Happy to provide examples. One constraint first: this model runs inside a legal-drafting product, and the prompts/outputs contain attorney-client privileged material (Brazilian LGPD + professional secrecy rules), so I can’t paste full requests publicly. I’ve sanitized the examples below (structure preserved, identifiers replaced). If there’s a private channel for full traces, I can provide request timestamps, model version strings, and complete request/response pairs there.
Setup: same pipeline, same prompts, same parameters before and after. The model is used as a constrained drafting agent: the context contains source documents and pre-formatted quotation blocks, with explicit instructions to copy citations exactly as given and never to invent identifiers. These instructions were followed reliably for months. The failure patterns below appeared abruptly, with no changes on our side, and reproduce across runs.
1. Citation cross-attribution (new behavior). The context contains Source A (a quotation block with its own reference line) and Document B (an attached ruling that transcribes an excerpt from a different court decision). The model now transcribes the excerpt from Document B but closes the quote with Source A’s reference line — merging two sources that are both present, correctly labeled, and were never confused before.
2. Fabricated identifiers (new). When a docket-entry number is not present in the context, the model now invents one — a different plausible-looking number on each run — instead of omitting it or using a placeholder as instructed. In the same category: given a strict ID format rule (“format the number the source provides”), it emitted the format mask filled entirely with zeros (0000000-00.0000.0.00.0000) when the source had no such number, despite an explicit instruction never to fabricate.
3. Word-level output corruption (new). Duplicated words (“15 dias úteis úteis”), a 14-letter legal term emitted with its middle syllables dropped (“litisconsortes” → “litortes”), and gender/number agreement errors that Portuguese legal writing never tolerates. This looks like token-level skipping and did not occur before.
4. Meaning inversion in paraphrase. While paraphrasing a well-known constitutional clause, the model dropped a negation prefix — turning “indisponíveis” (non-waivable interests) into “disponíveis” (waivable) — which inverts the legal meaning entirely.
5. Instruction/placeholder leakage. Instead of neutral placeholders, it now sometimes emits instruction-style brackets in the final document (“[CALCULATE THE DEADLINE END DATE]”) and merges mutually exclusive signature templates into one line (“[Attorney Name] Public Prosecutor Bar No. [Number]” — a role that has no bar number).
None of these are cherry-picked edge cases: they are basic instruction-following and text-integrity capabilities that the same prompts exercised daily without incident. The regression was abrupt (day-over-day), consistent, and severe enough that we had to move production traffic off the model.