solveedit_benchmark_2728 / code /src /pvr_eval /agent /evidence_program.py
wenjieshu's picture
Add files using upload-large-folder tool
b8ea052 verified
Raw History Blame Contribute Delete
36.5 kB
"""Strict compiler from one atomic contract check to an evidence program.
The resolver's prose is an audit artifact, not an executable interface. This
module turns the fields retained on :class:`AtomicCheckSpec` into a typed,
per-check program and refuses to call it executable when a required operation
or region cannot actually be resolved.
The compiler deliberately does not create physical model calls. It is the
semantic layer immediately before ``tool_plan``: executors may batch/cache the
typed requirements later, while preserving ``check_id`` provenance.
"""
from __future__ import annotations
from copy import deepcopy
from dataclasses import dataclass, field
import re
from typing import Any, Iterable, Mapping
from pvr_eval.core.schema import AtomicCheckSpec
from pvr_eval.agent.contract_metric_binding import canonical_rule_type, metric_capability
EVIDENCE_PROGRAM_SCHEMA_VERSION = "pvr.evidence_program.v1"
# These are operations for which the repository has an objective execution
# path. Callers should pass their runtime capability set when it differs.
DEFAULT_AVAILABLE_OPERATIONS = frozenset(
{
"crop",
"object_detection",
"mask_refinement",
"instance_matching",
"source_instance_selection",
"crop_classification",
"image_registration",
"change_analysis",
"ocr",
"text_alignment",
"text_quality",
"route_extraction",
"state_parsing",
"structured_check",
"visual_attribute",
"pose_estimation",
"local_correspondence",
"spatial_geometry",
"counting",
"region_resolution",
"source_state_classification",
"predicate_decision",
}
)
@dataclass(frozen=True)
class RegionDependency:
region_id: str
image_roles: tuple[str, ...]
proposal_method: str
resolver_operation: str
status: str
query_or_rule: str = ""
geometry: Any = None
reasons: tuple[str, ...] = ()
def to_dict(self) -> dict[str, Any]:
return {
"region_id": self.region_id,
"image_roles": list(self.image_roles),
"proposal_method": self.proposal_method,
"resolver_operation": self.resolver_operation,
"status": self.status,
"query_or_rule": self.query_or_rule,
"geometry": self.geometry,
"reasons": list(self.reasons),
}
@dataclass(frozen=True)
class EvidenceRequirement:
requirement_id: str
operation: str
tool_family: str
image_roles: tuple[str, ...]
region_ids: tuple[str, ...] = ()
depends_on: tuple[str, ...] = ()
source: str = "inferred"
raw: str = ""
optional: bool = False
status: str = "supported"
reasons: tuple[str, ...] = ()
def to_dict(self) -> dict[str, Any]:
return {
"requirement_id": self.requirement_id,
"operation": self.operation,
"tool_family": self.tool_family,
"image_roles": list(self.image_roles),
"region_ids": list(self.region_ids),
"depends_on": list(self.depends_on),
"source": self.source,
"raw": self.raw,
"optional": self.optional,
"status": self.status,
"reasons": list(self.reasons),
}
@dataclass(frozen=True)
class EvidenceArtifact:
artifact_id: str
description: str
producer_operations: tuple[str, ...]
status: str
reasons: tuple[str, ...] = ()
def to_dict(self) -> dict[str, Any]:
return {
"artifact_id": self.artifact_id,
"description": self.description,
"producer_operations": list(self.producer_operations),
"status": self.status,
"reasons": list(self.reasons),
}
@dataclass(frozen=True)
class ParameterPromotion:
checker_params: dict[str, Any]
promoted: dict[str, str]
conflicts: tuple[str, ...]
@dataclass(frozen=True)
class EvidenceProgram:
check_id: str
checker: str
status: str
checker_params: dict[str, Any]
promoted_parameters: dict[str, str]
parameter_conflicts: tuple[str, ...]
regions: tuple[RegionDependency, ...]
requirements: tuple[EvidenceRequirement, ...]
evidence_artifacts: tuple[EvidenceArtifact, ...]
unsupported_requirements: tuple[str, ...]
reasons: tuple[str, ...]
@property
def executable(self) -> bool:
return self.status == "executable"
def to_dict(self) -> dict[str, Any]:
return {
"schema_version": EVIDENCE_PROGRAM_SCHEMA_VERSION,
"check_id": self.check_id,
"checker": self.checker,
"status": self.status,
"executable": self.executable,
"checker_params": self.checker_params,
"promoted_parameters": self.promoted_parameters,
"parameter_conflicts": list(self.parameter_conflicts),
"regions": [item.to_dict() for item in self.regions],
"requirements": [item.to_dict() for item in self.requirements],
"evidence_artifacts": [item.to_dict() for item in self.evidence_artifacts],
"unsupported_requirements": list(self.unsupported_requirements),
"reasons": list(self.reasons),
}
_PARAM_ALIASES = {
"required_anchor_strings": "expected_all",
"pass_mask_containment": "min_mask_containment",
"partial_mask_containment": "partial_min_mask_containment",
"slot_expand_px_for_detection": "slot_expand_px",
"min_detector_score": "min_score",
"normalized_edit_similarity_min": "min_text_similarity",
"output_ocr_confidence_min": "min_ocr_confidence",
"minimum_match_confidence": "min_match_confidence",
"minimum_start_zone_intersection": "min_start_zone_intersection",
"minimum_end_zone_intersection": "min_end_zone_intersection",
"maximum_normalized_skeleton_gap": "max_normalized_skeleton_gap",
"minimum_route_corridor_coverage": "min_route_corridor_coverage",
"maximum_invariant_changed_area_ratio": "pass_max_changed",
"maximum_unexplained_architectural_change_fraction": "pass_max_changed",
"maximum_unexplained_region_change_ratio": "pass_max_changed",
}
def _checker_base(checker: str) -> str:
return str(checker).split("|", 1)[0].split(":", 1)[0].strip()
def _decision_parameters(params: Mapping[str, Any]) -> dict[str, Any]:
rule = params.get("decision_rule", {})
if isinstance(rule, Mapping):
nested = rule.get("parameters", {})
return dict(nested) if isinstance(nested, Mapping) else {}
return {}
def promote_checker_parameters(spec: AtomicCheckSpec) -> ParameterPromotion:
"""Promote decision parameters without overwriting explicit checker data."""
original = deepcopy(dict(spec.params or {}))
promoted: dict[str, str] = {}
conflicts: list[str] = []
for source_key, value in _decision_parameters(original).items():
target_key = _PARAM_ALIASES.get(str(source_key), str(source_key))
if target_key not in original:
original[target_key] = deepcopy(value)
promoted[str(source_key)] = target_key
elif original[target_key] != value:
conflicts.append(f"decision_parameter_conflict:{source_key}->{target_key}")
# Materialize concrete regions in the schema consumed by current checkers.
zones = deepcopy(original.get("zones")) if isinstance(original.get("zones"), Mapping) else {}
slot = str(original.get("slot", "")).strip()
slot_bbox = original.get("slot_bbox")
if slot and _is_bbox(slot_bbox) and slot not in zones:
zones[slot] = {"id": slot, "bbox": list(slot_bbox)}
promoted["slot_bbox"] = f"zones.{slot}"
for item in _planned_regions(original):
region_id = str(item.get("region_id") or item.get("id") or "").strip()
geometry = _region_geometry(item)
if region_id and geometry is not None and region_id not in zones:
zones[region_id] = {"id": region_id, **geometry}
promoted[f"planned_regions.{region_id}"] = f"zones.{region_id}"
if zones:
original["zones"] = zones
return ParameterPromotion(original, promoted, tuple(conflicts))
def _slug(value: str) -> str:
return re.sub(r"[^a-z0-9]+", "_", value.lower()).strip("_") or "requirement"
def _roles(value: Any) -> tuple[str, ...]:
text = str(value or "pred").strip().lower()
if text in {"both", "input_output", "input+output", "input_pred"}:
return ("input", "pred")
if text in {"output", "prediction", "model_output", "pred"}:
return ("pred",)
if text == "input":
return ("input",)
return ()
def _is_bbox(value: Any) -> bool:
return isinstance(value, (list, tuple)) and len(value) == 4 and all(isinstance(v, (int, float)) for v in value)
def _region_geometry(item: Mapping[str, Any]) -> dict[str, Any] | None:
for key in ("bbox", "bbox_xyxy", "polygon", "mask", "rect"):
value = item.get(key)
if value not in (None, [], {}, ""):
return {"bbox" if key in {"bbox_xyxy", "rect"} else key: deepcopy(value)}
return None
def _bbox_from_rule(value: Any) -> list[float] | None:
"""Read an explicitly supplied bbox from resolver prose.
Resolved plans sometimes keep a frozen zone as ``Use supplied bbox
[x0,y0,x1,y1]`` instead of duplicating it into a separate geometry field.
That is still contract geometry and must not be replaced by a detector
guess.
"""
match = re.search(
r"\bbbox\s*\[\s*(-?\d+(?:\.\d+)?)\s*,\s*(-?\d+(?:\.\d+)?)\s*,\s*"
r"(-?\d+(?:\.\d+)?)\s*,\s*(-?\d+(?:\.\d+)?)\s*\]",
str(value or ""),
flags=re.IGNORECASE,
)
if not match:
return None
bbox = [float(item) for item in match.groups()]
if bbox[2] <= bbox[0] or bbox[3] <= bbox[1]:
return None
return bbox
def _planned_regions(params: Mapping[str, Any]) -> list[dict[str, Any]]:
raw = params.get("planned_regions", [])
return [dict(item) for item in raw if isinstance(item, Mapping)] if isinstance(raw, list) else []
_REGION_METHODS = {
"grounding_dino_box": "object_detection",
"dino_box": "object_detection",
"yolo_box": "object_detection",
"yolo_world_box": "object_detection",
# DINO proposes a concrete box first; SAM may refine it downstream.
"sam_mask": "object_detection",
"ocr_layout": "ocr",
"route_mask": "route_extraction",
"grid_parser": "state_parsing",
"color_blob_parser": "visual_attribute",
"cv_segmentation": "state_parsing",
# Region proposal is objective-tool first. A VLM remains the fallback if
# open-vocabulary detection cannot materialize the requested region.
"vlm_region_box": "object_detection",
"vlm_box": "object_detection",
"manual": "object_detection",
"manual_annotation": "object_detection",
}
def _compile_regions(params: Mapping[str, Any], available: frozenset[str]) -> tuple[RegionDependency, ...]:
result: list[RegionDependency] = []
decision = _decision_parameters(params)
for index, item in enumerate(_planned_regions(params)):
region_id = str(item.get("region_id") or item.get("id") or f"region_{index}").strip()
method = _slug(str(item.get("proposal_method") or ""))
materializer = item.get("materializer") if isinstance(item.get("materializer"), Mapping) else {}
materializer_kind = str(materializer.get("kind") or "")
roles = _roles(item.get("source_image"))
query = str(item.get("query_or_rule") or "").strip()
geometry = _region_geometry(item)
if geometry is None and method == "existing_zone":
supplied_bbox = _bbox_from_rule(query)
if supplied_bbox is not None:
geometry = {"bbox": supplied_bbox}
reasons: list[str] = []
resolver = ""
if geometry is None and region_id == str(decision.get("slot", params.get("slot", ""))) and _is_bbox(decision.get("slot_bbox", params.get("slot_bbox"))):
geometry = {"bbox": list(decision.get("slot_bbox", params.get("slot_bbox")))}
elif geometry is None and region_id.endswith("_exact") and _is_bbox(decision.get("slot_bbox", params.get("slot_bbox"))):
geometry = {"bbox": list(decision.get("slot_bbox", params.get("slot_bbox")))}
if not roles:
reasons.append("unknown_region_image_role")
if geometry is not None:
status = "concrete" if roles else "blocked"
resolver = "existing_geometry"
elif method == "existing_zone":
resolver = "object_detection"
query = query or region_id.replace("_", " ")
status = "auto_resolvable" if resolver in available else "blocked"
if resolver not in available:
reasons.append(f"unavailable_region_resolver:{resolver}")
else:
resolver = {
"detector_segmenter": "object_detection",
"registered_change_derivation": "change_analysis",
"ocr_container_layout": "ocr",
"anchor_relative_zone": "state_parsing",
"shelf_grid_parser": "state_parsing",
"planar_free_space_parser": "state_parsing",
"derived_free_space_zone": "state_parsing",
"architecture_region_parser": "state_parsing",
}.get(materializer_kind, _REGION_METHODS.get(method, ""))
if not resolver:
status = "blocked"
reasons.append(f"unsupported_region_proposal_method:{method or 'missing'}")
elif resolver not in available:
status = "blocked"
reasons.append(f"unavailable_region_resolver:{resolver}")
elif not query:
if resolver == "object_detection":
query = region_id.replace("_", " ")
status = "auto_resolvable"
else:
status = "blocked"
reasons.append("auto_region_has_no_query_or_rule")
else:
status = "auto_resolvable"
result.append(
RegionDependency(
region_id=region_id,
image_roles=roles,
proposal_method=method,
resolver_operation=resolver,
status=status,
query_or_rule=query,
geometry=geometry,
reasons=tuple(reasons),
)
)
return tuple(result)
def normalize_tool_operation(value: Any) -> tuple[str, str, bool] | None:
"""Return ``(operation, family, optional)`` for a resolver tool phrase."""
text = _slug(str(value or ""))
if not text:
return None
optional = any(token in text for token in ("fallback", "audit_if", "targeted_vlm", "uncertain", "low_confidence"))
if "vlm" in text:
return ("vlm_audit", "vlm", True)
if any(token in text for token in ("grounding", "dino", "yolo", "open_vocabulary", "object_detection", "person_detection")):
return ("object_detection", "detector", optional)
if any(token in text for token in ("artifact_detection", "artifact_localization", "piece_detection", "card_detection", "cassette_detector", "tray_localization", "shelving_unit_detection", "calendar_region_detection")):
return ("object_detection", "detector", optional)
if "sam" in text or "instance_segmentation" in text or "mask_refin" in text:
return ("mask_refinement", "segmenter", optional)
if any(token in text for token in ("registration", "homography", "alignment_by_feature")):
return ("image_registration", "registration", optional)
if any(token in text for token in ("instance_matching", "appearance_matching", "identity_matching", "embedding_comparison")):
return ("instance_matching", "matcher", optional)
if any(token in text for token in ("ocr_string_alignment", "string_edit", "string_alignment")):
return ("text_alignment", "ocr_postprocess", optional)
if any(token in text for token in ("sharpness", "text_contrast", "legibility")):
return ("text_quality", "image_metrics", optional)
if "ocr" in text or "text_recognition" in text or "text_detection" in text:
return ("ocr", "ocr", optional)
if any(token in text for token in ("route_skeleton", "route_graph", "endpoint_matching", "connectivity", "centerline")):
return ("route_topology", "route_parser", optional)
if any(token in text for token in ("lane_arrow", "direction_classification", "movement_classification")):
return ("route_movement_classification", "route_parser", optional)
if any(token in text for token in ("route_line", "route_mask", "path_mask", "change_mask")):
operation = "change_analysis" if "change_mask" in text else "route_extraction"
return (operation, "change_analyzer" if operation == "change_analysis" else "route_parser", optional)
if any(token in text for token in ("lane_boundary", "crosswalk", "stop_line", "waiting_box_segmentation")):
return ("road_layout_parsing", "state_parser", optional)
if any(token in text for token in ("grid", "slot_parser", "board_parser", "rack_parser", "shelf_parser", "state_parser")):
return ("state_parsing", "state_parser", optional)
if any(token in text for token in ("color", "attribute_class", "symbol", "checkbox", "glyph", "transparency")):
return ("visual_attribute", "attribute_parser", optional)
if any(token in text for token in ("difference", "perceptual", "image_similarity", "edge_comparison", "residual_occupancy", "vacancy_comparison")):
return ("change_analysis", "change_analyzer", optional)
if any(token in text for token in ("preservation", "residual_edge", "residual_texture", "source_zone_residual", "allowed_edit_mask")):
return ("change_analysis", "change_analyzer", optional)
if any(token in text for token in ("containment", "overlap", "support_geometry", "distance", "centroid", "geometry")):
return ("spatial_geometry", "checker_internal", optional)
if any(token in text for token in ("counting", "residual_count", "enumerate")):
return ("counting", "checker_internal", optional)
if any(token in text for token in ("connected_component", "instance_deduplication", "duplicate_detection_suppression")):
return ("counting", "checker_internal", optional)
if any(token in text for token in ("source_zone_construction", "existing_zone_loader")):
return ("region_resolution", "checker_internal", optional)
if text.startswith("crop") or "roi_upscaling" in text:
return ("crop", "image_preprocess", optional)
return None
def _checker_requirements(spec: AtomicCheckSpec, params: Mapping[str, Any]) -> list[tuple[str, str, tuple[str, ...]]]:
checker = _checker_base(spec.checker)
rule = params.get("decision_rule", {})
decision_type = str(rule.get("type", "")) if isinstance(rule, Mapping) else ""
canonical_type = canonical_rule_type(decision_type)
out: list[tuple[str, str, tuple[str, ...]]] = []
if checker in {"object_presence_checker", "distance_checker"}:
out.append(("object_detection", "detector", ("pred",)))
if checker in {"object_zone_checker", "source_vacated_checker", "vehicle_slot_checker", "vehicle_penalty_checker"}:
out.extend(
[
("object_detection", "detector", ("pred",)),
("mask_refinement", "segmenter", ("pred",)),
("spatial_geometry", "checker_internal", ("pred",)),
]
)
if checker in {"object_preservation_checker", "source_vacated_checker"}:
out.extend([("object_detection", "detector", ("input", "pred")), ("instance_matching", "matcher", ("input", "pred"))])
if checker in {"vehicle_slot_checker", "vehicle_penalty_checker", "color_classifier"}:
out.append(("visual_attribute", "attribute_parser", ("pred",)))
if checker in {"text_ocr_checker", "date_window_digit_checker"}:
roles = ("input", "pred") if any(token in decision_type for token in ("preserv", "unchanged", "all_text")) else ("pred",)
out.append(("ocr", "ocr", roles))
if checker in {"route_mask_checker", "endpoint_checker", "forbidden_overlap", "line_connectivity_checker"}:
out.extend([("route_extraction", "route_parser", ("input", "pred")), ("route_topology", "route_parser", ("pred",))])
if _decision_parameters(params).get("expected_movements"):
out.append(("route_movement_classification", "route_parser", ("input", "pred")))
if checker in {"state_checker", "state_image_proxy_checker", "grid_state_parser", "watch_hand_angle_parser"}:
out.append(("state_parsing", "state_parser", ("pred",)))
if checker in {"preservation", "preservation_checker", "change_mask_zone", "input_output_similarity_guard"}:
out.extend([("image_registration", "registration", ("input", "pred")), ("change_analysis", "change_analyzer", ("input", "pred"))])
capability = metric_capability(canonical_type)
if capability is not None:
out.extend(
(item.operation, item.tool_family, item.image_roles)
for item in capability.evidence
)
return out
def _artifact_operations(text: str) -> tuple[str, ...]:
normalized = text.lower()
operations: list[str] = []
is_text_evidence = any(token in normalized for token in ("ocr", "string", "text", "line order"))
groups = [
(("ocr", "string", "text", "line order"), "ocr"),
(("detector", "detection", "label"), "object_detection"),
(("sam", "mask"), "mask_refinement"),
(("register", "corresponding roi"), "image_registration"),
(("match", "identity", "appearance embedding"), "instance_matching"),
(("containment", "overlap", "support-contact", "distance", "centroid"), "spatial_geometry"),
(("route", "skeleton", "graph", "path", "endpoint"), "route_topology"),
(("color", "transparency", "attribute"), "visual_attribute"),
(("count", "inventory"), "counting"),
(("sharpness", "contrast", "legibility"), "text_quality"),
(("difference", "changed", "residual occupancy"), "change_analysis"),
(("region", "zone", "polygon", "roi"), "region_resolution"),
]
for needles, operation in groups:
if any(needle in normalized for needle in needles):
operations.append(operation)
if "box" in normalized and not is_text_evidence:
operations.append("object_detection")
return tuple(dict.fromkeys(operations))
def _dedupe_requirements(items: Iterable[EvidenceRequirement]) -> tuple[EvidenceRequirement, ...]:
result: list[EvidenceRequirement] = []
seen: set[tuple[Any, ...]] = set()
for item in items:
key = (item.operation, item.image_roles, item.region_ids, item.optional, item.source)
if key not in seen:
seen.add(key)
result.append(item)
return tuple(result)
def compile_evidence_program(
spec: AtomicCheckSpec,
*,
available_operations: Iterable[str] | None = None,
contract_zones: Mapping[str, Any] | None = None,
) -> EvidenceProgram:
"""Compile one check and expose every objective execution blocker."""
available = frozenset(available_operations) if available_operations is not None else DEFAULT_AVAILABLE_OPERATIONS
promotion = promote_checker_parameters(spec)
params = deepcopy(promotion.checker_params)
predicate = params.get("executable_predicate") if isinstance(params.get("executable_predicate"), Mapping) else None
predicate_required = bool(params.get("require_executable_predicate"))
if predicate_required and (predicate is None or predicate.get("status") != "executable"):
predicate_reasons = (
["missing_executable_predicate"]
if predicate is None
else [str(item) for item in predicate.get("reasons", [])]
)
return EvidenceProgram(
check_id=spec.check_id,
checker=spec.checker,
status="blocked",
checker_params=params,
promoted_parameters=promotion.promoted,
parameter_conflicts=promotion.conflicts,
regions=(),
requirements=(),
evidence_artifacts=(),
unsupported_requirements=("executable_predicate",),
reasons=tuple(dict.fromkeys(("executable_predicate_not_compiled", *predicate_reasons))),
)
predicate_authoritative = bool(predicate_required and predicate is not None)
authoritative_conflicts = (
()
if predicate_authoritative and str(predicate.get("family") or "") == "structured_case_parser"
else promotion.conflicts
)
if predicate_authoritative:
# Tool operations and derived-region semantics now come from the
# executable predicate. Legacy prose remains in the contract only as
# review provenance and must not be reinterpreted by this compiler.
params["planned_regions"] = [
deepcopy(dict(item))
for item in predicate.get("regions", [])
if isinstance(item, Mapping) and (item.get("region_id") or item.get("id"))
]
# Top-level contract zones are part of the executable contract even though
# AtomicCheckSpec only carries per-check params. Make them visible to the
# region compiler and, later, to the fixed checker.
if contract_zones:
zones = deepcopy(dict(contract_zones))
zones.update(dict(params.get("zones", {})) if isinstance(params.get("zones"), Mapping) else {})
params["zones"] = zones
planned = []
for item in _planned_regions(params):
region = deepcopy(item)
region_id = str(region.get("region_id") or region.get("id") or "").strip()
zone = zones.get(region_id)
if _region_geometry(region) is None and isinstance(zone, Mapping):
geometry = _region_geometry(zone)
if geometry:
region.update(geometry)
planned.append(region)
if planned:
params["planned_regions"] = planned
regions = _compile_regions(params, available)
# Geometry recovered by the strict region compiler must also reach the
# fixed checker/CV executor. Keeping it only on RegionDependency would
# make the plan look concrete while the runtime still saw a prose-only
# placeholder.
concrete_regions = {item.region_id: item.geometry for item in regions if item.geometry is not None}
if concrete_regions:
zones = deepcopy(dict(params.get("zones", {}))) if isinstance(params.get("zones"), Mapping) else {}
for region_id, geometry in concrete_regions.items():
if region_id not in zones:
zones[region_id] = {"id": region_id, **deepcopy(dict(geometry))}
params["zones"] = zones
planned = []
for item in _planned_regions(params):
region = deepcopy(item)
region_id = str(region.get("region_id") or region.get("id") or "").strip()
geometry = concrete_regions.get(region_id)
if geometry is not None and _region_geometry(region) is None:
region.update(deepcopy(dict(geometry)))
planned.append(region)
if planned:
params["planned_regions"] = planned
region_ids = tuple(item.region_id for item in regions)
requirements: list[EvidenceRequirement] = []
def add(operation: str, family: str, roles: tuple[str, ...], *, source: str, raw: str = "", optional: bool = False) -> None:
index = len(requirements)
blocked_regions = tuple(item.region_id for item in regions if item.status == "blocked")
reasons: list[str] = []
if operation not in available:
reasons.append(f"unavailable_operation:{operation}")
if operation == "region_resolution":
if not region_ids:
reasons.append("region_resolution_has_no_typed_regions")
elif blocked_regions:
reasons.append("blocked_region_dependencies:" + ",".join(blocked_regions))
if region_ids and operation in {"spatial_geometry", "route_topology", "route_movement_classification", "change_analysis", "crop"} and blocked_regions:
reasons.append("blocked_region_dependencies:" + ",".join(blocked_regions))
status = "optional" if optional else ("supported" if not reasons else "unsupported")
uses_regions = operation in {"spatial_geometry", "route_topology", "route_movement_classification", "change_analysis", "crop", "region_resolution"}
requirements.append(
EvidenceRequirement(
requirement_id=f"req_{index:02d}_{_slug(operation)}",
operation=operation,
tool_family=family,
image_roles=roles,
region_ids=region_ids if uses_regions else (),
depends_on=tuple(f"region:{region_id}" for region_id in region_ids) if uses_regions else (),
source=source,
raw=raw,
optional=optional,
status=status,
reasons=tuple(reasons),
)
)
if predicate_authoritative:
for operation in predicate.get("operations", []):
if not isinstance(operation, Mapping):
continue
add(
str(operation.get("kind") or ""),
str(operation.get("tool_family") or "predicate_interpreter"),
tuple(str(item) for item in operation.get("image_roles", []) if str(item)),
source="executable_predicate",
raw=str(operation.get("id") or ""),
optional=bool(operation.get("optional")),
)
requirements_tuple = _dedupe_requirements(requirements)
unsupported = [
item.requirement_id
for item in requirements_tuple
if not item.optional and item.status != "supported"
]
reasons = list(authoritative_conflicts)
reasons.extend(
reason
for item in requirements_tuple
if not item.optional and item.status != "supported"
for reason in item.reasons
)
return EvidenceProgram(
check_id=spec.check_id,
checker=spec.checker,
status="executable" if not unsupported and not authoritative_conflicts else "blocked",
checker_params=params,
promoted_parameters=promotion.promoted,
parameter_conflicts=authoritative_conflicts,
regions=regions,
requirements=requirements_tuple,
evidence_artifacts=(),
unsupported_requirements=tuple(unsupported),
reasons=tuple(dict.fromkeys(reasons)),
)
for operation, family, roles in _checker_requirements(spec, params):
add(operation, family, roles, source="checker_and_decision_rule")
raw_tools = params.get("tool_sequence", [])
raw_tools = raw_tools if isinstance(raw_tools, list) else [raw_tools]
for raw in raw_tools:
normalized = normalize_tool_operation(raw)
if normalized is None:
add("unknown_declared_operation", "unknown", (), source="tool_sequence", raw=str(raw))
else:
operation, family, optional = normalized
roles = ("input", "pred") if operation in {"image_registration", "change_analysis", "instance_matching", "text_alignment", "text_quality"} else ("pred",)
add(operation, family, roles, source="tool_sequence", raw=str(raw), optional=optional)
# Region proposal jobs are first-class requirements, not invisible setup.
for region in regions:
if region.status == "auto_resolvable":
add(region.resolver_operation, "region_proposer", region.image_roles, source="planned_regions", raw=region.region_id)
raw_artifacts = params.get("evidence_required", [])
raw_artifacts = raw_artifacts if isinstance(raw_artifacts, list) else [raw_artifacts]
artifact_families = {
"object_detection": "detector",
"mask_refinement": "segmenter",
"image_registration": "registration",
"instance_matching": "matcher",
"spatial_geometry": "checker_internal",
"route_topology": "route_parser",
"visual_attribute": "attribute_parser",
"pose_estimation": "human_pose_parser",
"local_correspondence": "local_correspondence",
"counting": "checker_internal",
"text_quality": "image_metrics",
"change_analysis": "change_analyzer",
"ocr": "ocr",
"region_resolution": "checker_internal",
}
# evidence_required is executable input, not documentation. Promote every
# typed producer it names into a first-class requirement so the runtime
# cannot omit a tool merely because tool_sequence forgot to repeat it.
for value in raw_artifacts:
description = str(value).strip()
if not description:
continue
lower = description.lower()
for operation in _artifact_operations(description):
if operation in {"image_registration", "instance_matching", "change_analysis", "text_quality"}:
roles = ("input", "pred")
elif "input" in lower and ("output" in lower or "pred" in lower):
roles = ("input", "pred")
else:
roles = ("pred",)
add(operation, artifact_families.get(operation, "unknown"), roles, source="evidence_required", raw=description)
requirements_tuple = _dedupe_requirements(requirements)
required_ops = {item.operation for item in requirements_tuple if not item.optional and item.status == "supported"}
if regions and all(item.status != "blocked" for item in regions):
required_ops.add("region_resolution")
artifacts: list[EvidenceArtifact] = []
for index, value in enumerate(raw_artifacts):
description = str(value).strip()
if not description:
continue
producer_ops = _artifact_operations(description)
missing = [op for op in producer_ops if op not in required_ops]
reasons = []
if not producer_ops:
reasons.append("evidence_artifact_has_no_typed_producer")
if missing:
reasons.append("missing_supported_producers:" + ",".join(missing))
artifacts.append(
EvidenceArtifact(
artifact_id=f"evidence_{index:02d}",
description=description,
producer_operations=producer_ops,
status="supported" if not reasons else "unsupported",
reasons=tuple(reasons),
)
)
unsupported = [item.requirement_id for item in requirements_tuple if not item.optional and item.status != "supported"]
unsupported.extend(item.artifact_id for item in artifacts if item.status != "supported")
reasons: list[str] = list(promotion.conflicts)
reasons.extend(reason for item in regions if item.status == "blocked" for reason in item.reasons)
reasons.extend(reason for item in requirements_tuple if not item.optional and item.status != "supported" for reason in item.reasons)
reasons.extend(reason for item in artifacts if item.status != "supported" for reason in item.reasons)
plan_unresolved = params.get("plan_unresolved", [])
if plan_unresolved:
reasons.append("resolved_plan_has_unresolved_facts")
unsupported.append("resolved_plan_unresolved")
if not requirements_tuple:
reasons.append("no_typed_evidence_requirements")
unsupported.append("no_requirements")
status = "executable" if not unsupported and not promotion.conflicts else "blocked"
return EvidenceProgram(
check_id=spec.check_id,
checker=spec.checker,
status=status,
checker_params=params,
promoted_parameters=promotion.promoted,
parameter_conflicts=promotion.conflicts,
regions=regions,
requirements=requirements_tuple,
evidence_artifacts=tuple(artifacts),
unsupported_requirements=tuple(dict.fromkeys(unsupported)),
reasons=tuple(dict.fromkeys(reasons)),
)