Laya-LoRA Coding Companion: Asymmetric Dual-Process Test-Time Compute, Full-Scale Curriculum on 10 Structured Coding Datasets, and Serverless Multi-Tier Inference Infrastructure on GCP (Camera-Ready Version 3)

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We present the Laya-LoRA Coding Companion and GWAYA dual-process architecture, an open-weight, parameter-efficient System 1/System 2 compound AI system for autonomous code intelligence, quality assurance, and zero-trust anti-hallucination verification. Operating over a 149M-parameter ModernBERT-base encoder adapted via Low-Rank Adaptation (LoRA, rank r=8, α=16) on fused query-key-value projections and four multi-task task heads (578,353 trainable parameters, 0.388% of backbone), Laya achieves microsecond non-autoregressive decision inference (∼45 ms CPU latency). Formally verified under Lean 4 (lake build ANSE) ensuring parameter budget compliance (|Θ| ≤ 600,000, Invariant I2), the model was trained across an exhaustive 10-dataset curriculum comprising 78,503 structured records spanning security vulnerabilities (PyCode-Vul), code smells (SmellBench), unit testing (CodeRM-UnitTest), execution efficiency (EffiBench-X, SWE-Perf, RAPL Joules), formal theorem proving (Lean-Workbook, miniF2F), and execution trace alignment (CRUXEval, Magpie). Coupled with Qwen3.8-27B and serving as a parallel verifier for frontier agent outputs (Gemini 3.1 Pro & 3.8 Flash), the system achieves 96.0% overall gate accuracy (Wilson 95% CI: [86.5%, 98.9%]), 100.0% threat recall (25/25 malicious patterns blocked, zero false negatives), and resolves 54.0% of incoming queries on the fast reflex path, yielding an aggregate energy consumption of 11.48 Wh per 1,000 queries (—53.2% energy reduction compared to standalone 27B autoregressive generation at 24.50 Wh). This reproducibility package includes the camera-ready 12-page manuscript (incorporating Reviewer 4 Meta-Assessment resolutions), full XeLaTeX source code, verified benchmark receipts, formal Lean 4 verification proofs, GWAYA verifier & advisor integration code, model weights, and complete SHA-256 cryptographic manifests.