A checklist names 267 failure modes in 13 families. Each mode has one acceptance case. A mode counts as contained only when its case passes with the two lines alone, and the decision log shows that Provely acted.

What are the families?

Family	What fails	Modes
`SYN`	The text of the arguments: quotes, commas, fences, truncation	33
`SHP`	The response and the argument values: types, required members, enums	29
`SEL`	The choice of the tool: unknown names, case, prefixes	13
`STR`	The structure of the calls in one turn: ids, duplicates, order	16
`STM`	Streaming: a stream that stops or arrives out of order	10
`PRO`	The protocol rules of each model API	31
`HAL`	Invented argument values	19
`CLM`	Invented claims of completion	8
`FRM`	The framework between the model and the tool	41
`INV`	The invocation of the tool: timeouts, errors, repeats	23
`RES`	What the tool or the provider answers	16
`OUT`	Third-party outcome semantics, named only	2
`COX`	Coexistence with your own defences	26

Which cases pass on every path?

20 of 27 basic grid cases passed on all 15 paths. Measured on 2026-10-17: TypeScript at 82acaae6, Python at 9f9131ba.

single_quotes: Single quotes around the argument names and values.
truncated_json: A truncated argument object.
json_wrong_type_array: The arguments are an array that holds the object.
json_wrong_type_string: The arguments are a string.
json_wrong_type_number: The arguments are a number.
json_null: The arguments are null.
member_wrong_type: A member of the wrong type.
missing_required_member: A required member is missing.
extra_member: An extra member that no person asked for.
enum_case_mismatch: An enum value in another letter case.
stringified_json: The argument object is a JSON string inside JSON.
tool_name_case: The tool name in another letter case.
tool_name_prefix: The tool name with a namespace prefix.
max_tokens_cut: The token limit cuts the arguments.
empty_body: An empty tool call.
ZC-SEED-01: A fenced argument with a trailing comma.
ZC-SEED-02: A tool name that is not in the catalogue.
ZC-SEED-06: Two parallel calls to one write tool.
ZC-SEED-07: Two invocations of one agent in one process.
ZC-SEED-07c: Three identical reads of a tool that the operator marks as a read.

Which cases do not pass everywhere?

Case	What the model or tool does	Did not pass on
`duplicate_call_id`	Two calls with one call identifier	`py/anthropic-messages`, `ts/anthropic-messages`, `ts/openai-responses`
`partial_stream`	The stream stops in the middle of the arguments	`py/anthropic-messages`, `ts/anthropic-messages`; not applicable on 10
`ZC-SEED-03`	An amount with no source in the run	`py/anthropic-messages`, `py/langchain`, `py/langgraph`, `py/llamaindex`, `py/openai-agents`, `py/openai-chat`, `ts/anthropic-messages`, `ts/langchain`, `ts/langgraph`, `ts/llamaindex`, `ts/openai-agents`, `ts/openai-chat`, `ts/openai-responses`, `ts/vercel-ai-generate`, `ts/vercel-ai-stream`
`ZC-SEED-04a`	A tool that hangs, with a timeout that the agent code states	`py/anthropic-messages`, `py/llamaindex`, `py/openai-agents`, `py/openai-chat`, `ts/anthropic-messages`, `ts/openai-chat`, `ts/openai-responses`; not applicable on 4
`ZC-SEED-04b`	A tool that hangs, with a timeout that the operator sets	`py/anthropic-messages`, `py/openai-chat`, `ts/anthropic-messages`, `ts/openai-chat`, `ts/openai-responses`
`ZC-SEED-05`	The model repeats a write after a tool error	`py/anthropic-messages`, `py/openai-chat`, `ts/anthropic-messages`, `ts/openai-chat`, `ts/openai-responses`
`ZC-SEED-07b`	One invocation sends the same refund in two turns	`py/anthropic-messages`, `py/openai-chat`, `ts/anthropic-messages`, `ts/openai-chat`, `ts/openai-responses`

ZC-SEED-03 fails under the default posture by design: the default records an invented value and does not refuse it. Read Postures for its result under enforce.

A plain loop runs the tool in your own code. Provely cannot reach a tool that hangs or raises inside your own handler. On a plain loop, ZC-SEED-04a and ZC-SEED-05 are out of reach, and out of reach never counts as a pass.