One wrong punctuation mark cost Claude models more accuracy than misspelling 70% of the words

The two breaks that context did not repair both sat on a delimiter ^[9]. A closing quote ends the scope of "inside the quoted paragraph." A colon before a list ends the instruction and starts the data. Remove either one and the prompt still reads as English, with the boundary in a different place ^[10][12]. The other breaks, such as a comma moved around "except", were recovered from context by every Claude model ^[4][9].

Detection did not produce correction. Opus 5.5 wrote "the closing quote is missing, so I counted all" and still returned the wrong number ^[11]. It diagnosed the bug correctly and shipped it anyway. Raising effort from low to medium made no difference for any model on any task ^[18]. "Structural breaks are not a reasoning problem, so more thinking does not fix them," the author wrote ^[19]. A prepended line telling the model to expect spelling, grammar and punctuation mistakes had zero effect on Opus 5, Haiku and Sonnet, and added a few points on Opus 5.5 and Fable ^[20].

The spelling result holds under one condition. The script misspelled 35% or 70% of the words in the long prompts but never touched the words the answer depended on ^[3]. The finding covers noise around the key terms. It transfers to a production prompt only if those terms arrive intact. Scoring was lenient: a reply counted as correct if the ideal answer appeared anywhere in it, explanation included ^[6]. A harness that parses one bare field is stricter. The hard set has nine prompts ^[3]. Weighted equally, each is about 11 points of a column, so a loss of 8 to 23 points ^[8] is roughly one to two prompts failing ^[1]. That matches the two breaks context did not repair ^[9].

The seven local models, run through Ollama, took the damage from dictated input ^[5]. On the voice-to-text profile, gemma4 fell 34 points and granite 37 ^[2]. The author traced it to spelled-out numbers and, again, the missing colon before a list; every local model scored 0% on the dictated list task ^[15]. The Claude models mostly absorbed dictation. Beyond Sonnet's run of misreads, Fable made the same "two fifty" error once, and Haiku turned "extension two zero one" into 2001 once ^[16]. "These are not model bugs. Spoken numbers are ambiguous, and the model has to guess," the author wrote ^[17].

The test introduced each break deliberately and did not include prompts built from templates ^[1][4]. I think templated prompts are where the delimiter finding applies most. A template is the code that puts the quote or colon between the instruction and the inserted text, and a bug there repeats on every call. I'd check what a template emits around inserted data before spending any effort on the spelling of what users type. The prompt that began "shop sellin pens 3dolar each" scored 100% on the Claude models ^[21].