Week 8 – Full Dashboard, iOS App, and Gating

Abstract

Shipped a complete dashboard with charts and real-time updates, an iOS/iPadOS companion app, and the first gating experiments.

Summary

The busiest week in terms of features shipped. The dashboard went from basic CRUD to a full monitoring platform with charts and live updates. An iOS companion app was built from scratch. And the first gating mechanism (inspired by DeepSeek’s Engram) was implemented and tested.

Dashboard v2

Major dashboard improvements:

• Live charts: Real-time CE and accuracy plots per generation/iteration
• Tooltips: Hover over data points for detailed genome statistics
• WebSocket: Full duplex updates — experiments push progress to the dashboard
• Flow templates: Pre-configured optimization flows for common setups
• Stop/restart: Graceful shutdown chain with checkpoint save before stopping

The data model was redesigned (v2) with the database as the source of truth, removing the intermediate “Phase” abstraction layer.

iOS/iPadOS App

Built a SwiftUI companion app for monitoring experiments on the go:

• Experiment iteration lists with expandable detail sections
• Charts for tracking optimization progress
• Flow management (create, stop, resume)
• iPad adaptive layouts with 10-column iterations view
• Connects to the same FastAPI backend as the web dashboard

RAM-based Gating

Implemented content-based gating inspired by DeepSeek’s Engram architecture:

• RAMGating: Dedicated RAM neurons learn which clusters should be active for each input context
• Majority voting: N neurons per gate, gate opens if majority fires TRUE
• Staged training: Train RAM first (freeze), then train gating (Stage 2)
• Rust acceleration: rayon parallelization achieving 16x speedup

The gating model observes the same input bits as the main model but produces a binary mask over clusters. This is the RAM equivalent of attention-based gating.

Initial results with gating on the tiered architecture were mixed — it didn’t consistently improve over ungated scores. The hypothesis is that 50K+ clusters (one per token) is too many for the gating model to learn meaningful patterns. This motivates the bitwise architecture developed later.

Other Improvements

• Normalized harmonic fitness calculator
• Accuracy floor for GA/TS (prevents degenerate genomes)
• Checkpoint resumption improvements
• Memory leak fixes in Metal evaluators with buffer pooling

Next

The gating underperformance with tiered architecture, combined with data density insights, points toward a fundamentally different approach — the bitwise architecture with only 16 clusters.