reproducibility readiness: high) and ClaimReceipt for agent-evaluation receipts (reproducibility readiness: low).

Papers in the Evolve catalog

• 2026-09-18

  Refuse, Decompose, Refresh: A Claim-Safe Protocol for Closed-Loop AI Evaluation

• 2026-09-14first author

  Four Ledgers, Not One Score: Responsible Communication of LLM-Judge Calibration in Biomedical ML

• 2026-09-02

  ClaimReceipt: Verifying Evidence Sufficiency and Coverage in Agent Evaluations

Frequent co-authors