usedot/dot-loom-conductor-v2

Dot Loom Conductor v2

11,100 synthetic traces for training budget-constrained multi-model routers.

Trained adapter  ·  Policy explorer  ·  Generator and receipts

Each example describes one task, three anonymous worker profiles, hard call, credit, and latency budgets, and the highest-utility feasible Lean, Balanced, or Strict execution plan.

Property	Value
Examples	11,100
Train / validation / test	9,000 / 900 / 1,200
Policy balance	3,700 Lean / 3,700 Balanced / 3,700 Strict
Task families	19, disjoint across splits
Candidate plans evaluated per example	15
User or production data	0
Dataset seed	20260716

Load

from datasets import load_dataset

dataset = load_dataset("usedot/dot-loom-conductor-v2")
print(dataset)
print(dataset["train"][0]["target"])

Split design

Split	Examples	Lean	Balanced	Strict	Task families
Train	9,000	3,000	3,000	3,000	8
Validation	900	300	300	300	3
Test	1,200	400	400	400	8
Total	11,100	3,700	3,700	3,700	19

Task families are disjoint across splits. The test families are payment races, tenant isolation, webhook replay, OAuth integrity, SSRF egress, stream settlement, health triage, and contract risk. None appears in training or validation.

The JSONL payload is 46,900,561 bytes.

Schema

Each row includes both a serialized completion-training pair and inspectable source fields.

Field	Type	Meaning
prompt	string	Complete conductor prompt with input and output schema
target	string	Canonical compact JSON completion target
task	object	Risk, complexity, consequence, ambiguity, evidence need, reversibility, and input size
workers	list[object]	Three anonymous capability, reliability, cost, latency, provider, and strength profiles
constraints	object	Maximum calls, credits, latency, and target minimum quality
label	object	Policy, roles, receipts, reasons, verification, and access graph
oracle	object	Utility, predicted quality, pass rate, candidate count, and target attainment
family	string	Synthetic task family used for split isolation
split	string	Train, validation, or held-out test
schema_version	string	loom-conductor-v2

{
  "task": {
    "summary": "Audit a user-controlled fetcher for SSRF and metadata access.",
    "risk": 0.95,
    "complexity": 0.88,
    "consequence": 0.96,
    "evidence_need": 0.92
  },
  "constraints": {
    "max_calls": 3,
    "max_credits": 16.5,
    "max_latency_ms": 95000,
    "minimum_quality": 0.90
  },
  "label": {
    "policy": "strict",
    "writer": "worker_a",
    "reviewer": "worker_b",
    "finalizer": "worker_c",
    "independent_verification": true
  }
}

The example is abbreviated for readability. The dataset viewer exposes every complete field and canonical target.

Label construction

For every example, the generator enumerates all 15 role assignments:

• Three Lean writer-only plans
• Six ordered writer-reviewer plans
• Six ordered writer-reviewer-finalizer plans

Plans that exceed hard call, credit, or latency budgets are rejected. The disclosed outcome model scores task-adjusted quality, pass rate, cost, latency, role fitness, provider independence, reviewer regression risk, and finalizer synthesis. The highest-utility feasible plan becomes the label. If any feasible plan meets the target quality, below-target plans are excluded.

Worker profiles are sampled around three frozen empirical anchors from a six-case Dot Loom receipt. Model names never appear in training prompts. Worker identifiers and provider groups are shuffled so the model must route from capability and constraint fields instead of memorizing brands.

Independent label audit

A blinded independent OpenAI judge audited 90 stratified train and validation labels. Test labels were not sent to the judge.

Audit outcome	Count
Agreement	76 of 90
Proposed disagreements	14
Feasible alternatives	3
Invalid or over-budget alternatives	11
Feasible alternatives improving the disclosed oracle	0

The judge is an independent critic, not ground truth. Raw responses and deterministic revalidation are included in the public research package.

Privacy and provenance

The corpus contains no user prompts, production conversations, wallets, API keys, private code, or customer data. Task descriptions are synthetic. Worker identities and provider groups are anonymous and shuffled.

Split	SHA-256
Train	8e8244bb5ad7034f6d2a5949503526762895072875998662510d5a3a9b269739
Validation	b6579587f1123c32b58fe5e55ca9516a5e2d94c47e04a1a84405802dfa4f2795
Test	9cd3065a36787ce18042aa8bde7e10b1ccf3acd04093058cf78c77f338c5d049

Intended use

• Supervised fine-tuning for structured routing-plan generation
• Research on cost, latency, independence, verification, and model-allocation policies
• Comparing learned conductors with deterministic routers
• Testing deterministic enforcement around model-proposed plans

Limitations

• Labels are generated by a simulator calibrated to a six-case empirical receipt.
• Worker capability values are abstractions and can drift from live provider performance.
• Synthetic task descriptors are simpler than full production prompts.
• The held-out test measures routing-plan generation, not downstream answer correctness.
• The balanced label distribution does not represent natural production traffic.
• Production systems must validate learned plans with deterministic runtime enforcement.

Citation

@dataset{dotloomconductorv2_2026,
  title     = {Dot Loom Conductor v2},
  author    = {{Dot R\&D}},
  year      = {2026},
  month     = {7},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/usedot/dot-loom-conductor-v2}
}

License

Released under Apache 2.0.