SueMintony/ScienceVision-STEM-250-review

ScienceVision — expert annotation tasks, review bundle

285 3D science review records from the ScienceVision annotation platform. Submitted expert answers are the ground truth; records still in question or answer work remain explicitly answer-less. Sub-domains follow 2026-09-09.

1. Distribution

Domain totals: biology 87 · medical 63 · architecture 51 · mechanical-cad 51 · chemistry 33 · all 285

domain	sub_domain	tasks
biology	bio_molecular_contact	70
biology	bio_component_configuration	17
medical	med_lesion_localization	35
medical	med_imaging_geometry	28
architecture	arch_code_compliance	25
architecture	arch_cross_floor_alignment	9
architecture	arch_structural_defects (civil)	8
architecture	arch_consistency_alignment	6
architecture	arch_structural_deformation (civil)	3
mechanical-cad	cad_parts_features	29
mechanical-cad	cad_assembly_interfaces	12
mechanical-cad	cad_mechanisms_functions	8
mechanical-cad	cad_geometry_drawings	2
chemistry	chem_molecular_conformation	20
chemistry	chem_molecular_extent	13

The taxonomy also declares bio_interaction_typing, chem_crystal_periodicity, chem_synthetic_macromolecule, med_lesion_characterization, med_comparative_reasoning, arch_mep_coordination with nothing in them yet — the 2026-09-09 meeting created them and they await authoring. subdomains.json carries the whole table, empty buckets included.

2. Question directions by sub-domain

domain	sub_domain	question direction (task_type)	tasks
biology	bio_molecular_contact	molecular_contact_geometry	29
biology	bio_molecular_contact	chem_ligand_polymer_contact_geometry	21
biology	bio_molecular_contact	bio_docked_component_proximity	11
biology	bio_molecular_contact	bio_collision_overlap_magnitude	3
biology	bio_molecular_contact	bio_collision_pair_distance	3
biology	bio_molecular_contact	bio_docking_component_nearest_pair	1
biology	bio_molecular_contact	chem_binding_component_atom_context	1
biology	bio_molecular_contact	chem_binding_het_contact	1
biology	bio_component_configuration	bio_assembly_centrality	8
biology	bio_component_configuration	bio_chain_terminus_orientation	4
biology	bio_component_configuration	chem_polymer_chain_topology	4
biology	bio_component_configuration	bio_docking_component_absence	1
medical	med_lesion_localization	anatomy-* — 26 body-part slugs, kept as the anatomy_region field	35
medical	med_imaging_geometry	med_ct_geometry	28
architecture	arch_code_compliance	stair_ramp_spatial_validity	13
architecture	arch_code_compliance	utility_network_clash_detection	12
architecture	arch_cross_floor_alignment	cross_floor_alignment_validation	9
architecture	arch_structural_defects	bridge-damage-localization	4
architecture	arch_structural_defects	concrete-void-detection	4
architecture	arch_consistency_alignment	drawing-bim-consistency	4
architecture	arch_consistency_alignment	section_elevation_reasoning	2
architecture	arch_structural_deformation	girder-spatial-deformation	3
mechanical-cad	cad_parts_features	cad-handedness-chirality	8
mechanical-cad	cad_parts_features	cad-bom-unique-parts	6
mechanical-cad	cad_parts_features	cad-feature-fillet-chamfer	6
mechanical-cad	cad_parts_features	assembly-occluded-count	5
mechanical-cad	cad_parts_features	cad_assembly_component_identity	4
mechanical-cad	cad_assembly_interfaces	cad-joint-type	4
mechanical-cad	cad_assembly_interfaces	cad-mating-interfaces	4
mechanical-cad	cad_assembly_interfaces	cad_assembly_completeness	4
mechanical-cad	cad_mechanisms_functions	cad-firing-interval	4
mechanical-cad	cad_mechanisms_functions	cad-turnability	2
mechanical-cad	cad_mechanisms_functions	cad_mechanism_kinematics	2
mechanical-cad	cad_geometry_drawings	depth-collapsed-planarity	2
chemistry	chem_molecular_conformation	chem_conformation_explicit_dihedral	18
chemistry	chem_molecular_conformation	chem_conformation_continuous_signed_dihedral	1
chemistry	chem_molecular_conformation	chem_conformation_signed_absolute_dihedral	1
chemistry	chem_molecular_extent	chem_xyz_heavy_extent	6
chemistry	chem_molecular_extent	chem_xyz_farthest_heavy_pair	5
chemistry	chem_molecular_extent	chem_xyz_centroid_radius	2

med_lesion_localization also records the orthogonal body-part axis in anatomy_region: abdomen 13 · pelvis 8 · whole_body_multi_region 4 · thorax 4 · brain 2 · spine 1 · upper_extremity 1 · neck 1 · lower_extremity 1.

3. File structure

.
├── index.html            browse, search and filter every task; opens over file://
├── index.jsonl           one JSON row per task — the machine-readable index
├── index_data.js         the same rows inlined, so index.html needs no network
├── subdomains.json       the canonical taxonomy: buckets, labels, task types, counts
├── VERIFY.json           this build's audit — gates, censuses, pinned exporter commit
└── tasks/
    └── <domain>/<sub_domain>/<NNN>-<task_id>/
        ├── question.md   the sheet a human reads
        ├── task.json     the same task, for a script
        ├── assets/       3D geometry and volumes
        └── drawings/     2D engineering drawings — only tasks that have them

One folder per domain, and the whole domain is in it: architecture/ 51, biology/ 87, chemistry/ 33, mechanical_cad/ 51, medical/ 63.

A domain's tasks are all in its folder, whatever stage of review they are at. Not every domain is at the same point: mechanical-cad holds 34 pending_answer_qc, 14 answer_in_progress, 3 pending_question_qc — none of them approved yet. Every task in architecture, biology, chemistry, medical is approved. review_stage on each record says where a task stands:

review_stage	tasks	what it means
approved	234	the expert's answer passed answer QC
pending_answer_qc	34	answer submitted, QC not done
answer_in_progress	14	assignment issued, answer not submitted yet
pending_question_qc	3	the question itself is still in review, so no answer exists yet

Filter on it before consuming: jq 'select(.review_stage == "approved")'.

The path is the canonical classification, not the platform's. A task authored as chemistry sits under biology/ when its molecular system contains protein or nucleic acid; platform.domain records what the platform called it.

The <NNN>- ordinal is per bucket and stable within a build, not a global task number — adding a task renumbers its siblings. Of the 285 tasks, 80 ship a 2D drawing alongside the 3D file.

What ships beside each question

	values
kind	3d 319, 2d 319, aux 6
role	primary 411, engineering_drawing 125, mask 28, source_semantic_drawing_front 10, viewer_drawing_sheet_front 10, source_semantic_drawing_left 10, +5 more drawing roles
format	.jpg 205, .png 114, .nii.gz 91, .cif 87, .glb 46, .xyz 33, .step 27, .stp 24, .json 11, .seg 6

Files keep the names the platform gave them. role: mask is the one to watch: for a medical task the mask is the ground truth, shipped so a reviewer can check the measurement, and it must never reach an agent as input.

4. Key fields

task.json is sciencevision.review_bundle_task.v4. question.md is the same content rendered for a human; index.jsonl is a flattened subset for scripts.

Identity and placement

field	in this bundle	what it decides
task_id, question_id	285 distinct	public record keys; a platform X-copy assignment is published as X-qc
source_task_id, platform_task_id, source_question_id, assignment_id	provenance preserved	joins the public record back to the original platform assignment
review_stage	approved 234, pending_answer_qc 34, answer_in_progress 14, pending_question_qc 3	lifecycle state; it does not change the task's folder tree
folder	—	the task's own path inside the bundle

Classification

field	in this bundle	what it decides
domain	biology 87, medical 63, architecture 51, mechanical-cad 51, chemistry 33	the canonical domain, taken from the sub-domain and not from whoever authored the task
sub_domain	bio_molecular_contact 70, med_lesion_localization 35, cad_parts_features 29, med_imaging_geometry 28, arch_code_compliance 25, +10 more	the unique English key every consumer joins on
sub_domain_label_en / _zh	—	display names, in the meeting's own wording
sub_domain_group	civil_engineering 11	marks the down-weighted civil pair inside architecture
task_type	molecular_contact_geometry 29, med_ct_geometry 28, chem_ligand_polymer_contact_geometry 21, chem_conformation_explicit_dihedral 18, +61 more	the question generator's own family name — the key the taxonomy resolves from
taxonomy_version	2026-09-09 285	which ruling produced this filing
anatomy_region	abdomen 13, pelvis 8, whole_body_multi_region 4, thorax 4, brain 2, spine 1, upper_extremity 1, neck 1, lower_extremity 1	medical only: the body part, kept as a field because it stopped being a bucket
platform.domain, .sub_domain_slug, .sub_domain_label	medical 63, biology 60, chemistry 60, architecture 51, mechanical-cad 51	what the platform itself said, kept so a disagreement stays auditable

The question

field	in this bundle	notes
question.zh / .en	285 / 285 have at least one	the expert reads the Chinese; both are the same task
question.source	effective_question 282, task_surface 3	effective_question is the text question-QC signed off on
answer_shape.fields	declared on 154 / 285	the fields the answer is supposed to fill
viewer	molstar 120, niivue 63, occt 51, gltf 36, architecture_inspection 11, bim_coordination 4	the 3D viewer the task was authored against
title	non-empty on 78 / 285	short label, often absent

The answer, which is the ground truth

No generator-produced answer is carried. Where an expert has submitted an answer, that submission is the ground truth; records still in question or answer work have a null ground truth until an answer exists.

field	in this bundle	notes
ground_truth	268 / 285	a parsed object when the answer was structured, otherwise the text verbatim; null only where no answer exists yet
ground_truth_source	expert_submission 268	one source by design
answer_form	structured_matching_schema 96, prose 80, structured 79, absent 17, structured_key_mismatch 13	whether a machine can compare it field by field
evaluation_ready	true on 175 / 285	true exactly when answer_form is structured or structured_matching_schema
ground_truth_extraction	verbatim 152, none 97, embedded_json 20, repaired_syntax 16	how the value was obtained from the submission
answer_repaired	true on 16	the answer's syntax was fixed at build time; expert_submission.answer_raw keeps the original
expert_submission.answer / .reasoning	—	verbatim, as submitted

• structured_matching_schema — a JSON object whose keys are exactly the declared fields
• prose — free text — the answer is there, but not field by field
• structured — a JSON object, on a task that declares no field list
• absent — no answer at all
• structured_key_mismatch — a JSON object whose keys match none of the declared fields

embedded_json covers 20 answers that walked through the measurement in prose and ended with the JSON block the question asked for. That block becomes the ground truth and the prose stays in answer_raw; two or more candidate blocks stop the extraction rather than pick one.

The repair is mechanical and never changes a value: it removes the answer template's angle brackets, turns full-width commas inside a tuple into a JSON array, and drops a trailing comma. A unit is stripped only when the field name already declares that unit, so 3.408Å in contact_distance_A becomes 3.408 while 0.315m in a _mm field is refused rather than silently converted.

Risk markers and run context

field	in this bundle	notes
review_flags.warn_kinds	answer_key_may_not_match_question 43, answer_not_yet_qc_approved 34, twin_question_text_differs 23, twin_answers_disagree 14, answer_not_submitted 14, question_still_in_review 3, question_language_flipped_to_chinese 1	why a reviewer should look twice
review_flags.qc_independence	self 155, unknown 61, cross 53, admin 16	self means the same person authored and approved
review_flags.ever_returned	False 281, True 4	the task was sent back at least once
task_context.mcp_allow_list_effective	non-empty on 285 / 285	the tools the expert was allowed to use
task_context.mcp_allow_list_source	declared 201, tool_policy 84	tool_policy means the task declared no mcp_allow_list, so the effective list came from tool_policy.allowed_tools. Nothing is invented
task_context.twin	mechanical tasks only	the other expert on the same base task: how many answered, and whether they agreed
task_context.vision_class	pixel-grounded evidence trace 75, visibility and occlusion 22, constrained invariant inference 21	how far the task depends on looking rather than computing
annotator_pseudonym, expert_id	13 distinct	stable across tasks, so a reviewer can see which tasks share an annotator. No real name is in the bundle

What the build guarantees

Three fields are gated fail-closed — a task missing any of them stops the build — so these are guarantees rather than measurements:

• a question in at least one language: 285 / 285
• an effective mcp_allow_list: 285 / 285
• an expert answer, and therefore a ground_truth, on every task whose review stage means one exists (answer_in_progress and pending_question_qc have none yet)

The completed bundle is also verified: no expert real name appears anywhere in the bundle, every declared asset exists at its declared size and SHA-256, public ids and paths are unique and contain no copy, and every task resolves to a declared sub-domain whose domain matches the record's. VERIFY.json holds the results.

Is task.json a harbor task?

No, and not yet by design. task.json is a record for a human reviewer; the evaluation runner consumes sci3d_bench.task_schema.TaskSpec (schema_version 1.0). Three things stand between them:

1. 110 records are not field-comparable. 17 have no ground truth yet; prose answers and schema mismatches also ship evaluation_eligible: false until someone normalises them.
2. No scorer is assigned. Review needs none; TaskSpec requires one.
3. Some asset formats are outside ALLOWED_FORMATS, and _validate rejects an unknown one:

format	assets	why it is not a straight lift
jpg	205	a rendered drawing, so it lifts to prerendered_image, not to asset
png	114	a rendered drawing, so it lifts to prerendered_image, not to asset
cif	87	needs a decision: COD files parse as cifCore, RCSB as mmcif, so one blanket mmcif label would parse a third of them wrong
seg	6	a segmentation sidecar with no format in the schema at all

The input side already maps one to one: question, answer_shape → answer_schema, assets[] → asset / assets, task_context.trace_requirements → trace_requirements, tool_policy → tool_policy, evidence_requirements → required_evidence, vision_class → v_class, platform + provenance → provenance.

5. How to use it

Read it. Open index.html — double-click is enough, no server and no network. Search the question text, an answer or an id; filter by domain, sub-domain, review stage or viewer; click a row for the full question, the expert's answer and reasoning, the files, and that task's taxonomy provenance.

Script it. index.jsonl is one row per task:

# tasks per canonical bucket
jq -r '[.domain, .sub_domain] | @tsv' index.jsonl | sort | uniq -c

# ground truths a field-by-field scorer can use today
jq -r 'select(.evaluation_ready) | .task_id' index.jsonl

# the rest, with the reason
jq -r 'select(.evaluation_ready | not) | "\(.task_id)\t\(.answer_form)"' index.jsonl

# everything not yet approved
jq -r 'select(.review_stage != "approved") | .folder' index.jsonl

# tasks that ship a 2D engineering drawing
jq -r 'select(.asset_2d > 0) | "\(.task_id)\t\(.asset_2d)"' index.jsonl

# the tools one task's expert was allowed to use
jq -r 'select(.task_id == "<id>") | .mcp_allow_list[]' index.jsonl

Four things to watch for.

1. Review is still open on 51 records: 34 await answer QC, 14 await an answer, and 3 await question QC.
2. 110 records are not field-comparable yet — no answer, prose, or JSON whose keys miss the declared fields. answer_form says which case each one is.
3. A medical task's mask.nii.gz is the ground truth. Shipping it to a reviewer is deliberate; it must never be an agent's input.
4. No trajectories. Keyframes are not in this bundle, so an expert's operation sequence cannot be replayed from it.

6. Where it came from

Sources: ScienceVision-upload-0909/01_done_deliverable contributes 224 approved records. The 2026-09-10 ScienceVision-answer-qc-0910/ pull contributes architecture-approved 10 · mechanical-assembly-occluded 3 · mechanical-cad 48; overlap with the approved corpus is removed before writing. Input policies: architecture-approved base mode, kept 10/51 · mechanical-assembly-occluded base mode, kept 3/6 · mechanical-cad assignment mode, kept 48/48.

mechanical-cad ships at assignment granularity: both expert assignments remain visible, with public ids X and X-qc. mechanical-assembly-occluded ships at base granularity: its six source assignments become three question-review records. source_task_id, platform_task_id, source_question_id and assignment_id preserve the original values, including -copy, inside provenance fields.

The platform exporter is pinned to BabyVision-v2 commit 3ccee5b7096f5be56b963ee83882f9c3bd624adc. Question text and sub-domain come from the platform's own approved_export._selected_rows, not re-derived locally. Rebuild with:

python3 -m scripts.review_showcase.pack --out <this directory> \
  --qc-bundle <answer-qc pull>/architecture-approved \
  --qc-bundle <answer-qc pull>/mechanical-assembly-occluded \
  --qc-assignment-bundle <answer-qc pull>/mechanical-cad

A local build also holds a content-addressed assets/ pool, hard-linked to the per-task copies. It is not published, because nothing reads it.