Work

punchmark: a retrospective producer identifier for hosted LLM routes

Year: 2026
Type: software
Abstract: punchmark consumes archived benchmark response text (no logprobs, headers or timing) and returns a producer-identity ruling per (route, window): SAME-PRODUCER, SUBSTITUTED or UNDETERMINED at a caller-declared false-alarm rate, calibrated on same-route within-window resamples and cross-fitted nulls, with seeded-substitution power analysis reporting the minimum substituted fraction the archive could have resolved. It emits a one-line certificate attachable to a published score and an append-only, content-addressed rulings store gated in CI. Every verdict is closed-set and certifies the route label as served. It says nothing about model weights and makes no statement about the quality of any model.
Author: Ji, Yuxiang
Institution: Sunway College
Language: English
Open Access status: green
OpenAlex ID: W7201890741
Cites:
Cited by:
Related to: 10