Senior ML Architect / Research Lead

• primitive composition;

Senior NPTA ML Architect / Research LeadRole Purpose

Lead the design, training and technical validation of the Neural Pass Transition Array (NPTA), a purpose-built approximately 350M-parameter neural inference core used within the Surface inference architecture.

The NPTA is not a conventional language model. It is a trained neural transition fabric designed to execute compact, machine-level inference primitives over bounded structured state supplied by Surface.

The role is responsible for turning the NPTA functional specification into a trainable neural architecture capable of reliable atomic state transitions, context-boundary recognition, semantic discrimination and compositional reasoning.

The successful candidate must be comfortable treating transformer-derived neural architectures as programmable computational substrates rather than conversational models.

Core ResponsibilitiesNPTA Architecture

Own the neural architecture of the NPTA, including:

• parameter allocation within an approximately 350M-parameter target;
• layer count, hidden dimension and feed-forward dimensions;
• attention structure;
• RMSNorm and activation architecture;
• embedding and output-head structure;
• NCL vocabulary integration;
• output-region topology;
• intermediate representation design;
• training-only auxiliary heads;
• quantisation considerations;
• scaling experiments across smaller NPTA variants.

Establish experimentally whether the target architecture should remain close to the proven 28-layer transformer-derived baseline or diverge where NPTA workloads justify it.

Capability Decomposition

Work with the Surface architect to define which operations require learned neural inference and which must remain within deterministic Surface code.

Develop the atomic NPTA capability hierarchy covering areas including:

• relational inference;
• entity and binding resolution;
• logical and constraint inference;
• candidate discrimination;
• temporal and causal inference;
• spatial and qualitative physical reasoning;
• context sufficiency;
• missing-context identification;
• PCS virtual-row integration;
• result resolution.

Prevent conventional LLM responsibilities such as formatting, persistent memory, query planning, counting, retrieval and output management from migrating unnecessarily into NPTA weights.

Training Objective Design

Design training objectives appropriate to exact state-transition behaviour rather than natural-language generation.

Responsibilities include:

• primary token/symbol loss;
• canonical-output training;
• context-sufficiency objectives;
• context-miss classification;
• output-region classification;
• candidate-region contraction;
• intermediate-state supervision;
• auxiliary representation losses;
• failure-state supervision;
• long-horizon compositional training.

Training Curriculum

Define curriculum progression from simple primitive transitions to complex Surface-managed trajectories.

Expected stages include:

• syntax/state validity;
• single primitives;
• transition boundaries;
• missing-context recognition;
• virtual-row restoration;
• distractor resistance;
• primitive composition;
• cross-family composition;
• long-horizon trajectories;
• adversarial states;
• hard-case replay.

Model Scaling

Design and evaluate staged NPTA variants such as:

• 10M;
• 30M;
• 70M;
• 150M;
• 250M;
• 350M parameters.

Determine capability saturation, training-data requirements and optimal parameter allocation empirically rather than assuming conventional language-model scaling relationships apply.

Surface and DSP Co-Design

Work closely with Surface runtime and DSP engineers to ensure the trained model exposes useful internal structure at designated observability locations.

Investigate training methods that improve separability or predictability at:

• matrix multiplication boundaries;
• RMSNorm boundaries;
• linear input/output points;
• candidate narrowing stages;
• token formation;
• output-region resolution.

Ensure the neural design remains compatible with eventual bare-metal and ASIC execution.

Required Experience

Essential:

• substantial experience designing and training transformer or transformer-derived models from scratch;
• strong understanding of attention internals;
• RMSNorm and modern feed-forward architectures such as SwiGLU;
• embedding and output-head geometry;
• representation learning;
• curriculum learning;
• loss-function design;
• synthetic training;
• model scaling;
• training stability;
• gradient behaviour;
• quantisation;
• PyTorch or JAX;
• experience analysing intermediate neural representations.

Candidates whose experience is limited primarily to API use, prompt engineering, RAG, RLHF, instruction tuning or model fine-tuning will not have the required technical depth.

Highly Desirable Experience

• training compact high-performance models;
• structured reasoning architectures;
• recurrent or bounded-state neural systems;
• adaptive computation;
• hierarchical output heads;
• model/hardware co-design;
• deterministic or highly constrained decoding;
• neural architecture research;
• synthetic-world training;
• formal or symbolic/neural systems;
• CPU inference optimisation;
• ASIC-oriented ML design.

Technical Mindset

The successful candidate must be comfortable with the principle:

If an operation can be performed exactly and cheaply by Surface, a conventional processor, PCS or fixed logic, it should not automatically be learned by NPTA.

The role therefore requires judgment about where neural computation provides genuine value.

First 6-Month Deliverables

• NPTA architecture specification v1;
• initial NCL neural symbol interface;
• first 10M–30M experimental NPTA;
• capability scaling experiments;
• context-miss and PCS continuation demonstration;
• 70M-class compositional prototype;
• initial output-region training experiment;
• defined path to the 350M production candidate;
• documented architecture decisions and ablation results.

Success Measures

Success will be measured primarily through:

• exact transition correctness;
• unseen-state generalisation;
• primitive composition reliability;
• context-boundary accuracy;
• PCS continuation accuracy;
• robustness to irrelevant context;
• long-horizon trajectory survival;
• parameter efficiency;
• compute efficiency;
• suitability for Surface and DSP control.

Conventional conversational benchmark performance is not a primary objective.