Supported imports and compatibility

Kind: reference. This page lists supported from typevet… import paths and states compatibility expectations for the shipped wheel.

Parent theme: #29. Research baseline: #33. Layer ownership: #174.

typevet ships one distribution and one version (pyproject.toml / package metadata). Optional extras (cli) add dependencies only. They do not define a separate release line.

Import style

Prefer the package path that re-exports a name in that package’s __all__:

from typevet.domain import compile_json_schema, GenerationRequest
from typevet.adapters.outbound import LlamaCppGenerationAdapter
from typevet.runtime import ScoringJudgmentAdapter

Deep imports of the underlying module (for example typevet.domain.decision_compile) remain valid for the same objects. They are not the documented discovery path. Agents and new callers should start from the package __init__.py docstring and this page.

Import-linter contracts in pyproject.toml enforce hex layers. Do not import adapters from domain code or pull typevet.testing into adapters. See library-first architecture for the layer order.

Root typevet

The root __all__ declares these supported names:

Name	Role
AsyncGenerationPort	Structural protocol for async typed generation
BackendHttpError	llama.cpp HTTP error status with a body snippet
GenerationError	Base failure for a generation call
GenerationPort	Structural protocol for typed generation
GenerationRequest	Prompt, schema, and model ask
GenerationResult	Validated structured value
SchemaValidationError	Output failed the requested schema
TransportError	HTTP client failure before a response
__version__	Installed distribution version string
decide_categorical	M1 categorical decision via injected scoring port (native context= or inject_prefix=True)

__version__ comes from importlib.metadata.version("typevet"). An editable checkout without distribution metadata falls back to [project].version in pyproject.toml.

Convenience re-exports at the root mirror domain, ports, and runtime. For compiler types, judgment types, and scoring types, import from typevet.domain instead of the root.

typevet.domain

Name	Role
MAX_ENUM_CHOICES	Upper bound on enum size when compiling
MAX_PERMUTATIONS	Upper bound on enum permutation budget
Decision	One compiled TypeLLM field from JSON Schema
GenerationError	Base generation failure
BackendHttpError, TransportError	Generation failures raised by HTTP adapters
GenerationUnsupportedCapabilityError	The backend cannot honor the generation ask
GemmaTemplateError	Served template output is not a known Gemma or ChatML shape
GenerationRequest	Prompt, schema, and model ask
GenerationResult	Validated structured value
SchemaError	Invalid or unsupported schema for compilation
SchemaValidationError	Output failed the requested schema
compile_json_schema	Compile object schema to decisions
dependency_layers	Topological layers for decision dependencies
Noul, Choice, Score, Question	System One-shaped judgment questions
NoulAnswer, ChoiceAnswer, ScoreAnswer, Answer	Typed judgment answers; ScoreAnswer.score is the probability-weighted expected rubric level (float), not the modal level
JudgmentResponse, TokenUsage	Judgment call result and token metadata
JudgmentError, JudgmentValidationError	Judgment failure types
question_types	Map question ids to wire type names
CandidateScoringRequest, CandidateTokenSpec	Candidate logprob scoring ask
ImageInput	One image to condition a judgment on
MEDIA_MARKER, SUPPORTED_IMAGE_MIME_TYPES, count_media_markers	Media marker and accepted image mime types
CandidateScoringResult, ScoredCandidate, ScoringTermination	Scoring result and metadata
ScoreStage	Pre- vs post-sampling stage enum
ScoringError, ScoringValidationError, ScoringUnsupportedCapabilityError	Scoring failure types
build_and_validate_result	Fail-closed result assembly for scoring
CategoricalExecutionResult	Greedy categorical execute outcome
DecisionExecutionError	Categorical execute rejected inputs
execute_categorical_decision	Choice/Bool execution; calls the injected scoring port and performs no I/O itself
bind_control_candidates, judgment_original_labels	Control-token binding for native questions
normalize_noul, normalize_choice, normalize_score, normalize_question	Native question → Decision
question_record_to_property, question_records_to_json_schema, compile_question_records	Question records → JSON Schema or Decision values; defined in typevet.domain.question_schema (see question-schema-map.md)

typevet.ports

Name	Role
GenerationPort	Structural protocol for typed generation
AsyncGenerationPort	Structural protocol for async typed generation
JudgmentPort	Structural protocol for System One-shaped judgment
CandidateScoringPort, ScoringPort	Structural protocol for candidate logprobs (ScoringPort is an alias)
ModelFramingPort	Structural protocol for model turn framing

typevet.runtime

Thin orchestration facades over domain, ports, and outbound adapters (#148).

Name	Role
ScoringJudgmentAdapter	Sync JudgmentPort over CandidateScoringPort
judge_with_scoring	One-shot helper wrapping the adapter
decide_categorical	M1 categorical decision via injected scoring port
compose_scoring_prefix	Degraded ChatML scoring prefix composition
open_vllm_judgment	Context manager: judgment port over vLLM chat-completions scoring
VllmJudgmentSession	Session that open_vllm_judgment yields (port, client, model)
vllm_tokenize	Hook factory for vLLM /tokenize without special tokens
open_gemma_native_vision_judgment	Context manager: judgment port over the llama.cpp Gemma native vision factory
GemmaNativeVisionSession	Session that open_gemma_native_vision_judgment yields
probe_gemma_native_vision_support	Router probe for Gemma native vision support, without a long-lived port

open_vllm_judgment takes keyword client, model, and optional base_url, tokenize_content and scoring_port_wrapper. It sends no template probe. Scoring and /tokenize requests go to base_url, or to the client base_url when base_url is not set. The caller owns the client.

typevet.adapters.inbound.backend_settings.open_judgment selects the session from TYPEVET_BACKEND: the Gemma native vision factory for llama_cpp (the default), or open_vllm_judgment on the TYPEVET_VLLM__* client for vllm.

typevet.adapters.inbound

Name	Role
generate	Build a GenerationRequest and invoke a sync port
run_sync	Run an async generation coroutine from sync scripts
LlamaSettings	Frozen llama.cpp connection settings for composition roots
load_llama_settings	Read TYPEVET_LLAMA__* (and legacy aliases) from the environment
llama_cpp_adapter	Construct LlamaCppGenerationAdapter from LlamaSettings
VllmSettings	Frozen vLLM connection settings; api_key is not in repr
load_vllm_settings	Read TYPEVET_VLLM__* from the environment
vllm_http_client	Build the shared vLLM httpx.Client
async_vllm_generation_adapter	Construct AsyncVllmGenerationAdapter from VllmSettings
load_backend	Read TYPEVET_BACKEND
generation_adapter	Construct the generation adapter that TYPEVET_BACKEND selects

Environment names and CLI hookup notes live in configuration.md.

Signature: keyword-only prompt, schema, and model after the port argument. Prefer generate for library entry when you already hold a GenerationPort. Prefer run_sync(port.generate(request)) for AsyncGenerationPort in scripts instead of duplicating sync wrappers on each adapter.

The inbound package holds no command-line module. The eval runner command is in the typevet-evals workspace member. See Not in the wheel.

typevet.adapters.outbound

Name	Role
FakeGenerationAdapter	Offline adapter that validates a fixed or callable value
AsyncFakeGenerationAdapter	Offline async adapter that validates a fixed value
LlamaCppGenerationAdapter	OpenAI-compat llama.cpp router adapter
AsyncLlamaCppGenerationAdapter	Async OpenAI-compat llama.cpp router adapter
LlamaCppCandidateScoringAdapter	Pre-sampling /completion candidate scorer
VllmGenerationAdapter	vLLM structured-output generation adapter
AsyncVllmGenerationAdapter	Async vLLM generation adapter with a POST limit
VllmCandidateScoringAdapter	vLLM chat-completions logprob scorer
ChatContentFraming	Plain chat content framing for vLLM scoring

Constructors take explicit arguments only (no settings module on the adapter). See library-first architecture.

typevet.adapters.outbound.llama_cpp

This package groups the llama.cpp serving-backend modules.

Name	Role
LlamaCppGenerationAdapter	OpenAI-compat llama.cpp router adapter
AsyncLlamaCppGenerationAdapter	Async OpenAI-compat llama.cpp router adapter
LlamaCppCandidateScoringAdapter	Pre-sampling /completion candidate scorer

The package does not import the Gemma native vision factory. Import the factory from typevet.adapters.outbound.llama_cpp.gemma_native_vision_factory.

typevet.adapters.outbound.vllm

This package groups the vLLM serving-backend modules.

Name	Role
VllmGenerationAdapter	vLLM structured-output generation adapter
AsyncVllmGenerationAdapter	Async vLLM generation adapter with a POST limit
VllmCandidateScoringAdapter	vLLM chat-completions logprob scorer
ChatContentFraming	Plain chat content framing for vLLM scoring

The package does not import the vLLM judgment factory. Import the factory from typevet.adapters.outbound.vllm.judgment_factory, or import open_vllm_judgment from typevet.runtime.

typevet.adapters.outbound.gemma

This package holds Gemma and ChatML served-template constants, template classification and answer-binding helpers. It imports no serving backend.

Name	Role
CHATML_ASSISTANT_HEADER, CHATML_IM_START, CHATML_IM_END	ChatML turn control strings
GEMMA3_START_OF_TURN, GEMMA3_END_OF_TURN, GEMMA3_MODEL_TURN_HEADER	Gemma 3 turn control strings
GEMMA4_TURN_OPEN, GEMMA4_TURN_CLOSE, GEMMA4_MODEL_TURN_HEADER	Gemma 4 turn control strings
GEMMA4_CHANNEL_CLOSE, GEMMA4_THINK_TRIGGER, GEMMA4_NO_THINKING_PREFILL, GEMMA4_TOOL_RESPONSE	Gemma 4 channel, thinking and tool control strings
ServedTemplateClass	Served-template family enum
classify_served_template	Classify /apply-template output
stop_markers_for	Stop-marker substrings for one template class
label_embeds_control_fragment	Tell whether a label embeds a reserved control fragment
ThinkingDisposition	Thinking-channel disposition enum
AnswerAnchor, resolve_answer_anchor	Byte and token boundary before the first candidate token, and its resolver
bind_enum_label, bind_enum_labels	Bind enum labels to exact token ids and CandidateTokenSpec rows
termination_kind	Classify how an assistant completion ends
compose_scoring_prefix, compose_media_scoring_prefix	Scoring prefix text: degraded ChatML, or native turns with media

typevet.adapters.diagnostics

Name	Role
LogSettings, load_log_settings	Frozen log settings and the TYPEVET_LOG__* reader
configure, configure_from_environ	Apply log settings to stderr structlog
bind_run_id, new_run_id	Invocation id helpers
generation_call_event, http_request_event	Terminal diagnostic event context managers
diagnostic_model	Keep safe model aliases in event fields
REDACTED, SECRET_KEYS	Redaction placeholder and masked field names

Importing the package does not configure structlog. See diagnostic events.

typevet.testing

Name	Role
StaticGenerationFake	Fixed-value port double without adapter imports
ScriptedScoringFake	Scripted logprob double for CandidateScoringPort

Use for fast unit doubles and offline typed-judgment tutorials from an installed wheel. Use FakeGenerationAdapter when tests must exercise schema validation like production outbound code.

typevet.adapters

Organizational package only. It has no __all__ exports. Import from typevet.adapters.inbound, typevet.adapters.outbound, or typevet.adapters.diagnostics.

Not in the wheel

The typevet-evals workspace member in evals/ holds the evaluation code. Its import package is typevet_evals. The member is never published: its classifier is Private :: Do Not Upload. The library wheel holds only typevet. The member has no compatibility promise. Import it by submodule.

Module	Contents
typevet_evals.cli	Module entries eval_runner and cord_semantic_acceptance
typevet_evals.datasets	Dataset loaders, download helpers and the partner guard
typevet_evals.runner, typevet_evals.tpjep	Eval runner and TPJEP records, loader and runner
typevet_evals.cord, typevet_evals.psai_vision_consumer, typevet_evals.instruction_variant	Evaluation harness families
typevet_evals.throughput, typevet_evals.vllm_acceptance	Throughput and vLLM acceptance harnesses
typevet_evals.experiment_identity, typevet_evals.outcome_replay_metrics, typevet_evals.psai_vision_probability_evidence	Shared evaluation modules
typevet_evals.wheel_isolated, typevet_evals.gemma_native_vision_wheel_smoke	Wheel proof tooling

The library never imports typevet_evals. An import-linter contract enforces this rule. The root library tests may import typevet_evals. See ADR 0002.

Removed before 0.1.0

typevet removed these root modules before its first release (#256). They only re-exported names from their current home. Import from the current home.

Removed module	Current home
typevet.eval_runner_cli	typevet_evals.cli.eval_runner (workspace member, not in the wheel)
typevet.cord_semantic_acceptance_cli	typevet_evals.cli.cord_semantic_acceptance (workspace member, not in the wheel)
typevet.eval_runner, typevet.eval_runner_datasets, typevet.eval_runner_live_gate, typevet.eval_runner_report	typevet_evals.runner (workspace member, not in the wheel)
typevet.eval_tpjep_loader, typevet.eval_tpjep_outcome, typevet.eval_tpjep_records, typevet.eval_tpjep_runner	typevet_evals.tpjep (workspace member, not in the wheel)
Other typevet.eval_* loaders, download helpers, and guards	One submodule of typevet_evals.datasets (workspace member, not in the wheel)

typevet also moved the llama.cpp outbound modules into one package before its first release (#256). The six old module files are removed, and five of the old paths do not resolve. typevet.adapters.outbound.llama_cpp still resolves: it is now a package that re-exports the three adapters.

Removed module	Current home
typevet.adapters.outbound.llama_cpp (module)	typevet.adapters.outbound.llama_cpp.generation
typevet.adapters.outbound.async_llama_cpp	typevet.adapters.outbound.llama_cpp.generation_async
typevet.adapters.outbound.llama_cpp_http	typevet.adapters.outbound.llama_cpp.http_mapping
typevet.adapters.outbound.llama_cpp_multimodal	typevet.adapters.outbound.llama_cpp.multimodal
typevet.adapters.outbound.llama_cpp_scoring	typevet.adapters.outbound.llama_cpp.scoring
typevet.adapters.outbound.gemma_native_vision_factory	typevet.adapters.outbound.llama_cpp.gemma_native_vision_factory

typevet also moved the vLLM outbound modules into one package before its first release (#256). The six old module files do not resolve.

Removed module	Current home
typevet.adapters.outbound.vllm_content	typevet.adapters.outbound.vllm.content
typevet.adapters.outbound.vllm_generation	typevet.adapters.outbound.vllm.generation
typevet.adapters.outbound.vllm_generation_async	typevet.adapters.outbound.vllm.generation_async
typevet.adapters.outbound.vllm_http	typevet.adapters.outbound.vllm.http_mapping
typevet.adapters.outbound.vllm_scoring	typevet.adapters.outbound.vllm.scoring
typevet.adapters.outbound.vllm_judgment_factory	typevet.adapters.outbound.vllm.judgment_factory

typevet also moved these evaluation modules out of the library into the typevet-evals workspace member before its first release (#256). The wheel does not hold them, and the old paths do not resolve.

Removed module	Current home (workspace member, not in the wheel)
typevet.adapters.inbound.cord_semantic_acceptance_cli	typevet_evals.cli.cord_semantic_acceptance
typevet.adapters.inbound.eval_cli	typevet_evals.cli.eval_runner
typevet.testing.wheel_isolated	typevet_evals.wheel_isolated
typevet.evaluation.gemma_native_vision_wheel_smoke	typevet_evals.gemma_native_vision_wheel_smoke
typevet.evaluation.instruction_variant_consumer_* (8 modules)	typevet_evals.instruction_variant.*, with the prefix removed (for example …_offline → offline)
typevet.evaluation.outcome_replay_metrics	typevet_evals.outcome_replay_metrics
typevet.evaluation.collections_metrics, …collections_throughput, …collections_workload, …public_workload	typevet_evals.throughput (same module names)
typevet.evaluation.vllm_acceptance, …vllm_acceptance_sets, …vllm_acceptance_transport	typevet_evals.vllm_acceptance.core, .sets, .transport
typevet.evaluation.psai_vision_consumer_* (12 modules)	typevet_evals.psai_vision_consumer.*, with the prefix removed (for example …_harness → harness)
typevet.evaluation.consumer_http_accounting	typevet_evals.psai_vision_consumer.http_accounting
typevet.evaluation.psai_vision_probability_evidence	typevet_evals.psai_vision_probability_evidence
typevet.evaluation.cord_* (7 modules)	typevet_evals.cord.*, with the cord_ prefix removed (for example …_semantic_acceptance → semantic_acceptance)
typevet.evaluation.runner and its 4 modules	typevet_evals.runner (same module names)
typevet.evaluation.tpjep and its 4 modules	typevet_evals.tpjep (same module names)
typevet.evaluation.experiment_identity	typevet_evals.experiment_identity
typevet.evaluation.datasets and its 22 modules, with the clinc_*.json data files	typevet_evals.datasets (same module names)
typevet.evaluation	No replacement. The package is removed; each evaluation family is a typevet_evals subpackage or module

Command-line entry

typevet declares no console script. pyproject.toml has no [project.scripts] table, so installing the wheel does not put a typevet command on PATH.

The wheel ships no command. In a checkout, the eval runner and the CORD receipt acceptance check are Python module entries in the typevet-evals workspace member:

uv run python -m typevet_evals.cli.eval_runner --help
uv run python -m typevet_evals.cli.cord_semantic_acceptance path/to/receipt.json

Without TYPEVET_LLAMA__* router settings, the command prints a skip reason and exits 0. See live eval runner.

Typing and packaging

• src/typevet/py.typed marks the package as typed for consumers.
• Public API surface is the union of package __all__ lists and documented module docstrings checked by docvet.
• Version string authority is [project].version in pyproject.toml. typevet.__version__ reads it back from distribution metadata. See verify package typing and version.

0.1.0 compatibility assessment

Surface	Assessment	Notes
Library	Initial public hex surface	Root, domain, ports, runtime, inbound, outbound (with llama_cpp, vllm and gemma), diagnostics, testing
Evaluation	Not shipped	All evaluation code is in the typevet-evals workspace member. See Not in the wheel
CLI	Not shipped	No console script and no module entry in the wheel. In a checkout: python -m typevet_evals.cli.eval_runner and python -m typevet_evals.cli.cord_semantic_acceptance. The cli extra lists Typer only and no module imports it
MCP	Not shipped	No extra or entry point
Dependencies	httpx, jsonschema, structlog	Locked via uv.lock in development

Additive changes should extend __all__ and this page. Breaking renames or removed exports require an explicit compatibility note in a future release section on this page.

Verification limits

Contract tests with FakeGenerationAdapter do not prove a live llama.cpp router behaves correctly. Live tests use the live pytest marker and opt-in runs. See verification.