agent-apprenticeship 0.1.6 → 0.1.8

package/src/agent_apprenticeship_trace/public_sanitizer.py

@@ -19,6 +19,9 @@ PUBLIC_OMIT_KEYS={
     'source_url_or_ref','source_kind','source_url','source_ref','source_basis','source_license',
     'expected_pay','expected_apprentice_pay',
     'evaluation_mode','data_sharing_level','sensitive_info_masking',
+    'raw_llm_output','raw_model_output',
+    'limitations','known_limitations','limitations_count','when_not_to_use',
+    'not_useful_for','non_use_cases','rubric_limitations',
 }
 _ABS_LOCAL_PATH_RE=re.compile(r"(/Users/[^\s\"\']+|/home/[^\s\"\']+|/private/[^\s\"\']+|/tmp/[^\s\"\']+|/var/folders/[^\s\"\']+)")
 _USAGE_RE=re.compile(r"(you['’]?ve hit your usage limit|usage limit|rate limit|quota)", re.I)
@@ -38,6 +41,36 @@ PUBLIC_TEXT_REPLACEMENTS=[
     ("evaluation_mode", "mentor_mode"),
     ("worker_attempt", "apprentice_attempt"),
     ("worker_agent", "apprentice_agent"),
+    ("Gatekeeping", "Review Routing"),
+    ("gatekeeping", "review routing"),
+    ("Worker " + "Agent", "Apprentice Agent"),
+    ("worker " + "agent", "Apprentice Agent"),
+    ("Deep" + "Seek", "model provider"),
+    ("files_created_" + "list" + "ing_inconsistency", "files_created_index_inconsistency"),
+    ("Cur" + "ated", "Selected"),
+    ("cur" + "ated", "selected"),
+    ("Contribution " + "Bundle", "Experience Compilation"),
+    ("contribution " + "bundle", "Experience Compilation"),
+    ("Export Full " + "Training Package", "Export Full Experience Compilation"),
+    ("Training Data " + "Optimization", "Experience Compiler"),
+    ("T" + "DO", "Experience Compiler"),
+    ("Honest " + "limitations", "Generation notes"),
+    ("honest " + "limitations", "generation notes"),
+    ("Limitations", "Generation Notes"),
+    ("limitations", "generation notes"),
+    ("limitation", "generation note"),
+    ("Applic" + "ability", "Best Use"),
+    ("applic" + "ability", "fit"),
+    ("Required " + "Conditions", "Required Inputs"),
+    ("Evidence " + "Coverage", "Source Evidence"),
+    ("Evidence " + "coverage", "Source evidence"),
+    ("evidence " + "coverage", "source evidence"),
+    ("Coverage " + "Notes", "Generation Notes"),
+    ("Data Coverage " + "Notes", "Generation Notes"),
+    ("Transfer " + "Boundaries", "Transfer Notes"),
+    ("Missing Evidence " + "Notes", "Generation Notes"),
+    ("Quality " + "Notes", "Quality Signals"),
+    ("quality " + "notes", "quality signals"),
 ]
 
 def sha256_text(text: str | None) -> str | None:
@@ -89,7 +122,10 @@ def _sanitize_obj(obj: Any, prompt_text: str | None=None) -> Any:
     for k,v in obj.items():
         if k in CONTROLLER_TELEMETRY_KEYS or k in PUBLIC_OMIT_KEYS or k.startswith('stop_on_'):
             continue
-        out[k]=_sanitize_obj(v, prompt_text)
+        safe_key=sanitize_public_text(str(k), prompt_text) or str(k)
+        if safe_key in PUBLIC_OMIT_KEYS or safe_key.startswith('stop_on_'):
+            continue
+        out[safe_key]=_sanitize_obj(v, prompt_text)
     return out
 
 def sanitize_public_obj(obj: dict[str, Any], prompt_text: str | None=None) -> dict[str, Any]:

package/src/agent_apprenticeship_trace/rubric_generation.py

@@ -3,7 +3,7 @@ import json
 from pathlib import Path
 from pydantic import BaseModel
 from .schemas import TaskIntakeSpec, RubricItem, RubricSpec, RubricQualityReport
-from .config import get_settings
+from .config import apprentice_agent_display_name, get_settings
 from .io import read_json
 from .openai_structured import get_model_provider_status, run_structured_role
 from .public_sanitizer import sanitize_public_obj, sha256_text
@@ -22,14 +22,15 @@ def _mentor_provider_id() -> str:
 def deterministic_rubric(spec: TaskIntakeSpec) -> tuple[RubricSpec, RubricQualityReport]:
     reqs = spec.output_requirements or ['final deliverable']
     weight = 1.0/len(reqs)
-    items=[RubricItem(rubric_item_id=f'ri_{i+1}', criterion_name=r, criterion_description=f'Output satisfies {r}', weight=weight, score_min=0, score_max=1, pass_threshold=0.7, observable_evidence=[r], required_artifacts=[r], scoring_method='deterministic', worker_visible=True, verifier_only=False, hidden_reference_required=False, failure_modes=['missing','incorrect'], partial_credit_rules=['partial if substantially present'], edge_cases=[], anti_cheat_notes=[], metadata_json={}) for i,r in enumerate(reqs)]
-    rub=RubricSpec(rubric_id=f'rubric_{spec.task_id}', task_id=spec.task_id, task_family_id=None, rubric_version='v0.1', rubric_items=items, total_weight=1.0, pass_threshold=0.7, worker_visible_rubric_ref='rubric/worker_visible_rubric.md', verifier_private_rubric_ref='rubric/verifier_private_rubric.json', hidden_reference_policy='No hidden references for deterministic seed.', scoring_aggregation='weighted_sum', required_artifacts=reqs, disqualifying_errors=['secret leak'], partial_credit_allowed=True, grader_kind='deterministic', rubric_generation_source='deterministic_seed', rubric_generation_agent_provider=None, rubric_generation_agent_model=None, rubric_generation_confidence=0.4, metadata_json={'rubric_source':'deterministic_fallback','rubric_limitations':['LLM rubric generation unavailable'],'rubric_confidence':'low'})
-    qr=RubricQualityReport(rubric_id=rub.rubric_id, task_id=spec.task_id, criteria_count=len(items), total_weight=1.0, weights_sum_valid=True, has_observable_evidence=True, has_required_artifacts=True, has_partial_credit_rules=True, has_disqualifying_errors=True, has_hidden_reference_policy=True, has_worker_visible_view=True, has_verifier_private_view=True, ambiguous_criteria_count=0, unverifiable_criteria_count=0, rubric_quality_score=0.85, quality_flags=[], blockers=[], metadata_json={})
     settings=get_settings()
+    apprentice_name = apprentice_agent_display_name(settings)
+    items=[RubricItem(rubric_item_id=f'ri_{i+1}', criterion_name=r, criterion_description=f'Output satisfies {r}', weight=weight, score_min=0, score_max=1, pass_threshold=0.7, observable_evidence=[r], required_artifacts=[r], scoring_method='hybrid', worker_visible=True, verifier_only=False, hidden_reference_required=False, failure_modes=['missing','incorrect'], partial_credit_rules=['partial if substantially present'], edge_cases=[], anti_cheat_notes=[], metadata_json={'draft_author':'apprentice_agent','trace_quality_checks':['artifact_refs point to produced files','actual_outputs.json summarizes deliverables'],'environment_response_checks':['trace includes observed environment responses for meaningful tool/file actions']}) for i,r in enumerate(reqs)]
+    rub=RubricSpec(rubric_id=f'rubric_{spec.task_id}', task_id=spec.task_id, task_family_id=None, rubric_version='v0.1', rubric_items=items, total_weight=1.0, pass_threshold=0.7, worker_visible_rubric_ref='rubric/worker_visible_rubric.md', verifier_private_rubric_ref='rubric/verifier_private_rubric.json', hidden_reference_policy='No hidden references for this Apprentice-generated draft.', scoring_aggregation='weighted_sum', required_artifacts=reqs, disqualifying_errors=['secret leak'], partial_credit_allowed=True, grader_kind='hybrid', rubric_generation_source='agent_assisted', rubric_generation_agent_provider=None, rubric_generation_agent_model=None, rubric_generation_confidence=0.55, metadata_json={'rubric_source':'apprentice_agent_draft','draft_author':'apprentice_agent','apprentice_agent_id':settings.worker_agent,'apprentice_agent_name':apprentice_name,'mentor_audited':False,'mentor_audit_source':None,'audit_status':'unaudited','rubric_limitations':['Mentor audit has not completed for this rubric draft.'],'rubric_confidence':'medium','verifier_checks':reqs,'artifact_checks':reqs,'trace_quality_checks':['one meaningful action per trace step','artifact_refs and deliverable_refs are package-relative','actual_outputs.json is present'],'environment_response_checks':['environment_response_pairs can be extracted from observed trace output/state_change'],'reward_modeling_notes':['Use mentor/verifier/grader evidence when available; structural checks alone are not semantic quality labels.'],'process_supervision_notes':['Use trace step outcomes and evaluator feedback as labels when available.'],'revision_preference_notes':['Prefer revised attempts only when mentor/verifier/grader evidence supports improvement.'],'economic_value_checks':['Use expected economic value fields from task intake as context, not as outcome proof.']})
+    qr=RubricQualityReport(rubric_id=rub.rubric_id, task_id=spec.task_id, criteria_count=len(items), total_weight=1.0, weights_sum_valid=True, has_observable_evidence=True, has_required_artifacts=True, has_partial_credit_rules=True, has_disqualifying_errors=True, has_hidden_reference_policy=True, has_worker_visible_view=True, has_verifier_private_view=True, ambiguous_criteria_count=0, unverifiable_criteria_count=0, rubric_quality_score=0.8, quality_flags=['mentor_audit_pending'], blockers=[], metadata_json={'draft_author':'apprentice_agent','mentor_audited':False})
     if settings.rubric_mode in {'hybrid','llm_default'} and _mentor_provider_can_attempt() and settings.llm_rubric_generation_enabled:
         provider=_mentor_provider_id()
-        rub.grader_kind='hybrid'; rub.rubric_generation_agent_provider=provider; rub.rubric_generation_agent_model=settings.model_provider_model or settings.openai_model; rub.metadata_json.update({'llm_rubric_generation_enabled': True, 'llm_unavailable': True, 'rubric_source':'deterministic_fallback', 'provider': provider})
-        qr.quality_flags.append('deterministic_fallback_used')
+        rub.grader_kind='hybrid'; rub.rubric_generation_agent_provider=provider; rub.rubric_generation_agent_model=settings.model_provider_model or settings.openai_model; rub.metadata_json.update({'llm_rubric_generation_enabled': True, 'mentor_audit_status': 'not_completed', 'provider': provider})
+        qr.quality_flags.append('mentor_audit_not_completed')
     return rub, qr
 
 def deterministic_rubric_quality_check(rubric: RubricSpec) -> RubricQualityReport:
@@ -40,7 +41,7 @@ def _rubric_prompt(spec: TaskIntakeSpec) -> str:
     public_spec=sanitize_public_obj(spec.model_dump(mode='json'))
     return """Return only valid JSON. Do not include markdown. Do not add extra top-level fields; place extras under metadata_json.extra_model_fields.
 Required skeleton: {"rubric_spec":{"rubric_id":"rubric_<task_id>","task_id":"...","rubric_version":"0.1","rubric_items":[{"rubric_item_id":"ri_1","criterion_name":"...","criterion_description":"...","weight":1.0,"score_min":0,"score_max":1,"pass_threshold":0.7,"observable_evidence":[],"required_artifacts":[],"scoring_method":"llm_rubric_judge","worker_visible":true,"verifier_only":false,"hidden_reference_required":false,"failure_modes":[],"partial_credit_rules":[],"edge_cases":[],"anti_cheat_notes":[],"metadata_json":{}}],"total_weight":1.0,"pass_threshold":0.7,"worker_visible_rubric_ref":"rubric/worker_visible_rubric.md","verifier_private_rubric_ref":"rubric/verifier_private_rubric.json","hidden_reference_policy":"no_hidden_reference_available","scoring_aggregation":"weighted_sum","required_artifacts":[],"disqualifying_errors":[],"partial_credit_allowed":true,"grader_kind":"llm_rubric_judge","rubric_generation_source":"task_specific_agent_draft","metadata_json":{}},"rubric_quality_report":{"rubric_id":"...","task_id":"...","criteria_count":1,"total_weight":1.0,"weights_sum_valid":true,"has_observable_evidence":true,"has_required_artifacts":true,"has_partial_credit_rules":true,"has_disqualifying_errors":false,"has_hidden_reference_policy":true,"has_worker_visible_view":true,"has_verifier_private_view":true,"ambiguous_criteria_count":0,"unverifiable_criteria_count":0,"rubric_quality_score":0.75,"quality_flags":[],"blockers":[],"metadata_json":{}}}.
-Create a task-specific rubric with weighted rubric_items that sum to 1.0. Include explicit required_artifacts, observable_evidence, success criteria, failure modes, grader/verifier/evaluator guidance in metadata, limitations, and public-safe metadata. Do not leak hidden/reference answers to worker-visible fields.
+Audit and improve an Apprentice Agent-generated task rubric draft. Keep the Apprentice Agent as the draft author; your role is Mentor audit/improvement. Create a task-specific rubric with weighted rubric_items that sum to 1.0. Include explicit required_artifacts, observable_evidence, success criteria, failure modes, trace_quality_checks, environment_response_checks, reward_modeling_notes, process_supervision_notes, revision_preference_notes, economic_value_checks, and grader/verifier/evaluator guidance in metadata. Do not rely on Apprentice Agent self-evaluation as outcome evidence. Do not leak hidden/reference answers to Apprentice-visible fields. Do not include commerce metadata or off-brand selection language.
 TaskIntakeSpec JSON:
 """ + json.dumps(public_spec, sort_keys=True)
 
@@ -56,9 +57,9 @@ def generate_rubric(spec: TaskIntakeSpec, role_root: Path | None=None) -> tuple[
                 parsed=read_json(role_root/'rubric_agent/parsed_output.json')
                 rub=RubricSpec.model_validate(parsed['rubric_spec'])
                 qr=deterministic_rubric_quality_check(rub)
-                rub.grader_kind='hybrid'; rub.rubric_generation_agent_provider=rr.provider; rub.rubric_generation_agent_model=rr.model; rub.rubric_generation_confidence=rub.rubric_generation_confidence or 0.75
-                rub.metadata_json.update({'rubric_source':'llm','provider':rr.provider,'model':rr.model,'llm_prompt_ref_internal':str(role_root/'rubric_agent/prompt.md'),'llm_response_ref_internal':str(role_root/'rubric_agent/raw_output.txt'),'prompt_template_id':'rubric_generation_agent_v0','prompt_template_version':'0.1','prompt_hash':sha256_text(prompt),'public_response_summary':'Model-generated rubric passed deterministic quality checking.'})
-                qr.metadata_json.update({'rubric_source':'llm','role_result_ref_internal':str(role_root/'rubric_agent/role_result.json')})
+                rub.grader_kind='hybrid'; rub.rubric_generation_source='agent_assisted'; rub.rubric_generation_agent_provider=rr.provider; rub.rubric_generation_agent_model=rr.model; rub.rubric_generation_confidence=rub.rubric_generation_confidence or 0.75
+                rub.metadata_json.update({'rubric_source':'apprentice_agent_draft','draft_author':'apprentice_agent','mentor_audited':True,'mentor_audit_source':'mentor_model','audit_status':'mentor_audited','provider':rr.provider,'model':rr.model,'llm_prompt_ref_internal':str(role_root/'rubric_agent/prompt.md'),'llm_response_ref_internal':str(role_root/'rubric_agent/raw_output.txt'),'prompt_template_id':'rubric_generation_agent_v0','prompt_template_version':'0.1','prompt_hash':sha256_text(prompt),'public_response_summary':'Mentor Model audited and improved an Apprentice-generated rubric draft.'})
+                qr.metadata_json.update({'draft_author':'apprentice_agent','mentor_audited':True,'mentor_audit_source':'mentor_model','role_result_ref_internal':str(role_root/'rubric_agent/role_result.json')})
                 return rub, qr
             if settings.rubric_mode == 'llm_required' or settings.llm_fail_closed:
                 raise RuntimeError(rr.error_message or 'Model rubric generation failed')

package/src/agent_apprenticeship_trace/schemas.py

@@ -3,6 +3,64 @@ from typing import Literal, Any
 from pydantic import BaseModel, ConfigDict, Field, model_validator
 
 DictAny = dict[str, Any]
+EnvironmentDomain = Literal[
+    "terminal",
+    "browser",
+    "computer_use",
+    "web",
+    "search",
+    "codebase",
+    "software",
+    "os",
+    "android",
+    "mcp",
+    "api",
+    "database",
+    "spreadsheet",
+    "file",
+    "multimodal",
+    "device",
+    "robotics",
+    "lab",
+    "industrial",
+    "physical",
+    "mixed",
+    "unknown",
+]
+InteractionSurface = Literal[
+    "api",
+    "mcp_tool",
+    "mcp_resource",
+    "mcp_prompt",
+    "browser",
+    "computer_use",
+    "cli",
+    "file",
+    "codebase",
+    "database",
+    "spreadsheet",
+    "search",
+    "cloud_console",
+    "saas_app",
+    "async_job",
+    "queue",
+    "webhook",
+    "scheduler",
+    "messaging",
+    "email",
+    "collaboration",
+    "document",
+    "pdf",
+    "table",
+    "multimodal",
+    "device",
+    "industrial",
+    "physical",
+    "mixed",
+    "other_tool",
+    "other_surface",
+    "unknown",
+]
 
 class StrictModel(BaseModel):
     model_config = ConfigDict(extra="forbid")
@@ -30,6 +88,7 @@ class ActualOutputs(StrictModel):
     status: Literal["success","partial","failed","timeout","error"]
     output_summary: str
     primary_output_ref: str | None = None
+    input_artifact_refs: list[str] = Field(default_factory=list)
     deliverable_refs: list[str] = Field(default_factory=list)
     final_message_ref: str | None = None
     artifact_refs: list[str] = Field(default_factory=list)
@@ -50,6 +109,7 @@ class AgentTraceStep(StrictModel):
     action: Literal["user_message","agent_step","output","error"]
     operation: Literal["plan","analyze","search","read","write","edit","execute","verify","download","install","ask_user","answer","select","grade","evaluate","revise","other"] | None = None
     tool: str | None = None
+    environment_domain: EnvironmentDomain | None = None
     execution_mode: Literal["serial","parallel"] | None = None
     parallel_group: str | None = None
     observation: str | None = None
@@ -73,6 +133,252 @@ class AgentTraceStep(StrictModel):
     ended_at: str | None = None
     retry_of: int | None = None
     artifact_refs: list[str] = Field(default_factory=list)
+    actor_kind: Literal["user","apprentice_agent","mentor_model","verifier","grader","evaluator","subagent","tool","system","expert","unknown"] | None = None
+    actor_role: str | None = None
+    agent_id: str | None = None
+    agent_name: str | None = None
+    subagent_id: str | None = None
+    subagent_name: str | None = None
+    subagent_role: str | None = None
+    parent_agent_id: str | None = None
+    delegation_id: str | None = None
+    delegation_source_step: int | None = None
+    handoff_reason: str | None = None
+    handoff_payload_ref: str | None = None
+    handoff_output_refs: list[str] = Field(default_factory=list)
+    subagent_output_refs: list[str] = Field(default_factory=list)
+    coordination_pattern: Literal["sequential","parallel","debate","planner_worker","reviewer_reviser","swarm","unknown"] | None = None
+    communication_channel: Literal["trace","file","message","tool_call","api","shared_workspace","unknown"] | None = None
+    depends_on_agent_outputs: list[str] = Field(default_factory=list)
+    modality: Literal["text","image","screenshot","audio","video","pdf","spreadsheet","table","diagram","mixed","unknown"] | None = None
+    modalities: list[str] = Field(default_factory=list)
+    media_type: str | None = None
+    mime_type: str | None = None
+    artifact_role: Literal["input","output","evidence","intermediate","preview","transcript","ocr","annotation","verification"] | None = None
+    artifact_hash: str | None = None
+    artifact_preview_ref: str | None = None
+    ocr_text_ref: str | None = None
+    transcript_ref: str | None = None
+    annotation_ref: str | None = None
+    bbox_refs: list[str] = Field(default_factory=list)
+    frame_refs: list[str] = Field(default_factory=list)
+    timecode_refs: list[str] = Field(default_factory=list)
+    table_refs: list[str] = Field(default_factory=list)
+    sheet_refs: list[str] = Field(default_factory=list)
+    derived_text_summary: str | None = None
+    interaction_surface: InteractionSurface | None = None
+    surface_type: str | None = None
+    connected_surfaces: list[str] = Field(default_factory=list)
+    resource_id: str | None = None
+    resource_type: str | None = None
+    object_id: str | None = None
+    object_type: str | None = None
+    request_id: str | None = None
+    operation_id: str | None = None
+    job_id: str | None = None
+    event_id: str | None = None
+    session_id: str | None = None
+    transaction_id: str | None = None
+    correlation_id: str | None = None
+    idempotency_key: str | None = None
+    status_before: str | None = None
+    status_after: str | None = None
+    state_before_ref: str | None = None
+    state_after_ref: str | None = None
+    raw_request_ref: str | None = None
+    raw_response_ref: str | None = None
+    error_code: str | None = None
+    evidence_refs: list[str] = Field(default_factory=list)
+    verification_refs: list[str] = Field(default_factory=list)
+    redaction_notes: list[str] = Field(default_factory=list)
+    api_provider: str | None = None
+    api_service: str | None = None
+    api_endpoint: str | None = None
+    api_method: str | None = None
+    api_version: str | None = None
+    http_status: int | None = None
+    request_payload_ref: str | None = None
+    response_payload_ref: str | None = None
+    request_schema_ref: str | None = None
+    response_schema_ref: str | None = None
+    auth_mode: str | None = None
+    rate_limit_state: str | None = None
+    pagination_state: str | None = None
+    retry_count: int | None = None
+    mcp_server: str | None = None
+    mcp_server_version: str | None = None
+    mcp_capability_type: Literal["tool","resource","prompt"] | None = None
+    mcp_tool_name: str | None = None
+    mcp_resource_uri: str | None = None
+    mcp_prompt_name: str | None = None
+    mcp_arguments_ref: str | None = None
+    mcp_result_ref: str | None = None
+    mcp_schema_ref: str | None = None
+    mcp_error_code: str | None = None
+    mcp_error_message: str | None = None
+    mcp_call_id: str | None = None
+    database_type: str | None = None
+    database_name: str | None = None
+    schema_name: str | None = None
+    table_name: str | None = None
+    query_type: str | None = None
+    query_ref: str | None = None
+    query_hash: str | None = None
+    row_count: int | None = None
+    affected_row_count: int | None = None
+    isolation_level: str | None = None
+    migration_id: str | None = None
+    explain_plan_ref: str | None = None
+    before_snapshot_ref: str | None = None
+    after_snapshot_ref: str | None = None
+    rollback_ref: str | None = None
+    db_error_code: str | None = None
+    db_error_message: str | None = None
+    workbook_ref: str | None = None
+    sheet_name: str | None = None
+    table_ref: str | None = None
+    cell_range: str | None = None
+    column_count: int | None = None
+    formula_refs: list[str] = Field(default_factory=list)
+    pivot_refs: list[str] = Field(default_factory=list)
+    chart_refs: list[str] = Field(default_factory=list)
+    data_validation_refs: list[str] = Field(default_factory=list)
+    before_table_ref: str | None = None
+    after_table_ref: str | None = None
+    diff_ref: str | None = None
+    calculation_mode: str | None = None
+    spreadsheet_error: str | None = None
+    command: str | None = None
+    command_ref: str | None = None
+    cwd_ref: str | None = None
+    exit_code: int | None = None
+    stdout_ref: str | None = None
+    stderr_ref: str | None = None
+    file_path: str | None = None
+    file_paths: list[str] = Field(default_factory=list)
+    file_operation: str | None = None
+    patch_ref: str | None = None
+    git_branch: str | None = None
+    git_commit_before: str | None = None
+    git_commit_after: str | None = None
+    test_command: str | None = None
+    test_result_ref: str | None = None
+    build_result_ref: str | None = None
+    lint_result_ref: str | None = None
+    dependency_changes_ref: str | None = None
+    process_id: str | None = None
+    timeout_seconds: int | None = None
+    search_provider: str | None = None
+    query: str | None = None
+    query_ref: str | None = None
+    search_scope: str | None = None
+    filters: DictAny | None = None
+    result_count: int | None = None
+    selected_result_refs: list[str] = Field(default_factory=list)
+    citation_refs: list[str] = Field(default_factory=list)
+    ranking_signal: str | None = None
+    search_time_range: str | None = None
+    search_error: str | None = None
+    ui_surface: Literal["browser","desktop_app","mobile_app","terminal","ide","cloud_console","file_manager","spreadsheet_app","unknown"] | None = None
+    app_name: str | None = None
+    app_version: str | None = None
+    os_name: str | None = None
+    device_type: str | None = None
+    window_title: str | None = None
+    page_title: str | None = None
+    url_or_resource_id: str | None = None
+    visible_state_ref: str | None = None
+    screenshot_ref: str | None = None
+    dom_snapshot_ref: str | None = None
+    accessibility_tree_ref: str | None = None
+    element_selector: str | None = None
+    coordinates: str | None = None
+    action_type: str | None = None
+    input_text: str | None = None
+    before_state_ref: str | None = None
+    after_state_ref: str | None = None
+    network_ref: str | None = None
+    api_call_ref: str | None = None
+    software_state_ref: str | None = None
+    service_name: str | None = None
+    workspace_id: str | None = None
+    project_id: str | None = None
+    tenant_id: str | None = None
+    record_id: str | None = None
+    ticket_id: str | None = None
+    object_state_before_ref: str | None = None
+    object_state_after_ref: str | None = None
+    permission_context: str | None = None
+    region: str | None = None
+    environment_name: str | None = None
+    deployment_id: str | None = None
+    audit_log_ref: str | None = None
+    change_request_ref: str | None = None
+    queue_name: str | None = None
+    message_id: str | None = None
+    webhook_id: str | None = None
+    callback_url_ref: str | None = None
+    schedule_id: str | None = None
+    trigger_type: str | None = None
+    enqueue_time: str | None = None
+    start_time: str | None = None
+    completion_time: str | None = None
+    delivery_status: str | None = None
+    polling_status: str | None = None
+    callback_status: str | None = None
+    job_status_before: str | None = None
+    job_status_after: str | None = None
+    async_result_ref: str | None = None
+    dead_letter_ref: str | None = None
+    platform_name: str | None = None
+    channel_id: str | None = None
+    thread_id: str | None = None
+    sender_role: str | None = None
+    recipient_role: str | None = None
+    message_type: str | None = None
+    message_ref: str | None = None
+    attachment_refs: list[str] = Field(default_factory=list)
+    decision_refs: list[str] = Field(default_factory=list)
+    approval_state: str | None = None
+    followup_required: bool | None = None
+    document_ref: str | None = None
+    document_type: str | None = None
+    page_refs: list[str] = Field(default_factory=list)
+    section_refs: list[str] = Field(default_factory=list)
+    text_extract_ref: str | None = None
+    table_extract_refs: list[str] = Field(default_factory=list)
+    figure_refs: list[str] = Field(default_factory=list)
+    annotation_refs: list[str] = Field(default_factory=list)
+    summary_ref: str | None = None
+    document_diff_ref: str | None = None
+    physical_environment_type: str | None = None
+    device_id: str | None = None
+    device_model: str | None = None
+    sensor_readings_ref: str | None = None
+    actuator_command: str | None = None
+    control_mode: str | None = None
+    physical_state_before_ref: str | None = None
+    physical_state_after_ref: str | None = None
+    telemetry_ref: str | None = None
+    calibration_ref: str | None = None
+    safety_state: Literal["normal","warning","fault","emergency_stop","unknown"] | None = None
+    safety_limit_refs: list[str] = Field(default_factory=list)
+    operator_override: bool | None = None
+    human_supervisor_present: bool | None = None
+    simulation_or_real_world: Literal["simulation","real_world","hybrid","unknown"] | None = None
+    latency_ms: float | None = None
+    tolerance: str | None = None
+    rollback_action: str | None = None
+    emergency_stop: bool | None = None
+    compliance_refs: list[str] = Field(default_factory=list)
+    cross_system_dependencies: list[str] = Field(default_factory=list)
+    handoff_refs: list[str] = Field(default_factory=list)
+    state_sync_refs: list[str] = Field(default_factory=list)
+    consistency_checks: list[str] = Field(default_factory=list)
+    system_of_record: str | None = None
+    surface_capture_status: Literal["complete","partial","unavailable","not_applicable"] | None = None
+    surface_capture_notes: str | None = None
+    missing_surface_fields: list[str] = Field(default_factory=list)
     metadata_json: DictAny = Field(default_factory=dict)
 
     @model_validator(mode="after")
@@ -91,6 +397,9 @@ class AgentTrace(StrictModel):
     schema_version: str = "aa-trace-v0.1"
     trace_id: str
     collection_id: str | None = None
+    run_id: str | None = None
+    package_id: str | None = None
+    bundle_id: str | None = None
     prior_trace_id: str | None = None
     trace_mode: Literal["live","retraced","hybrid"]
     task: str
@@ -107,6 +416,13 @@ class AgentTrace(StrictModel):
     skills: list[str] | None = None
     memory: str | None = None
     agent_config: DictAny | None = None
+    environment_domain: EnvironmentDomain | None = None
+    actual_outputs_ref: str | None = None
+    input_artifact_refs: list[str] = Field(default_factory=list)
+    artifact_refs: list[str] = Field(default_factory=list)
+    deliverable_refs: list[str] = Field(default_factory=list)
+    output_summary: str | None = None
+    final_output_summary: str | None = None
     learning: str | None = None
     termination_reason: Literal["task_complete","verifier_passed","verifier_failed","max_iterations_reached","agent_blocked","timeout","error_unrecoverable","partial_then_stopped","provider_usage_limit","other"]
     steps: list[AgentTraceStep]
@@ -227,7 +543,7 @@ class TaskIntakeQualityReport(StrictModel):
 class RubricItem(StrictModel):
     rubric_item_id: str; criterion_name: str; criterion_description: str; weight: float; score_min: float; score_max: float; pass_threshold: float
     observable_evidence: list[str]; required_artifacts: list[str]
-    scoring_method: Literal["llm_rubric_judge","deterministic","schema_match","regex","unit_test","hybrid","human_future"]
+    scoring_method: Literal["llm_rubric_judge","deterministic","schema_match","regex","unit_test","hybrid","human_future","structural_guardrail"]
     worker_visible: bool; verifier_only: bool; hidden_reference_required: bool
     failure_modes: list[str] = Field(default_factory=list); partial_credit_rules: list[str] = Field(default_factory=list); edge_cases: list[str] = Field(default_factory=list); anti_cheat_notes: list[str] = Field(default_factory=list); metadata_json: DictAny = Field(default_factory=dict)
 
@@ -236,7 +552,7 @@ class RubricSpec(StrictModel):
     worker_visible_rubric_ref: str; verifier_private_rubric_ref: str; hidden_reference_policy: str
     scoring_aggregation: Literal["weighted_sum","sum","all_required","custom"]
     required_artifacts: list[str]; disqualifying_errors: list[str] = Field(default_factory=list); partial_credit_allowed: bool
-    grader_kind: Literal["llm_rubric_judge","deterministic","hybrid","human_future"]
+    grader_kind: Literal["llm_rubric_judge","deterministic","hybrid","human_future","structural_guardrail","unavailable"]
     rubric_generation_source: Literal["agent_assisted","family_template","task_specific_agent_draft","expert_override","deterministic_seed"]
     rubric_generation_agent_provider: str | None = None; rubric_generation_agent_model: str | None = None; rubric_generation_confidence: float | None = None; metadata_json: DictAny = Field(default_factory=dict)
     @model_validator(mode="after")
@@ -252,7 +568,7 @@ class RubricQualityReport(StrictModel):
 class RubricItemScore(StrictModel):
     rubric_item_id: str; criterion_name: str; score: float; max_score: float; passed: bool; evidence_refs: list[str] = Field(default_factory=list); failure_mode: str | None = None; notes: str | None = None; confidence: float | None = None; artifact_presence_ok: bool | None = None; semantic_correctness_score: float | None = None; reasoning_summary: str | None = None; improvement_suggestion: str | None = None
 class GraderResult(StrictModel):
-    grader_result_id: str; task_id: str; attempt_id: str; attempt_kind: str; rubric_id: str; grader_kind: Literal["llm_rubric_judge","llm","model","deterministic","hybrid","human_future"]; score_source: Literal["llm","llm_semantic","model_judged","deterministic","deterministic_artifact_contract","deterministic_fallback","hybrid","human_future"]; score: float; max_score: float; passed: bool; rubric_item_scores: list[RubricItemScore]; failed_criteria: list[str]; passed_criteria: list[str]; evidence_refs: list[str]; confidence: float; reasoning_summary: str | None = None; limitations: list[str] = Field(default_factory=list); hidden_reference_used: bool; hidden_reference_leaked: bool; artifact_contract_score: float | None = None; semantic_score: float | None = None; model_score: float | None = None; legacy_semantic_score: float | None = None; legacy_score_source: str | None = None; final_score: float | None = None; model: str | None = None; provider: str | None = None; deterministic_precheck_ref: str | None = None; llm_prompt_ref_internal: str | None = None; llm_response_ref_internal: str | None = None; public_prompt_hash: str | None = None; public_response_summary: str | None = None; score_reliability: Literal["verified","unverified","needs_review","failed_verification"] | None = None; verifier_status: str | None = None; verifier_confidence: float | None = None; verifier_issue_count: int | None = None; verifier_issues_summary: str | None = None; metadata_json: DictAny = Field(default_factory=dict)
+    grader_result_id: str; task_id: str; attempt_id: str; attempt_kind: str; rubric_id: str; grader_kind: Literal["llm_rubric_judge","llm","model","deterministic","hybrid","human_future","structural_guardrail","unavailable"]; score_source: Literal["llm","llm_semantic","model_judged","deterministic","deterministic_artifact_contract","deterministic_fallback","hybrid","human_future","structural_guardrail","unavailable"]; score: float; max_score: float; passed: bool; rubric_item_scores: list[RubricItemScore]; failed_criteria: list[str]; passed_criteria: list[str]; evidence_refs: list[str]; confidence: float; reasoning_summary: str | None = None; limitations: list[str] = Field(default_factory=list); hidden_reference_used: bool; hidden_reference_leaked: bool; artifact_contract_score: float | None = None; semantic_score: float | None = None; model_score: float | None = None; legacy_semantic_score: float | None = None; legacy_score_source: str | None = None; final_score: float | None = None; model: str | None = None; provider: str | None = None; deterministic_precheck_ref: str | None = None; llm_prompt_ref_internal: str | None = None; llm_response_ref_internal: str | None = None; public_prompt_hash: str | None = None; public_response_summary: str | None = None; score_reliability: Literal["verified","unverified","needs_review","failed_verification"] | None = None; verifier_status: str | None = None; verifier_confidence: float | None = None; verifier_issue_count: int | None = None; verifier_issues_summary: str | None = None; metadata_json: DictAny = Field(default_factory=dict)
 class VerifierResult(StrictModel):
     verifier_result_id: str; task_id: str; attempt_id: str; attempt_kind: str; grader_result_id: str | None = None; verification_status: Literal["verified","partially_verified","failed","not_run"]; artifact_contract_ok: bool; evidence_grounding_ok: bool; score_consistency_ok: bool; hidden_reference_leaked: bool; issues: list[str] = Field(default_factory=list); confidence: float; verifier_notes: str | None = None; semantic_evidence_grounding_ok: bool | None = None; unsupported_claims: list[str] = Field(default_factory=list); leakage_check_ok: bool | None = None; model: str | None = None; provider: str | None = None; metadata_json: DictAny = Field(default_factory=dict)
 class EvaluatorFeedback(StrictModel):
@@ -271,3 +587,62 @@ class RewardModelingExample(StrictModel):
     example_id: str; task_id: str; attempt_id: str; rubric_ref: str; output_refs: list[str]; attempt_summary: str; rubric_item_scores: list[RubricItemScore]; final_score: float; passed: bool; failure_modes: list[str]; grader_notes: str | None = None; evidence_refs: list[str]; score_source: str; grader_kind: str; confidence: float; score_reliability: str | None = None; verifier_status: str | None = None; verifier_confidence: float | None = None; verifier_issue_count: int | None = None; verifier_issues_summary: str | None = None; metadata_json: DictAny=Field(default_factory=dict)
 class RevisionPreferencePair(StrictModel):
     pair_id: str; task_id: str; rubric_ref: str; baseline_attempt_ref: str; revised_attempt_ref: str; chosen_attempt_id: str; rejected_attempt_id: str; baseline_score: float; revised_score: float; score_delta: float; criteria_improved: list[str]; criteria_regressed: list[str]; preference_reason: str; score_source: str; grader_kind: str; confidence: float; metadata_json: DictAny=Field(default_factory=dict)
+
+class TDOJudgeResult(StrictModel):
+    tdo_round: int
+    tdo_judge_source: Literal["mentor_model", "apprentice_self_judge"]
+    mentor_audited: bool
+    verdict: Literal["accept", "improve", "reject"]
+    weak_pattern: str | None = None
+    strong_pattern: str | None = None
+    gap_interpretation: str | None = None
+    rubric_concerns: list[str] = Field(default_factory=list)
+    grounding_concerns: list[str] = Field(default_factory=list)
+    data_quality_concerns: list[str] = Field(default_factory=list)
+    grpo_suitability: Literal["high", "medium", "low", "unknown"]
+    learning_signal_quality: Literal["high", "medium", "low", "unknown"]
+    reuse_value: Literal["high", "medium", "low"]
+    quality_score: float
+    fidelity_score: float
+    difficulty_score: float
+    learnability_score: float
+    reuse_value_score: float
+    verdict_reason: str
+    suggestion_for_generator: str | None = None
+    metadata_json: DictAny = Field(default_factory=dict)
+
+    @model_validator(mode="after")
+    def validate_suggestion(self):
+        if self.verdict in {"improve", "reject"} and not self.suggestion_for_generator:
+            raise ValueError("suggestion_for_generator is required on improve/reject")
+        return self
+
+class TDOReport(StrictModel):
+    tdo_id: str
+    source_package_id: str
+    source_task_id: str | None = None
+    created_at: str
+    generator_mode: str
+    tdo_judge_source: Literal["mentor_model", "apprentice_self_judge"]
+    mentor_audited: bool
+    rounds_run: int
+    stop_reason: str
+    tdo_status: Literal["completed", "failed", "disabled", "partial"]
+    rows_generated_by_type: DictAny = Field(default_factory=dict)
+    rows_accepted_by_type: DictAny = Field(default_factory=dict)
+    rows_rejected_by_type: DictAny = Field(default_factory=dict)
+    average_quality_score: float | None = None
+    average_fidelity_score: float | None = None
+    average_difficulty_score: float | None = None
+    average_learnability_score: float | None = None
+    average_reuse_value_score: float | None = None
+    schema_validation_passed: bool
+    sanitization_passed: bool
+    weak_strong_proxy_available: bool
+    average_solver_gap_proxy: float | None = None
+    grpo_suitability_distribution: DictAny = Field(default_factory=dict)
+    omitted_outputs: list[str] = Field(default_factory=list)
+    not_generated_counts: DictAny = Field(default_factory=dict)
+    generation_notes: list[str] = Field(default_factory=list)
+    recommended_uses: list[str] = Field(default_factory=list)
+    metadata_json: DictAny = Field(default_factory=dict)