wtd-mixed-v1

wtd-mixed-v1

WTD recommends workflow-shape candidates from the workflow-shape-retrieval Intelligence Package. The model is a workflow-shape librarian: it retrieves known topologies, ranks them, and explains why they fit. It does not generate task payloads or execute workflows.

The production use case is an agent or runtime that has an intent or draft workflow graph and wants a reusable topology to adapt into executable ttasks. The target workflow runtime is @ianphil/ttasks-ts, with downstream runtime consumption through ianphil/ttasks-wtd.

Runtime contract

• schema: wtd_runtime_bundle_v1
• version: 0.4.3
• mode: structural
• patterns: 30
• text intent retrieval: True
• metadata fallback retrieval: True
• draft DAG structural retrieval: True
• heuristic ranking: True
• structural encoder status: exported
• heuristic ranker status: exported
• reserved encoder path: encoder.onnx
• reserved latents path: latents.f16

Retrieval surfaces

• Text retrieval uses text-index.json and exported workflow-shape metadata.
• Metadata fallback retrieval works without ONNX and is useful when structural inference is unavailable.
• Draft-DAG structural retrieval uses encoder.onnx and latents.f16 when the bundle advertises draftDagStructuralRetrieval: true.

How structural retrieval works

For draft-DAG retrieval, the runtime converts a workflow graph into four tensor inputs: node_features, adjacency, node_mask, and struct_features. encoder.onnx maps those tensors to a query latent vector. The runtime then compares that vector with each row in latents.f16 using nearest-neighbor distance and returns candidates ordered by closest structural match.

The raw structural score is:

1 / (1 + latent_distance)

Heuristic ranking

The exported ranker is deterministic and inspectable. It does not replace ONNX retrieval; it lightly reranks the raw candidate pool and attaches feature-level explanations. The v0.4.1 ranker calibration keeps retrieval_score dominant and uses node, edge, and depth compatibility as small tie-breakers so strong raw structural matches do not get buried.

Use ranker-config.json for weights and ranker-schema.json for feature names and explanation fields.

Runtime eval coverage

This bundle includes fixed eval metadata for consumer parity checks:

wmd verify-runtime --input <bundle> --json
wmd topology-eval-runtime \
  --input <bundle> \
  --manifest evals/topology-drafts.json \
  --k 15 \
  --json

The v0.4.1 topology gate is 15/15 ranked top-3, 15/15 ranked top-10, and zero raw top-1 matches degraded below ranked top-3.

Usage

wmd verify-runtime --input <bundle> --json
wmd retrieve-runtime \
  --input <bundle> \
  --draft-dag examples/dataset-ingest-draft.json \
  --k 5 \
  --json
wmd benchmark-runtime \
  --input <bundle> \
  --query "dataset ingestion eval publish candidate" \
  --draft-dag examples/dataset-ingest-draft.json \
  --runs 3 \
  --json

Files

• manifest.json

• release.json

• patterns.json

• labels.json

• text-index.json

• checksums.json

• encoder.onnx — encoder-only VAE path for draft-DAG tensors

• latents.f16 — pattern centroid latent matrix in patterns.json order

• text-projection.f32 — hash text projection for runtime node featurization

• ranker-config.json — deterministic heuristic ranker weights

• ranker-schema.json — ranker feature and explanation contract

• evals/gold-queries.json — text retrieval fixture suite

• evals/draft-dag-gold.json — draft-DAG fixture suite

• evals/topology-drafts.json — topology-native draft eval manifest

Limitations

WTD recommends topology. It does not generate executable task payloads. The structural encoder is included for draft DAG latent retrieval.

Use retrieved candidates as starting shapes, then adapt task names, payloads, dependencies, and validation logic for the current workflow.