A pre-vocalic connected speech context is said to enable the resyllabification of word-final consonants into an onset, thus conditioning alternations. We present EPG data on English word-final /l/, measuring the extent of alveolar contact and the rate of vocalisation, the extent of dorsal retraction (representing “darkness”), and the timing of alveolar contact relative to dorsal retraction. Two dialects of British English are considered, namely Scottish Standard English and Southern Standard British English. Results are that /l/ alternation is systematic: the tongue tip contact of word-final /l/, quite categorically for some speakers, is more onset-like in pre-vocalic and more coda-like in pre-consonantal contexts. This alternation is not along the lines predicted by a segmental resyllabification account, however. First, the segmental identity of the following consonant (/b/ or /h/) may be as powerful a factor in conditioning the presence or absence of alveolar contact for some speakers. Second, glottalisation of lexically vowel-initial words regularly occurs, but does not seem to condition the appearance (or otherwise) of tongue tip contact. Third, the tongue dorsum remains retracted and does not adopt an onset-like form or timing even when /l/ is pre-vocalic. Thus categorical resyllabification of a word-final /l/ segment based on phonotactic acceptability is rejected as a mechanism controlling English L-sandhi in connected speech. Instead, we propose a gestural-episodic model, in which individual gestures display different levels of coherence in lexical syllable roles, while in connected speech, segmental sequences are influenced by similarity to well-rehearsed lexical sequences, if they exist.