Opus 5.5 can't write random bits
Random 300-bit sequences contain 8 zeroes in a row 44.2% of the time. In the model's 87 valid lines it happened 0 times (expected 38.5 ± 4.6, p ≈ 10−22). Two problems explain it.
Issue 1It switches between 0 and 1 too often
The model flips to the other bit on 69% of steps instead of 50%, so runs are cut short: the longest run of zeroes per line averages 2.45, against about 7.7 for random bits.
Issue 2It repeats its own lines
After the first few lines, the model mostly rewrote earlier ones shifted by a bit or two. 82 of 87 lines share a 40-bit stretch with another line, and the 87 lines hold only about 38 distinct sequences.
The prompt
Everything above comes from this single prompt, sent once to a fresh Opus 5.5 agent with no earlier conversation. The agent wrote all lines in one file write and used no other tools.
Generate 100 independent, uniformly random bit sequences, each exactly 300 characters long, using only the characters 0 and 1. One sequence per line, nothing else: no numbering, no spaces, no commentary. You must produce the bits yourself, directly. Do NOT use Bash, Python, any random number generator, or any code or tool to create or check them. Write all 100 lines in a single Write tool call to this file: llm_sequences.txt Do not read or edit any other file. When done, reply only with "done".
The original prompt gave a full local path; it is shortened here to the file name.
The agent's full reply: “done (Note: as instructed, I didn't check line lengths or bit balance with any tool. Many later lines fell into repeated patterns.)”