Datasets:

sxcn
/

HERA-Bench

"""Content scoring with the evaluated public alias/search corrections."""
from copy import deepcopy
import hashlib
import json
from common import read
from evaluate import Episode

def require(condition, message):
    if not condition:
        raise ValueError(message)

def encode(name):
    result = name.replace(".", "__")
    if len(result) <= 64:
        return result
    suffix = "__h" + hashlib.sha256(name.encode()).hexdigest()[:12]
    return result[:64-len(suffix)] + suffix

def canonical(value):
    return json.dumps(value, sort_keys=True, ensure_ascii=False, allow_nan=False, separators=(',', ':'))

def decode_observed_tools(answer, env, public_names, execution_log, encode):
    """Pure, narrow decoding; argument/error/effect authentication stays in oracle."""
    require(type(public_names) is list and all(type(n) is str and n for n in public_names),
            'Invalid public tool inventory')
    require(len(set(public_names)) == len(public_names), 'Duplicate public tool names')
    qualified = {env + '.' + name: name for name in public_names}
    # Check every declared tool, including tools the agent never chose to call.
    all_aliases = {}
    for full, local in qualified.items():
        for alias in (local, full, encode(full)):
            require(alias not in all_aliases or all_aliases[alias] == local, 'Ambiguous public tool alias')
            all_aliases[alias] = local
    require(type(execution_log) is list, 'Missing official execution log')
    observed = set()
    for event in execution_log:
        require(type(event) is dict and type(event.get('tool')) is str, 'Malformed tool event')
        name = event['tool']
        if name in ('task_documents.list_documents', 'task_documents.read_document'):
            continue
        require(name in qualified, 'Cross-environment or nonpublic observed tool')
        observed.add(qualified[name])
    aliases = {alias: local for alias, local in all_aliases.items() if local in observed}
    result = deepcopy(answer)
    witnesses = result['missing_evidence']
    require(type(witnesses) is list, 'Malformed missing_evidence')
    changes = []
    for index, witness in enumerate(witnesses):
        # Other witness schemas/fields are left to the frozen schema and oracle.
        if not isinstance(witness, dict) or 'tool' not in witness:
            continue
        name = witness['tool']
        require(type(name) is str and name in aliases, 'Unknown, cross-environment, or unobserved witness tool')
        local = aliases[name]
        if local != name:
            witness['tool'] = local
            changes.append({'json_pointer': f'/missing_evidence/{index}/tool', 'original': name,
                            'normalized': local, 'observed_qualified_name': env + '.' + local})
    restored = deepcopy(result)
    for change in changes:
        restored['missing_evidence'][int(change['json_pointer'].split('/')[2])]['tool'] = change['original']
    require(canonical(restored) == canonical(answer), 'Decoding modified a non-tool field')
    return result, changes, {'observed_qualified_tools': sorted(env + '.' + n for n in observed),
                            'allowed_alias_to_domain_name': aliases}

def complete_public_coverage(episode, events):
    required = {r['record_id'] for r in episode.initial['records'] if episode.bundle.oracle._search_match(r)}
    for event in events:
        if event['tool'] != 'search_catalog_records':
            continue
        args, obs = event['arguments'], event['observation']
        if not set(args).issubset({'query', 'theme', 'reusability', 'media', 'sort'}):
            continue
        if str(args.get('query', '')).strip().lower() not in {'harbor', '*'}:
            continue
        if any(k in args and args[k] != v for k, v in {'theme': 'art', 'reusability': 'open', 'media': True}.items()):
            continue
        if obs.get('ok') is not True:
            continue
        found = obs.get('result', {})
        items = found.get('items', [])
        ids = {x.get('record_id') for x in items}
        if (type(found.get('total_results')) is int and found['total_results'] == len(items) == len(ids)
                and required.issubset(ids)):
            return {'search_arguments': deepcopy(args), 'observed_ids': sorted(ids),
                    'required_ids': sorted(required), 'complete_superset': True}
    return None

def organization_search_evidence(episode, events):
    oracle = episode.bundle.oracle
    derived, problem = oracle._derive(episode.initial)
    if problem:
        return None
    org = derived['org']
    observed_identity = False
    required = {d['id']: d for d in derived['searched']}
    fields = ('id', 'name', 'owner_org', 'state', 'series_key', 'edition',
              'coverage_start', 'coverage_end', 'metadata_modified')
    for event in events:
        args, obs = event['arguments'], event['observation']
        if obs.get('ok') is not True:
            continue
        result = obs.get('result')
        if event['tool'] == 'organization_autocomplete' and isinstance(result, list):
            exact = [r for r in result if isinstance(r, dict) and r.get('title') == oracle.TARGET_TITLE]
            if len(exact) == 1 and oracle._has_fields(exact[0], org, ('id', 'name', 'title')):
                observed_identity = True
        if event['tool'] != 'package_search' or not observed_identity:
            continue
        # Accept only the two real, observed identifiers; never guess an alias.
        if (set(args) != {'owner_org', 'tag', 'include_private'}
                or args['owner_org'] not in (org['id'], org['name'])
                or args['tag'] != oracle.TAG or args['include_private'] is not False
                or not isinstance(result, dict)):
            continue
        records, count = result.get('results'), result.get('count')
        if not isinstance(records, list) or not all(isinstance(r, dict) for r in records):
            continue
        ids = [r.get('id') for r in records]
        if (not all(isinstance(i, str) for i in ids) or type(count) is not int
                or count != len(records) or len(set(ids)) != count or set(ids) != set(required)):
            continue
        if not all(oracle._has_fields(next(r for r in records if r['id'] == key), value, fields)
                   for key, value in required.items()):
            continue
        # The sealed oracle counts metadata carried by an accepted search too.
        # Recheck its exact fields/resource-collection predicate on this search,
        # so use of a name cannot erase valid metadata that was already read.
        selected_metadata = all(any(
            oracle._has_fields(record, ds, fields)
            and oracle._resource_collection_exact(record.get('resources'),
                [r for r in derived['all_resources'] if r.get('package_id') == ds['id']])
            for record in records) for ds, _ in derived['pairs'])
        return {'observed_organization': {k: org[k] for k in ('id', 'name', 'title')},
                'search_arguments': deepcopy(args), 'complete_observed_dataset_ids': sorted(ids),
                'required_dataset_ids': sorted(required), 'complete_same_population': True,
                'selected_dataset_and_resource_metadata_complete': selected_metadata}
    return None


def parse_answer(text):
    """Parse exactly one JSON report, rejecting duplicate keys and nonfinite values."""
    def unique(items):
        result = {}
        for key, value in items:
            require(key not in result, 'Duplicate JSON key: ' + key)
            result[key] = value
        return result
    def invalid(value):
        raise ValueError('Nonfinite JSON value: ' + value)
    return json.loads(text, object_pairs_hook=unique, parse_constant=invalid)


def grade_submission(root, number, side, submission):
    """Grade a controller-recorded report, trace and state using the released rules.
    `events` use domain-local tool names. `final_state` is the domain state,
    not a model's assertion about whether the task succeeded.
    """
    episode = Episode(root, number, side)
    try:
        require(isinstance(submission, dict), 'Submission must be an object')
        require(('answer' in submission) != ('final_output' in submission),
                'Provide exactly one of answer or final_output')
        answer = (parse_answer(submission['final_output']) if 'final_output' in submission
                  else parse_answer(canonical(submission['answer'])))
    except (ValueError, TypeError, KeyError) as exc:
        return {'passed': False, 'format_passed': False, 'content_passed': None,
                'error_kind': 'report_format', 'errors': [str(exc)]}
    try:
        events, final_state = submission['events'], submission['final_state']
        require(isinstance(events, list), 'Missing controller event list')
        require(all(isinstance(e, dict) and set(e) == {'tool', 'arguments', 'observation'}
                    and isinstance(e['tool'], str) and isinstance(e['arguments'], dict)
                    for e in events), 'Invalid controller event format')
    except (ValueError, TypeError, KeyError) as exc:
        return {'passed': False, 'format_passed': None, 'content_passed': False,
                'error_kind': 'invalid_trace', 'errors': [str(exc)]}

    score = episode.grade(answer, events, final_state)
    if score.get('format_passed') is False:
        return score
    info = episode.info
    if isinstance(answer, dict) and answer.get('missing_evidence') and info['kind'] == 'bundle':
        try:
            tools = read(episode.root / 'private/bundles' / info['slot'] / 'public/tools.json')
            log = [{'tool': info['env_name'] + '.' + event['tool']} for event in events]
            normalized, changes, _ = decode_observed_tools(
                answer, info['env_name'], [tool['name'] for tool in tools], log, encode)
            if changes:
                score = episode.grade(normalized, events, final_state)
        except (ValueError, RuntimeError, KeyError, TypeError) as exc:
            return dict(score, passed=False, content_passed=False,
                        normalization_error={'type': type(exc).__name__, 'message': str(exc)[:600]})

    catalog_failure = 'missing complete evidence for the prescribed search result'
    identifier_failure = 'missing complete public tagged organization dataset search evidence'
    metadata_failure = 'missing complete chosen-dataset and resource metadata evidence'
    failure = {'europeana': catalog_failure, 'ckan': identifier_failure}.get(info['slot'])
    if failure is None or failure not in score.get('failures', []):
        return score
    # These narrow equivalences apply only after every tool observation and
    # the final state have been authenticated by deterministic replay.
    state = deepcopy(episode.initial)
    for event in events:
        step = episode.bundle.transition(state, event['tool'], event['arguments'])
        if step['observation'] != event['observation']:
            return score
        state = step['state']
    if state != final_state:
        return score
    removable = set()
    if info['slot'] == 'europeana' and complete_public_coverage(episode, events):
        removable.add(catalog_failure)
        score['checks'].append('complete public search superset covers required catalog scope')
    elif info['slot'] == 'ckan':
        evidence = organization_search_evidence(episode, events)
        if evidence:
            removable.add(identifier_failure)
            if evidence['selected_dataset_and_resource_metadata_complete']:
                removable.add(metadata_failure)
            score['checks'].append('observed organization identifier resolves to complete required search population')
    if removable:
        score['failures'] = [item for item in score['failures'] if item not in removable]
        score['passed'] = score['content_passed'] = not bool(score['failures'])
    return score

	"""Content scoring with the evaluated public alias/search corrections."""
	from copy import deepcopy
	import hashlib
	import json
	from common import read
	from evaluate import Episode

	def require(condition, message):
	if not condition:
	raise ValueError(message)

	def encode(name):
	result = name.replace(".", "__")
	if len(result) <= 64:
	return result
	suffix = "__h" + hashlib.sha256(name.encode()).hexdigest()[:12]
	return result[:64-len(suffix)] + suffix

	def canonical(value):
	return json.dumps(value, sort_keys=True, ensure_ascii=False, allow_nan=False, separators=(',', ':'))

	def decode_observed_tools(answer, env, public_names, execution_log, encode):
	"""Pure, narrow decoding; argument/error/effect authentication stays in oracle."""
	require(type(public_names) is list and all(type(n) is str and n for n in public_names),
	'Invalid public tool inventory')
	require(len(set(public_names)) == len(public_names), 'Duplicate public tool names')
	qualified = {env + '.' + name: name for name in public_names}
	# Check every declared tool, including tools the agent never chose to call.
	all_aliases = {}
	for full, local in qualified.items():
	for alias in (local, full, encode(full)):
	require(alias not in all_aliases or all_aliases[alias] == local, 'Ambiguous public tool alias')
	all_aliases[alias] = local
	require(type(execution_log) is list, 'Missing official execution log')
	observed = set()
	for event in execution_log:
	require(type(event) is dict and type(event.get('tool')) is str, 'Malformed tool event')
	name = event['tool']
	if name in ('task_documents.list_documents', 'task_documents.read_document'):
	continue
	require(name in qualified, 'Cross-environment or nonpublic observed tool')
	observed.add(qualified[name])
	aliases = {alias: local for alias, local in all_aliases.items() if local in observed}
	result = deepcopy(answer)
	witnesses = result['missing_evidence']
	require(type(witnesses) is list, 'Malformed missing_evidence')
	changes = []
	for index, witness in enumerate(witnesses):
	# Other witness schemas/fields are left to the frozen schema and oracle.
	if not isinstance(witness, dict) or 'tool' not in witness:
	continue
	name = witness['tool']
	require(type(name) is str and name in aliases, 'Unknown, cross-environment, or unobserved witness tool')
	local = aliases[name]
	if local != name:
	witness['tool'] = local
	changes.append({'json_pointer': f'/missing_evidence/{index}/tool', 'original': name,
	'normalized': local, 'observed_qualified_name': env + '.' + local})
	restored = deepcopy(result)
	for change in changes:
	restored['missing_evidence'][int(change['json_pointer'].split('/')[2])]['tool'] = change['original']
	require(canonical(restored) == canonical(answer), 'Decoding modified a non-tool field')
	return result, changes, {'observed_qualified_tools': sorted(env + '.' + n for n in observed),
	'allowed_alias_to_domain_name': aliases}

	def complete_public_coverage(episode, events):
	required = {r['record_id'] for r in episode.initial['records'] if episode.bundle.oracle._search_match(r)}
	for event in events:
	if event['tool'] != 'search_catalog_records':
	continue
	args, obs = event['arguments'], event['observation']
	if not set(args).issubset({'query', 'theme', 'reusability', 'media', 'sort'}):
	continue
	if str(args.get('query', '')).strip().lower() not in {'harbor', '*'}:
	continue
	if any(k in args and args[k] != v for k, v in {'theme': 'art', 'reusability': 'open', 'media': True}.items()):
	continue
	if obs.get('ok') is not True:
	continue
	found = obs.get('result', {})
	items = found.get('items', [])
	ids = {x.get('record_id') for x in items}
	if (type(found.get('total_results')) is int and found['total_results'] == len(items) == len(ids)
	and required.issubset(ids)):
	return {'search_arguments': deepcopy(args), 'observed_ids': sorted(ids),
	'required_ids': sorted(required), 'complete_superset': True}
	return None

	def organization_search_evidence(episode, events):
	oracle = episode.bundle.oracle
	derived, problem = oracle._derive(episode.initial)
	if problem:
	return None
	org = derived['org']
	observed_identity = False
	required = {d['id']: d for d in derived['searched']}
	fields = ('id', 'name', 'owner_org', 'state', 'series_key', 'edition',
	'coverage_start', 'coverage_end', 'metadata_modified')
	for event in events:
	args, obs = event['arguments'], event['observation']
	if obs.get('ok') is not True:
	continue
	result = obs.get('result')
	if event['tool'] == 'organization_autocomplete' and isinstance(result, list):
	exact = [r for r in result if isinstance(r, dict) and r.get('title') == oracle.TARGET_TITLE]
	if len(exact) == 1 and oracle._has_fields(exact[0], org, ('id', 'name', 'title')):
	observed_identity = True
	if event['tool'] != 'package_search' or not observed_identity:
	continue
	# Accept only the two real, observed identifiers; never guess an alias.
	if (set(args) != {'owner_org', 'tag', 'include_private'}
	or args['owner_org'] not in (org['id'], org['name'])
	or args['tag'] != oracle.TAG or args['include_private'] is not False
	or not isinstance(result, dict)):
	continue
	records, count = result.get('results'), result.get('count')
	if not isinstance(records, list) or not all(isinstance(r, dict) for r in records):
	continue
	ids = [r.get('id') for r in records]
	if (not all(isinstance(i, str) for i in ids) or type(count) is not int
	or count != len(records) or len(set(ids)) != count or set(ids) != set(required)):
	continue
	if not all(oracle._has_fields(next(r for r in records if r['id'] == key), value, fields)
	for key, value in required.items()):
	continue
	# The sealed oracle counts metadata carried by an accepted search too.
	# Recheck its exact fields/resource-collection predicate on this search,
	# so use of a name cannot erase valid metadata that was already read.
	selected_metadata = all(any(
	oracle._has_fields(record, ds, fields)
	and oracle._resource_collection_exact(record.get('resources'),
	[r for r in derived['all_resources'] if r.get('package_id') == ds['id']])
	for record in records) for ds, _ in derived['pairs'])
	return {'observed_organization': {k: org[k] for k in ('id', 'name', 'title')},
	'search_arguments': deepcopy(args), 'complete_observed_dataset_ids': sorted(ids),
	'required_dataset_ids': sorted(required), 'complete_same_population': True,
	'selected_dataset_and_resource_metadata_complete': selected_metadata}
	return None


	def parse_answer(text):
	"""Parse exactly one JSON report, rejecting duplicate keys and nonfinite values."""
	def unique(items):
	result = {}
	for key, value in items:
	require(key not in result, 'Duplicate JSON key: ' + key)
	result[key] = value
	return result
	def invalid(value):
	raise ValueError('Nonfinite JSON value: ' + value)
	return json.loads(text, object_pairs_hook=unique, parse_constant=invalid)


	def grade_submission(root, number, side, submission):
	"""Grade a controller-recorded report, trace and state using the released rules.

	`events` use domain-local tool names. `final_state` is the domain state,
	not a model's assertion about whether the task succeeded.
	"""
	episode = Episode(root, number, side)
	try:
	require(isinstance(submission, dict), 'Submission must be an object')
	require(('answer' in submission) != ('final_output' in submission),
	'Provide exactly one of answer or final_output')
	answer = (parse_answer(submission['final_output']) if 'final_output' in submission
	else parse_answer(canonical(submission['answer'])))
	except (ValueError, TypeError, KeyError) as exc:
	return {'passed': False, 'format_passed': False, 'content_passed': None,
	'error_kind': 'report_format', 'errors': [str(exc)]}
	try:
	events, final_state = submission['events'], submission['final_state']
	require(isinstance(events, list), 'Missing controller event list')
	require(all(isinstance(e, dict) and set(e) == {'tool', 'arguments', 'observation'}
	and isinstance(e['tool'], str) and isinstance(e['arguments'], dict)
	for e in events), 'Invalid controller event format')
	except (ValueError, TypeError, KeyError) as exc:
	return {'passed': False, 'format_passed': None, 'content_passed': False,
	'error_kind': 'invalid_trace', 'errors': [str(exc)]}

	score = episode.grade(answer, events, final_state)
	if score.get('format_passed') is False:
	return score
	info = episode.info
	if isinstance(answer, dict) and answer.get('missing_evidence') and info['kind'] == 'bundle':
	try:
	tools = read(episode.root / 'private/bundles' / info['slot'] / 'public/tools.json')
	log = [{'tool': info['env_name'] + '.' + event['tool']} for event in events]
	normalized, changes, _ = decode_observed_tools(
	answer, info['env_name'], [tool['name'] for tool in tools], log, encode)
	if changes:
	score = episode.grade(normalized, events, final_state)
	except (ValueError, RuntimeError, KeyError, TypeError) as exc:
	return dict(score, passed=False, content_passed=False,
	normalization_error={'type': type(exc).__name__, 'message': str(exc)[:600]})

	catalog_failure = 'missing complete evidence for the prescribed search result'
	identifier_failure = 'missing complete public tagged organization dataset search evidence'
	metadata_failure = 'missing complete chosen-dataset and resource metadata evidence'
	failure = {'europeana': catalog_failure, 'ckan': identifier_failure}.get(info['slot'])
	if failure is None or failure not in score.get('failures', []):
	return score
	# These narrow equivalences apply only after every tool observation and
	# the final state have been authenticated by deterministic replay.
	state = deepcopy(episode.initial)
	for event in events:
	step = episode.bundle.transition(state, event['tool'], event['arguments'])
	if step['observation'] != event['observation']:
	return score
	state = step['state']
	if state != final_state:
	return score
	removable = set()
	if info['slot'] == 'europeana' and complete_public_coverage(episode, events):
	removable.add(catalog_failure)
	score['checks'].append('complete public search superset covers required catalog scope')
	elif info['slot'] == 'ckan':
	evidence = organization_search_evidence(episode, events)
	if evidence:
	removable.add(identifier_failure)
	if evidence['selected_dataset_and_resource_metadata_complete']:
	removable.add(metadata_failure)
	score['checks'].append('observed organization identifier resolves to complete required search population')
	if removable:
	score['failures'] = [item for item in score['failures'] if item not in removable]
	score['passed'] = score['content_passed'] = not bool(score['failures'])
	return score