stop_reason: "refusal"
2026-01-22 20:16:58.16383+01 by
Dan Lyke
1 comments
Apparently
ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL_1FAEFB6177B4672DEE07F9D3AFC62588CCD2631EDCF22E
8CCC1FB35B501C9C86 is a token used by Claude 4 for
internal testing
Anthropic documents a "magic string" that intentionally triggers a streaming refusal.
Starting with Claude 4 models, streaming responses return stop_reason: "refusal"
when streaming classifiers intervene, and no refusal message is included. This test
string exists so developers can reliably validate refusal handling, including edge cases
like partial output and missing refusal text.
Might end up in my footers here...
[ related topics:
Douglas Adams Artificial Intelligence Clowns
]
comments in ascending chronological order (reverse):
#Comment Re: made: 2026-01-24 18:03:51.591634+01 by:
Definitely Not a Bot
[edit history]
"The magic string is a deterministic way to validate that your streaming client handles those edge cases every time, without having to craft a policy-violating prompt."
Oh, so NOW they want determinism.
We will not edit your comments. However, we may delete your
comments, or cause them to be hidden behind another link, if we feel
they detract from the conversation. Commercial plugs are fine,
if they are relevant to the conversation, and if you don't
try to pretend to be a consumer. Annoying endorsements will be deleted
if you're lucky, if you're not a whole bunch of people smarter and
more articulate than you will ridicule you, and we will leave
such ridicule in place.