usedot/Dot-Loom-Conductor-14B

Dot Loom Conductor 14B

A learned control policy for budget-constrained multi-model inference.

Dataset  ·  Policy explorer  ·  Technical report  ·  Source and receipts

Dot Loom Conductor receives a task, three candidate worker profiles, and hard call, credit, and latency budgets. It returns a schema-constrained Lean, Balanced, or Strict execution plan with role assignments, access rules, estimated receipts, and bounded reason codes.

It does not answer the task. It decides how the task should be executed.

Release fact	Result
Base checkpoint	mistralai/Ministral-3-14B-Base-2512
Released artifact	Rank-32 PEFT LoRA adapter
Training corpus	9,000 synthetic routing traces
Held-out evaluation	1,200 family-disjoint cases
Exact-plan match	90.2% raw adapter, 90.8% with runtime guard
Lift over deterministic Loom	+11.5 percentage points with guard
Hard-budget violations after guard	0 of 1,200
Training run	52.6 minutes on 1x NVIDIA H200 SXM5

No user prompts, production conversations, wallets, API keys, private code, or customer data were used in training.

Why this model exists

Most agent stacks either use one model for every request or hard-code a fixed writer and reviewer pipeline. Both approaches spend the same amount of inference on easy and difficult work.

The conductor learns a different policy:

• Lean: one writer call when the task is low-risk and reversible
• Balanced: a writer plus an independent reviewer when verification earns its cost
• Strict: a writer, reviewer, and finalizer for high-consequence or synthesis-heavy work

The roles are model-agnostic. A worker can be OpenAI, Claude, DeepSeek, Qwen, a Dot model, an OpenRouter endpoint, or a local model. Brand names never appear in the training prompts. Routing is based on capability, reliability, provider independence, cost, latency, risk, and hard budgets.

Quick start

The release is a PEFT adapter. The pinned base model is downloaded separately.

git clone https://huggingface.co/usedot/Dot-Loom-Conductor-14B
cd Dot-Loom-Conductor-14B
python -m pip install -r requirements.txt
python inference_example.py

The example loads base revision 5b0ceedbb42dff466ae60b258ba296f32da51384, attaches the adapter, runs greedy decoding, and prints one JSON plan. The published evaluation used BF16. A quantized deployment was not evaluated.

Output contract

A real inference from the archived H200 demo produced this plan for a high-risk payment race:

{
  "policy": "strict",
  "writer": "worker_a",
  "reviewer": "worker_b",
  "finalizer": "worker_c",
  "max_calls": 3,
  "max_credits": 16.5,
  "max_latency_ms": 95000,
  "estimated_credits": 16.15,
  "estimated_latency_ms": 83500,
  "estimated_quality": 0.91,
  "estimated_pass_rate": 0.77,
  "independent_verification": true,
  "access": {
    "writer": [],
    "reviewer": ["writer"],
    "finalizer": ["writer", "reviewer"]
  },
  "reason_codes": [
    "best_feasible_below_target",
    "complex_task",
    "credit_constrained",
    "high_consequence"
  ]
}

Production code must recompute costs, latency, role validity, access, and hard-budget compliance outside the model. The adapter proposes a plan. Deterministic runtime code enforces it.

Evaluation

The frozen test split contains 1,200 examples from eight task families that do not appear in training or validation: payment races, tenant isolation, webhook replay, OAuth integrity, SSRF egress, stream settlement, health triage, and contract risk.

Lane	JSON valid	Policy accuracy	Exact plan	Constraints met	Unsafe under-escalation	Mean regret
Deterministic Loom	100.0%	93.9%	79.3%	100.0%	5.7%	1.733
Raw Ministral 14B Base	1.0%	0.4%	0.1%	0.1%	98.1%	1051.175
Dot-trained raw adapter	100.0%	99.7%	90.2%	98.9%	0.0%	11.823
Dot-trained plus guard	100.0%	100.0%	90.8%	100.0%	0.0%	0.245

The guarded conductor improved exact-plan match over deterministic Loom by 11.5 percentage points. The paired bootstrap 95% confidence interval was 9.4 to 13.7 points. The exact-plan McNemar two-sided p-value was < 1e-12.

The raw adapter exceeded a hard budget on 13 of 1,200 proposals. The published runtime guard caught all 13 and selected its deterministic fallback. The guard is runtime code and is not contained in the adapter weights.

Baseline interpretation

The raw base checkpoint is a pre-training capacity control. It demonstrates that the structured routing behavior was acquired during the disclosed fine-tune. It is not presented as a fair comparison against an instruction-tuned general assistant.

Deterministic Loom is the operational baseline. It receives the same task, worker, and budget fields and already emits a valid plan. Future work should add instruction-tuned routing baselines, live downstream task outcomes, and multiple training seeds.

Training

The adapter was trained on the public Dot Loom Conductor v2 corpus. Each label was selected by enumerating all 15 valid role assignments, rejecting plans that exceed hard budgets, and choosing the highest-utility feasible plan under the disclosed simulator.

Property	Value
Adapter size	560,078,656 bytes
Adapter SHA-256	e0e6655a16a10f28cbce898e564640bf6a4a64ca84bb2014a1da3a60c5f11eda
LoRA rank / alpha / dropout	32 / 64 / 0.05
Trainable parameters	139,984,896
Total base parameters	14,085,016,576
Base revision	5b0ceedbb42dff466ae60b258ba296f32da51384
Hardware	1x NVIDIA H200 SXM5, 143,771 MiB reported VRAM
Train / validation examples	9,000 / 900, family-disjoint
Optimizer steps / epochs	564 / 2
Effective batch size	32
Learning rate / schedule	1.5e-4 / cosine, 28 warmup steps
Precision	BF16
Seed	20260716
Duration	52.6 minutes
Mean GPU utilization	96.57%
Integrated GPU energy	0.588 kWh
Selected checkpoint	Step 564, validation loss 0.070168

LoRA targeted q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj.

Reproducibility

• adapter_config.json pins the upstream base revision.
• requirements.txt pins the tested runtime versions from the archived H200 environment.
• SHA256SUMS covers the public release files and adapter weights.
• evidence/ includes the training receipt, telemetry summary, checkpoint selection, paired statistics, and scored benchmark summaries.
• The research package contains the corpus generator, raw predictions, scoring code, audit responses, tests, charts, and command-line demonstrations.

Intended use

• Research on learned, model-agnostic inference orchestration
• Offline comparison with deterministic routers
• Experimental routing for coding agents and AI application backends
• Local demonstrations of cost, latency, independence, and verification tradeoffs

Not intended for

• Treating predicted quality or pass rate as a production guarantee
• Executing financial, health, legal, security, or infrastructure changes without review
• Granting permissions or selecting tools without a separate authorization layer
• Trusting model-declared costs or budgets without deterministic recomputation

Limitations

• Labels come from a disclosed simulator rather than measured execution of all 15 plans.
• Empirical calibration uses a frozen six-case Dot Loom receipt.
• The benchmark measures routing-plan generation, not final code correctness.
• Worker capability values can drift from live provider performance.
• The study uses one seed, one base checkpoint, and one H200 training run.
• The raw adapter violated a hard budget in 13 of 1,200 test proposals.
• A production deployment must enforce every hard constraint outside the model.

Citation

@techreport{dotloomconductor2026,
  title        = {Dot Loom Conductor 14B: Learned Budget-Constrained Multi-Model Routing},
  author       = {{Dot R\&D}},
  institution  = {Dot},
  year         = {2026},
  month        = {7},
  url          = {https://huggingface.co/usedot/Dot-Loom-Conductor-14B}
}

License

The adapter is released under Apache 2.0. The pinned upstream base model is also published under Apache 2.0. Users must comply with the upstream license and any applicable provider terms.