Maestro Flash is a design proposal, not a trained model. It has no weights and no model results. The architecture, the planned 90-day work and a small central processing unit (CPU) experiment are set out below to make the proposal testable, and to keep a pipeline demonstration separate from a capability claim. The full design, with a source for each choice, is in the Maestro Flash technical report.
First principles
A typed decision is a probability over a closed set of outputs: a label, a bounded score, a true-or-false value, or a span of the input. Nothing has to be generated, so one forward pass can answer it. The design follows from what that one pass has to hold.
Everything shares one input budget. An encoder reads a fixed number of tokens. When options are written into the same sequence as the state, each added option takes room from the state and from the other options. That is the shape of Laya’s reported Banking77 limit. Encoding options separately, and caching them, removes the competition.
Ordered labels carry distance. On a five-point scale, predicting 4 when the truth is 5 is a smaller error than predicting 1. An unordered softmax treats both errors the same. A cumulative-link head with K-1 thresholds for K levels builds the order into the output.
Structure can be assembled instead of generated. A JavaScript Object Notation (JSON) object is a tree, and in a declared schema each leaf can be filled by one of the typed decisions above. Predicting the leaves in parallel and joining them by rule gives valid structure without a token loop.
A probability is only useful if it is honest. Confidence decides which outputs can run without review. So calibration is part of the pipeline: training against proper scoring rules, then temperatures fit on held-out data for each question type, option count and language.
Architecture proposal
The proposed production backbone is mmBERT-base, 322M parameters and 1,800+ languages, with Ettin-encoder-400M as an ablation candidate. The selection is not a claim that either will win. mmBERT provides multilingual coverage; Ettin provides a controlled encoder-versus-decoder precedent in arXiv:2507.11412. The plan is to choose from recorded stage-0 measurements rather than assume an outcome.
One shared encoder trunk would support four lightweight heads:
| Head | Proposal |
|---|---|
choice |
Encode options separately and cache static option representations, so option-cardinality does not consume the state-context budget. This directly tests a fix for Laya’s reported Banking77 0.425-0.492 limit against Jev’s published 0.870. |
score |
Use cumulative-link ordinal regression with K-1 cumulative logits for K levels, rather than unordered softmax. The target is Laya’s reported 0.372 on the five-class Stanford Sentiment Treebank (SST-5), a weak result. |
span |
Use start/end pointers over the state for extractive fields and supporting text. |
json |
Decompose schema leaves into choice, score, noul (a calibrated true-or-false probability) and span decisions, predict them in parallel, then deterministically assemble JSON. Nested arrays of objects are out of scope for v1. |
The data plan begins with public task-shaped datasets, adds IndicTrans2 translation into 11 languages, adds transliteration in the style of MuRIL (Multilingual Representations for Indian Languages), and reserves a human-verified subset. Machine-translated training text is not equivalent to native text, so native-text evaluation is the decision metric. The proposal calls for per (question type, option count, language) temperatures, extending the calibration grouping used by Laya.
Training is staged: supervised multi-task fine-tuning; soft-distribution distillation from Qwen3 or Gemma 3; proper-scoring-rule reinforcement learning (RL) with a group-relative baseline in the style of Group Relative Policy Optimization (GRPO); then held-out temperature fitting. The 8 × H100 × 168 hours = 1,344 H100-hours planning allocation is 96 for backbone ablation, 384 supervised fine-tuning, 192 data generation, 240 distillation, 288 RL, 96 calibration and evaluation, and 48 buffer. It is a budget, not a measured cost.
The evaluation plan measures accuracy, soft accuracy, Brier score, expected calibration error (ECE), ordinal mean absolute error (MAE), median (p50) latency, and option-order flip rate across MASSIVE, cross-lingual natural language inference (XNLI), Banking77, AG News, DAIR Emotion, an application-workflow benchmark, and a proposed Indic benchmark. All values are TARGET measurements to be collected, not results.
90-day plan
Days 1–7 establish licensing, a minimal reproducible skeleton, and a shared harness. The first exit condition is that the harness can approximately reproduce Laya’s published figures on its own checkpoints. If it cannot, the harness is fixed before training anything new.
Days 8–21 implement the four heads, compare mmBERT-base with continued-pretrained Ettin-encoder-400M, and start the Indic data pipeline. The Day 18 decision records both winner and loser. Days 22–42 run supervised fine-tuning and distillation. The Day 28 go/no-go is that Banking77 must exceed Laya’s 0.425–0.492 range; otherwise the proposed option encoding is root-caused rather than carried forward.
Days 43–63 run proper-scoring-rule RL and calibration ablations. The Day 63 criterion is lower post-refit ECE at matched accuracy than the stage-1-only checkpoint. A negative result is reported as such. Days 64–84 build the Indic typed-decision benchmark for Hindi, Bengali, Marathi, Tamil, Telugu, and Kannada, with an extended lower-verification tier for Gujarati, Urdu, Punjabi, Odia, and Malayalam. Days 85–90 prepare the benchmark report, model card, code, and a release decision based on the accumulated criteria rather than an aggregate comparison.
First CPU experiment
The first experiment is deliberately small: a numpy-only bag-of-words plus softmax choice head on 80 local AG News training rows and 40 test rows, using 400 vocabulary tokens. It ran in a fresh Python 3.9.6 environment with numpy 2.0.2, on 2026-09-23. It is a pipeline skeleton, not Maestro Flash.
The exact logged setup was 60 gradient-descent steps, lr=0.5, l2=0.001, CPU only. Its held-out report split had n=20. Accuracy was 0.5000, compared with the 0.2500 four-class random baseline. Brier score changed from 0.7013 to 0.6323; ECE changed from 0.1982 to 0.1143; a separate calibration split selected T = 0.25. Wall-clock time was 0.02s.
These values are real logged results for the toy program only. It used no pretrained encoder, no graphics processing unit (GPU), and no Maestro Flash weights. The calibration split has only n=20, so the fitted temperature is illustrative rather than a statistically meaningful estimate. The experiment demonstrates data loading, typed output, proper-score evaluation, and held-out temperature calibration. It says nothing about the performance of the proposed 322M-400M model.
Constraints
The proposal relies on several hypotheses: decoupled option encoding may fix high-cardinality choice failure; cumulative-link heads may improve ordinal scores; one multilingual backbone with language-aware calibration may reduce routing complexity; and non-autoregressive group construction may make GRPO-style RL useful for encoder decisions. These are experiments to run, not conclusions.
Relevant background includes mmBERT, arXiv:2509.06888, Ettin, arXiv:2507.11412, IndicTrans2, arXiv:2305.16307, GRPO, arXiv:2402.03300, proper scoring rules for language generation, arXiv:2405.18906, and Reinforcement Learning with Calibration Rewards (RLCR), arXiv:2507.16806.