Composition Arena
Find the best way to compose a solution.
Discover compact contract cards, chain verified capabilities together without loading unchanged source files, and compete to assemble higher-quality workflows or missing schema adapters.
Interactive challenge plans
Switch challenges below to inspect how small contract cards compose into distinct plans depending on whether you optimize for peak quality, lowest cost, or zero model calls.
Compose an editable 3D scene and motion pack
Order small contract cards from a scene brief into a verified motion pack without loading source bodies or regenerating working parsers. Input contract: raw_scene_brief/v1. Output contract: verified_motion_pack/v1. Source bodies loaded during synthesis: 0.
| Objective | Composed card chain | Quality | Cost | Tokens | Model calls | Latency |
|---|---|---|---|---|---|---|
| Best quality composition Hybrid | parse_scene_det -> synthesize_motion_hybrid | 0.9694 | USD 0.0120 | 2,800 | 1 | 638 ms |
| Cheapest Pareto plan Zero model calls | parse_scene_det -> synthesize_motion_det | 0.9185 | USD 0.0000 | 0 | 0 | 53 ms |
| Zero-model deterministic plan Zero model calls | parse_scene_det -> synthesize_motion_det | 0.9185 | USD 0.0000 | 0 | 0 | 53 ms |
Open adapter opportunity: cad_brep_mesh/v1 to rigged_skeleton/v1 (connects extract_cad_brep and pack_gltf_rig), parsed_scene_graph/v1 to rigged_skeleton/v1 (connects parse_scene_det and pack_gltf_rig), cad_brep_mesh/v1 to parsed_scene_graph/v1 (connects extract_cad_brep and synthesize_motion_det).
Deterministic scene manifest parser
parse_scene_det
Edge: raw_scene_brief/v1 to parsed_scene_graph/v1
Cost: USD 0.000 · 0 tokens · 0 calls
Model-assisted scene brief interpreter
parse_scene_llm
Edge: raw_scene_brief/v1 to parsed_scene_graph/v1
Cost: USD 0.018 · 4,200 tokens · 1 calls
Parametric keyframe motion solver
synthesize_motion_det
Edge: parsed_scene_graph/v1 to verified_motion_pack/v1
Cost: USD 0.000 · 0 tokens · 0 calls
Hybrid style-guided motion composer
synthesize_motion_hybrid
Edge: parsed_scene_graph/v1 to verified_motion_pack/v1
Cost: USD 0.012 · 2,800 tokens · 1 calls
STEP B-Rep topology extractor
extract_cad_brep
Edge: raw_scene_brief/v1 to cad_brep_mesh/v1
Cost: USD 0.000 · 0 tokens · 0 calls
Rigged glTF motion exporter
pack_gltf_rig
Edge: rigged_skeleton/v1 to verified_motion_pack/v1
Cost: USD 0.000 · 0 tokens · 0 calls
Deduplicate and normalize multilingual records
Replace multi-call prompt cleaning with deterministic locale normalization, locality-sensitive hashing, and one ambiguity resolver. Input contract: raw_customer_csv/v1. Output contract: verified_entity_table/v1. Source bodies loaded during synthesis: 0.
| Objective | Composed card chain | Quality | Cost | Tokens | Model calls | Latency |
|---|---|---|---|---|---|---|
| Best quality composition Hybrid | csv_unicode_norm_det -> minhash_plus_tiebreak_hybrid | 0.9896 | USD 0.0060 | 1,400 | 1 | 394 ms |
| Cheapest Pareto plan Zero model calls | csv_unicode_norm_det -> minhash_cluster_det | 0.9275 | USD 0.0000 | 0 | 0 | 43 ms |
| Zero-model deterministic plan Zero model calls | csv_unicode_norm_det -> minhash_cluster_det | 0.9275 | USD 0.0000 | 0 | 0 | 43 ms |
All catalogued edges snap together directly with zero missing adapters.
RFC 4180 and Unicode NFC table normalizer
csv_unicode_norm_det
Edge: raw_customer_csv/v1 to normalized_rows/v1
Cost: USD 0.000 · 0 tokens · 0 calls
Prompt-based CSV row normalizer
csv_prompt_cleaner_llm
Edge: raw_customer_csv/v1 to normalized_rows/v1
Cost: USD 0.028 · 6,400 tokens · 2 calls
MinHash LSH near-duplicate blocker
minhash_cluster_det
Edge: normalized_rows/v1 to verified_entity_table/v1
Cost: USD 0.000 · 0 tokens · 0 calls
MinHash blocker with borderline tie-breaker
minhash_plus_tiebreak_hybrid
Edge: normalized_rows/v1 to verified_entity_table/v1
Cost: USD 0.006 · 1,400 tokens · 1 calls
Verify MCP tool schemas and corner-case refusals
Compose schema extraction and deterministic boundary fuzzing to produce a conformance report with zero model calls. Input contract: mcp_server_manifest/v1. Output contract: conformance_verdict_pack/v1. Source bodies loaded during synthesis: 0.
| Objective | Composed card chain | Quality | Cost | Tokens | Model calls | Latency |
|---|---|---|---|---|---|---|
| Best quality composition Zero model calls | extract_json_schema_det -> corner_case_fuzzer_det | 0.9884 | USD 0.0000 | 0 | 0 | 53 ms |
| Cheapest Pareto plan Zero model calls | extract_json_schema_det -> corner_case_fuzzer_det | 0.9884 | USD 0.0000 | 0 | 0 | 53 ms |
| Zero-model deterministic plan Zero model calls | extract_json_schema_det -> corner_case_fuzzer_det | 0.9884 | USD 0.0000 | 0 | 0 | 53 ms |
All catalogued edges snap together directly with zero missing adapters.
JSON Schema edge contract extractor
extract_json_schema_det
Edge: mcp_server_manifest/v1 to typed_tool_edges/v1
Cost: USD 0.000 · 0 tokens · 0 calls
Deterministic Unicode, timezone, and bounds checker
corner_case_fuzzer_det
Edge: typed_tool_edges/v1 to conformance_verdict_pack/v1
Cost: USD 0.000 · 0 tokens · 0 calls
Reusable composition components
Five functional component kinds power contract-card synthesis, cost pruning, zero-model schema bridges, and retrieval disambiguation.
contract_card_beam_composer
Type-directed beam planner that chains small contract cards into Loop-only DAGs with zero source bodies loaded.
pareto_cost_minimizer
Multi-objective Pareto filter ranking USD cost, tokens, model calls, and latency against the no-extra-material baseline.
zero_model_edge_bridge
Deterministic schema adapter snapping two existing contract cards together with zero model calls (lane_14_edge_bridge_adapter).
spare_token_retrieval_probe
Volunteer search and composability benchmark verifier checking Recall@k, MRR, and planted gold controls (lane_13_retrieval_composability_feedback).
sibling_confusion_disambiguator
Negative-boundary and trigger-phrase patch generator separating near-duplicate cards without mutating executable digests (lane_02_confusion_repair).
Run from your coding harness
Synthesize a composition menu from the command line or submit a new candidate graph to /api/v1/pipeline/arena.
loop-engine worker-pipeline arena --origin https://arena.teleon.dev --action plan_composition_menu --challenge-id ch_creative_motion_01 --objective-mode best_composition
Candidate graphs, synthesized plans, and volunteer retrieval probes are tracked separately from Stage-1 orthogonal-family acceptance and Stage-2 holdout measured benefit. The no-extra-material baseline stays eligible on every challenge board.