nemo_gym.atif_reverification

Strict ATIF projection for external rollout reverification.

The adapter validates a deliberately bounded canonical ATIF v1.7 subset, joins each trajectory to an explicit materialized Gym task, and constructs the Responses-shaped payload used by Gym’s existing stateless verifier path.

Provider-native payloads and producer-private status records carried in extra are producer evidence, not part of the ATIF interchange contract. Exporters own their projection into the canonical fields consumed here.

Module Contents

Classes

Name	Description
`AtifProjectionError`	Raised when an ATIF trajectory cannot be projected without changing what is scored.
`AtifReverifyManifestEntry`	Explicit join between one external trajectory and one Gym rollout.
`LoadedAtifTrajectory`	A validated trajectory together with immutable source provenance.
`ProjectedAtifVerifyPayload`	Verifier payload plus provenance that must survive result persistence.

Functions

Name	Description
`_apply_materialized_request_context`	-
`_created_at`	-
`_has_multimodal_content`	-
`_item_id`	-
`_model_name`	-
`_observation_output`	Project both standard ATIF content and Relay’s structured result extension.
`_optional_nonnegative_count`	-
`_reject_nonfinite_json_numbers`	-
`_reject_unsupported_training_metadata`	-
`_relay_usage_details`	Aggregate optional Relay token details without turning unknown values into zero.
`_step_has_usage_counts`	-
`_text_parts`	-
`_trajectory_content_sha256`	Hash semantic ATIF content without ignored producer metadata.
`_trajectory_identity_seed`	-
`_usage`	-
`_validate_declared_gym_source_identity`	-
`_validate_projected_output_ids`	-
`_validate_supported_trajectory`	-
`atif_trajectory_to_response`	Project one complete text-only ATIF v1.7 trajectory into a Gym response.
`build_atif_verify_payload`	Replace a materialized task’s native response with its ATIF projection.
`index_materialized_inputs`	Index materialized tasks without guessing identity from row order.
`load_atif_manifest`	Load an explicit trajectory-to-rollout manifest from JSONL.
`load_atif_trajectory`	Read and validate one ATIF document without involving the Relay runtime.
`project_atif_manifest_entries`	Project a one-to-one manifest while rejecting ambiguous batch joins.
`project_atif_manifest_entry`	Resolve one manifest entry without relying on filenames or session IDs.

Data

_RESPONSE_REQUEST_FIELDS

API

class nemo_gym.atif_reverification.AtifProjectionError()

Bases: ConfigError

Raised when an ATIF trajectory cannot be projected without changing what is scored.

class nemo_gym.atif_reverification.AtifReverifyManifestEntry()

Bases: BaseModel

Explicit join between one external trajectory and one Gym rollout.

expected_sha256

str | None = Field(default=None, pattern='^[0-9a-f]{64}$')

model_config

= ConfigDict(extra='forbid', populate_by_name=True)

rollout_index

int

task_index

int = Field(alias='_ng_task_index', ge=0, strict=True)

trajectory_path

Path

class nemo_gym.atif_reverification.LoadedAtifTrajectory(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7,
    source_path: pathlib.Path,
    source_sha256: str
)

Dataclass

A validated trajectory together with immutable source provenance.

source_path

Path

source_sha256

str

trajectory

AtifTrajectoryV1_7

class nemo_gym.atif_reverification.ProjectedAtifVerifyPayload(
    payload: dict[str, typing.Any],
    trajectory_id: str | None,
    session_id: str | None,
    source_path: pathlib.Path,
    source_sha256: str,
    trajectory_content_sha256: str,
    task_index: int,
    rollout_index: int,
    schema_version: str,
    projection_status: typing.Literal['complete'] = 'complete'
)

Dataclass

Verifier payload plus provenance that must survive result persistence.

projection_status="complete" means Gym built the bounded scoring payload without guessing or dropping a field this adapter maps. It does not claim a full-fidelity conversion of the ATIF document or producer metadata in extra; the source hashes preserve that provenance.

payload

dict[str, Any]

projection_status

Literal['complete'] = 'complete'

rollout_index

int

schema_version

str

session_id

str | None

source_path

Path

source_sha256

str

task_index

int

trajectory_content_sha256

str

trajectory_id

str | None

nemo_gym.atif_reverification._apply_materialized_request_context(
    response: nemo_gym.openai_utils.NeMoGymResponse,
    materialized_input: typing.Mapping[str, typing.Any]
) -> nemo_gym.openai_utils.NeMoGymResponse

nemo_gym.atif_reverification._created_at(
    agent_steps: list[typing.Any]
) -> float

nemo_gym.atif_reverification._has_multimodal_content(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> bool

nemo_gym.atif_reverification._item_id(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7,
    components: str = ()
) -> str

nemo_gym.atif_reverification._model_name(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> str

nemo_gym.atif_reverification._observation_output(
    result: typing.Any,
    field_name: str
) -> str | openai.types.responses.response_function_call_output_item_list_param.ResponseFunctionCallOutputItemListParam

Project both standard ATIF content and Relay’s structured result extension.

nemo_gym.atif_reverification._optional_nonnegative_count(
    namespace: typing.Mapping[str, typing.Any],
    field_name: str,
    step_id: int
) -> int | None

nemo_gym.atif_reverification._reject_nonfinite_json_numbers(
    value: typing.Any,
    field_name: str
) -> None

nemo_gym.atif_reverification._reject_unsupported_training_metadata(
    step: typing.Any
) -> None

nemo_gym.atif_reverification._relay_usage_details(
    agent_steps: list[typing.Any]
) -> tuple[int | None, int | None]

Aggregate optional Relay token details without turning unknown values into zero.

nemo_gym.atif_reverification._step_has_usage_counts(
    step: typing.Any
) -> bool

nemo_gym.atif_reverification._text_parts(
    value: typing.Any,
    field_name: str
) -> list[str]

nemo_gym.atif_reverification._trajectory_content_sha256(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> str

Hash semantic ATIF content without ignored producer metadata.

Provider payloads and status records carried in extra do not affect the projected verifier response. Exclude them from the content identity so they cannot perturb generated response or item IDs when trajectory_id is absent. Retain the few extension fields that this adapter consumes.

nemo_gym.atif_reverification._trajectory_identity_seed(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> str

nemo_gym.atif_reverification._usage(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> nemo_gym.openai_utils.NeMoGymResponseUsage | None

nemo_gym.atif_reverification._validate_declared_gym_source_identity(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7,
    manifest_key: tuple[int, int]
) -> None

nemo_gym.atif_reverification._validate_projected_output_ids(
    output: list[typing.Any]
) -> None

nemo_gym.atif_reverification._validate_supported_trajectory(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> None

nemo_gym.atif_reverification.atif_trajectory_to_response(
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> nemo_gym.openai_utils.NeMoGymResponse

Project one complete text-only ATIF v1.7 trajectory into a Gym response.

nemo_gym.atif_reverification.build_atif_verify_payload(
    materialized_input: dict[str, typing.Any],
    trajectory: nemo_gym.atif_v1_7.AtifTrajectoryV1_7
) -> dict[str, typing.Any]

Replace a materialized task’s native response with its ATIF projection.

The materialized task remains authoritative for the system/user input, verifier metadata, agent routing, and rollout correlation. ATIF contributes only the response produced by the external agent.

nemo_gym.atif_reverification.index_materialized_inputs(
    rows: typing.Iterable[typing.Mapping[str, typing.Any]]
) -> dict[tuple[int, int], dict[str, typing.Any]]

Index materialized tasks without guessing identity from row order.

nemo_gym.atif_reverification.load_atif_manifest(
    path: pathlib.Path
) -> list[nemo_gym.atif_reverification.AtifReverifyManifestEntry]

Load an explicit trajectory-to-rollout manifest from JSONL.

nemo_gym.atif_reverification.load_atif_trajectory(
    path: pathlib.Path
) -> nemo_gym.atif_reverification.LoadedAtifTrajectory

Read and validate one ATIF document without involving the Relay runtime.

nemo_gym.atif_reverification.project_atif_manifest_entries(
    entries: typing.Iterable[nemo_gym.atif_reverification.AtifReverifyManifestEntry],
    materialized_inputs: typing.Mapping[tuple[int, int], typing.Mapping[str, typing.Any]],
    manifest_directory: pathlib.Path
) -> list[nemo_gym.atif_reverification.ProjectedAtifVerifyPayload]

Project a one-to-one manifest while rejecting ambiguous batch joins.

nemo_gym.atif_reverification.project_atif_manifest_entry(
    entry: nemo_gym.atif_reverification.AtifReverifyManifestEntry,
    materialized_inputs: typing.Mapping[tuple[int, int], typing.Mapping[str, typing.Any]],
    manifest_directory: pathlib.Path
) -> nemo_gym.atif_reverification.ProjectedAtifVerifyPayload

Resolve one manifest entry without relying on filenames or session IDs.

nemo_gym.atif_reverification._RESPONSE_REQUEST_FIELDS = frozenset({'background', 'instructions', 'max_output_tokens', 'max_tool_calls', ...