Technical report: learned budget-constrained multi-model routing
Dot R&D ยท July 2026
### Abstract Dot Loom Conductor 14B is a rank-32 LoRA adapter for `mistralai/Ministral-3-14B-Base-2512`. Given a task, three anonymous candidate-model profiles, and hard limits on calls, credits, and latency, it emits a structured Lean, Balanced, or Strict execution plan. The plan assigns writer, reviewer, and finalizer roles, specifies an access graph, and declares expected quality, pass rate, cost, and latency. We trained the adapter on 9,000 synthetic examples, selected the final checkpoint using 900 family-disjoint validation examples, and evaluated it on 1,200 examples from eight unseen task families. The corpus contains no user prompts or production conversations. Labels come from an exhaustive search over 15 possible role assignments using a disclosed simulator calibrated to a frozen six-case Dot Loom receipt. The raw adapter achieved 99.7% policy accuracy and 90.2% exact-plan accuracy. A deterministic runtime guard caught all 13 hard-budget violations, producing 100% constraint compliance and 90.8% exact-plan accuracy in the guarded lane. The benchmark measures routing-plan generation, not downstream answer correctness. ### Research question Can a locally hosted 14B model learn to allocate inference depth across one-model, writer-reviewer, and writer-reviewer-finalizer plans while respecting explicit operational budgets? ### Model and training - Base: `mistralai/Ministral-3-14B-Base-2512` - Base revision: `5b0ceedbb42dff466ae60b258ba296f32da51384` - Adapter: BF16 LoRA, rank 32, alpha 64, dropout 0.05 - Targets: attention and MLP projection layers - Trainable parameters: 139,984,896 of 14,085,016,576 - Hardware: one NVIDIA H200 SXM5 with 143,771 MiB reported VRAM - Training: 2 epochs, 564 optimizer steps, effective batch size 32 - Learning rate: `1.5e-4`, cosine decay, 28 warmup steps - Seed: `20260716` - Duration: 52.6 minutes - Selected checkpoint: step 564, validation loss `0.070168` - Adapter SHA-256: `e0e6655a16a10f28cbce898e564640bf6a4a64ca84bb2014a1da3a60c5f11eda` ### Corpus and oracle The v2 corpus contains 11,100 examples: 9,000 train, 900 validation, and 1,200 test. Each split is balanced across Lean, Balanced, and Strict labels. Task families are disjoint across splits. For each example, the generator enumerates three writer-only plans, six writer-reviewer plans, and six writer-reviewer-finalizer plans. It rejects plans exceeding hard call, credit, or latency limits, estimates task-adjusted quality and reliability, applies cost and latency penalties, and selects the highest-utility feasible plan. If any feasible plan meets the requested quality target, below-target plans are excluded. An independent OpenAI judge reviewed 90 stratified train and validation labels while provider and worker names remained hidden. The judge agreed with 76 labels and proposed 14 alternatives. Deterministic validation found only three alternatives feasible; none improved the disclosed oracle. The judge is an independent critic, not ground truth. ### Held-out results | Lane | Policy accuracy | Exact plan | Constraints met | Unsafe under-escalation | Mean regret | |---|---:|---:|---:|---:|---:| | Deterministic Loom | 93.9% | 79.3% | 100.0% | 5.7% | 1.733 | | Raw Ministral 14B Base | 0.4% | 0.1% | 0.1% | 98.1% | 1051.175 | | Dot-trained raw adapter | 99.7% | 90.2% | 98.9% | 0.0% | 11.823 | | Dot-trained plus guard | 100.0% | 90.8% | 100.0% | 0.0% | 0.245 | Compared with deterministic Loom, the guarded conductor improved exact-plan accuracy by 11.5 percentage points, paired-bootstrap 95% CI 9.4 to 13.7. The exact-plan McNemar two-sided p-value was below `1e-12`. The runtime guard fell back on 13 of 1,200 proposals. The raw Base checkpoint is a pre-training capacity control, not a comparison with an instruction-tuned assistant. Deterministic Loom is the operational baseline because it receives the same structured task, worker, and budget information. ### Deployment boundary The learned model proposes a policy. Deterministic code must parse the JSON, recompute calls, credits, latency, access, and provider independence, reject invalid role assignments, and fall back when a hard constraint is violated. The guarded numbers include this external enforcement. The guard is not encoded in the adapter weights. ### Limitations - The labels depend on a disclosed simulator, not execution of every candidate plan. - The empirical calibration contains six backend cases. - The study uses one seed, one base checkpoint, and one training run. - Synthetic task descriptors are simpler than production prompts. - Worker measurements may drift as providers and model versions change. - Routing-plan accuracy does not establish downstream answer quality. - Instruction-tuned conductor baselines and multi-seed replication remain future work. ### Artifacts - [Model and adapter](https://huggingface.co/usedot/Dot-Loom-Conductor-14B) - [Dataset](https://huggingface.co/datasets/usedot/dot-loom-conductor-v2) - [Reference policy explorer](https://huggingface.co/spaces/usedot/Dot-Loom-Conductor-Lab) - [Source, tests, raw predictions, and receipts](https://github.com/usedotai/dot-loom/tree/main/research/ministral-conductor) ### Citation ```bibtex @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} } ```