cadrumo.adapters.outbound.llm.preconditions module

LLM-owned terminal refusal identities and fact-only verdict construction.

The LLM package owns refusals about dispatch, model output, and the transient consent token. These are deliberately distinct from provisioning’s hardware and installed-dependency decisions: a full on-host slot, an omitted off-host model, and malformed model output are not memory-headroom predicates. Each outcome is terminal because this package has no safe way to bind an executable CLI action from those observations.

class LLMPreconditionCondition(*values)[source]

Bases: StrEnum

Closed failed-condition identities owned by LLM dispatch and readers.

COLUMN_MAPPING_HEADERS_PRESENT
COLUMN_MAPPING_RESPONSE_PARSEABLE
COLUMN_MAPPING_RESPONSE_SCHEMA_VALID
EVIDENCE_IMAGES_PRESENT
EVIDENCE_OFF_HOST_DISPATCH_PERMITTED
EVIDENCE_RESPONSE_JSON_OBJECT
EVIDENCE_RESPONSE_SCHEMA_VALID
EVIDENCE_TEXT_PRESENT
EVIDENCE_TOKEN_BOUND
EVIDENCE_TOKEN_EPHEMERAL
EVIDENCE_TRANSCRIPTION_NONEMPTY
LOCAL_INFERENCE_SLOT_AVAILABLE
OFF_HOST_MODEL_NAMED
PROMPT_DEFINITION_ID_VALID
PROVIDER_CREDENTIALS_PRESENT
PROVIDER_SELECTION_VALID
REQUEST_PROMPT_NONEMPTY
VISION_INPUT_SUPPORTED
LLMPreconditionFact

Locale-neutral scalar fact allowed in an LLM terminal verdict.

llm_no_recovery_verdict(condition, *, facts, provenance, outcome=NoRecoveryOutcome.OPERATOR_DECISION)[source]

Build one fact-only terminal LLM outcome at the application boundary.

The import is deferred because application modules consume cadrumo.adapters.outbound.llm; importing the application carrier while defining this package would close that cycle. The returned object is nevertheless the canonical application-owned verdict type used by the shared CLI resolver.

Return type:

PreconditionVerdict

Parameters: