Failures and limits

Every result has two fields that together decide ok:

• decision is "block" when any finding blocks, and "allow" otherwise.
• complete is false when a configured transform or judgment did not finish.

ok is decision == "allow" and complete. Branch on ok, or forward onward. onward is a refusal whenever ok is false, including on fail-open paths, and it never contains the checked text.

Backend errors

A judge call can fail: a timeout, a rate limit, a TypeSafe error, a network failure, a backend that returns nothing, or answers that are missing or don't match the question. The judge lists the BackendError reason for each. Choose the behavior with Guard(..., on_backend_error=...):

Mode	What happens	decision	complete	ok
"raise" (default)	The check raises BackendError	n/a	n/a	n/a
"block"	Adds a backend_error block finding	block	False	False
"allow"	Adds a backend_error flag finding and omits the failed scores	allow unless something else blocks	False	False

"allow" is an explicit fail-open. It doesn't add a block for the failure, but the result is still incomplete, so ok is false.

Deadlines

deadline_s (default 30 seconds) is a per-check duration. A check sends its requests to the judge in parallel, up to Limits.max_concurrency at once across the guard. Requests still running at the deadline count as backend errors: DeadlineExceeded, a subclass of BackendError. It follows on_backend_error with the label deadline_exceeded. AsyncGuard cancels those requests, and also cancels them when the caller cancels the check. Guard stops waiting for them and returns.

Transforms check the deadline between steps. A transform cannot interrupt a synchronous third-party call that is already running, such as a Presidio analysis. jes applies the deadline outcome when the call returns.

Resource caps

Set the caps with Guard(..., limits=Limits(...)):

from jes import Guard, Limits

guard = Guard(policies, model="jev-latest", limits=Limits(max_input_bytes=65_536))

Size and count problems are findings, not exceptions. They block with complete=False before more work is scheduled. Examples:

Finding	Cause
input_too_long	The text exceeds max_input_bytes
normalized_too_long, invalid_unicode	A transform grew the text to over four times its limit, or the text has a lone surrogate
context_too_long, too_many_context_items	The context exceeds its caps
text_too_long	A whole_text judgment does not fit the judge's max_request_bytes
too_many_chunks, too_many_items, too_many_requests	Chunk, item, or request caps were reached
too_many_redactions, redaction_store_full	One check replaced over max_redactions values, or the store is full
restored_output_too_large	Restoring placeholders grew the reply to over four times max_input_bytes

Two findings only flag: context_dropped, when an "optional" context does not fit next to the text, and history_truncated, when the oldest history entries were left out to make it fit.

The seven caps and their defaults are in Limits.

Exceptions

Exception	When
PolicyError	Construction: duplicate names, missing threshold, an invalid limit, and so on
PolicyExecutionError	During a check: a custom policy broke its contract. Always raised.
RedactionError	A store or scope mismatch, or a serialized blob fails to authenticate
BackendError	A transport failure or malformed answer, when on_backend_error="raise"
DeadlineExceeded	A subclass of BackendError

The runnable version is Failure and limits cookbook page.