timeframe-heuristic-winnower
Selects segment boundaries by agreement across upstream views. Every view in the input MMIF that carries time-anchored annotations is a source of candidate boundary events (a view whose TimeFrames declare different labelsetUri values is one source per value): the endpoints of its Sentence annotations when it has any, otherwise the endpoints of its TimeFrame annotations (word-level frames aligned to Tokens are skipped). Events are widened into a window of windowMs, and a boundary is a point where at least the share minAgreementRatio of the sources agree. Among boundaries closer than minSeparationMs, the best-corroborated set is selected exactly by dynamic programming. Output is one segment TimeFrame per interval between consecutive boundaries. The method needs no model and no GPU. The JCDL 2026 paper (https://doi.org/10.1145/3805696.3846037) used these defaults with minAgreementRatio=0.8 over five sources (whisper-wrapper, pyannotesd-wrapper, pyscenedetect-wrapper, inaspeechsegmenter-wrapper, swt-detection).