Monthly Summary — 2026-03

March 2026 was a transformative month defined by massive architectural shifts across robotics, bioinformatics, and desktop engineering. The month saw a successful transition from legacy, monolithic, and platform-dependent structures—such as macOS-specific UI components and symlink-heavy genomic repositories—to modular, cross-platform, and high-throughput frameworks. Key advancements included the evolution of robotics policies through the Error-Recovery Benchmark v5 and VLA training optimizations, the stabilization of spatial transcriptomics pipelines via joint-training architectures, and the successful migration of the TokenMonitor desktop suite to a Tauri/Svelte cross-platform stack. Despite challenges involving silent ML failures, HPC resource contention, and complex IPC synchronization, the month concluded with the establishment of robust, production-ready, and highly automated scientific and agentic workflows.

Monthly Overview

Metric Value
Active Days 29 / 31
Total Conversations 104
Projects 97
Tasks Completed 136
Tasks In Progress 16
Total Tokens 1,766,867,959
Total Cost $1,010.08
Claude Code Token 1,405,670,771
Claude Code Cost $806.68
Codex Tokens 361,197,188
Codex Cost $203.40
Daily Average Cost $33.67

Project Progress

Error-Recovery Benchmark & VLA Training (25 days active) — 🔄 active

Development of robotics policy evaluation pipelines (Pi0.5/Phoenix) and v5 Error Skill architecture. Includes distributed training on A800 clusters, implementing semantic error skills, and transition from physics-force injection to quota-driven scheduling/context replay.

Key Milestones:

  • Implementation of BatchedVLAServer
  • Successful 8-GPU training orchestration
  • Resolution of BC-RNN success rate collapse
  • Implementation of 11 semantic skills
  • Integration of context replay engine
  • Development of quota-driven distribution scheduler
  • Gradient scaling stabilization for VLA loss
  • Architecting progressive conditioning experiment matrix
  • v4 legacy archival and v5 architecture design
  • Deterministic scene generation implementation
  • Successful BOSS/LIBERO integration
  • Unified evaluation logging architecture

MIHD Spatial Transcriptomics Pipeline (21 days active) — 🔄 active

Optimization of multi-modal fusion pipelines (scGPT/UNI2/STAIG) for spatial omics, focusing on coordinate mapping accuracy, embedding variance conservation, and transitioning from independent feature reduction to joint-training alignment to prevent mathematical divergence.

Key Milestones:

  • Restoration of ARI metrics from near-zero
  • Deployment of two-stage embedding/fusion pipeline
  • Validation of cross-tissue RM-IDEAL benchmarks
  • Global symlink removal and absolute path migration
  • Standardization of DLPFC/HD/RM-IDEAL output hierarchy
  • Alignment failure diagnosis
  • KDTree optimization for memory scaling
  • Joint training roadmap establishment
  • Identification of embedding space orthogonality issues

Life Copilot (formerly CalendarPro) (9 days active) — 🔄 active

Transforming a scheduling tool into an autonomous agent featuring session-aware intent routing, persistent state management, and a dual-tier MultiIntent architecture.

Key Milestones:

  • Architecture overhaul to Phased Butler System
  • Integration of dynamic mismatch-driven learning
  • Expansion to 360+ stabilized test cases
  • MultiIntent architecture deployment
  • Codebase cleanup (~2400 LoC)
  • Project rebranding

TokenMonitor & Desktop UI (15 days active) — 🔄 active

Transitioning from macOS-centric development to a cross-platform Tauri/Svelte desktop suite, resolving Win32/WebView2 synchronization and eliminating native dependency friction.

Key Milestones:

  • Resolution of Code 257/sandbox permission errors
  • Implementation of hardware-accelerated NSVisualEffectView masking
  • Full macOS dependency stripping
  • Windows/Linux multi-OS matrix build establishment
  • Stabilization of tray-anchored geometry
  • TokenMonitor v0.6.0 Core Architecture & CI/CD Stabilization

Gadget Research Toolkit & DevOps (14 days active) — 🔄 active

Decomposing monolithic Python research pipelines into a modular, multi-language (Swift, Rust, TS) Hub-and-Spoke framework and managing bilingual documentation.

Key Milestones:

  • Development of two-phase export/merge state machines
  • Automated Hugo deployment pipeline
  • Monolithic pipeline decomposition
  • Implementation of scalable academic-style code-summarization skills
  • Bilingual (English/Chinese) static site architecture deployment

Academic Researcher Profiling Tool (4 days active) — ✅ completed

Developed a modular pipeline for scholarly lineage mapping using LLM-driven ingestion, Semantic Scholar integration, and Bayesian student discovery.

Key Milestones:

  • Deployment of cascading LLM fallback recovery chain
  • CLI Unification
  • Citation graph integration

Key Achievements

  1. Spatial Transcriptomics Performance Restoration (MIHD Spatial Transcriptomics Pipeline) — Recovered ARI scores from near-zero to >0.54 by patching coordinate mapping mismatches, scGPT checkpoint attribute persistence, and unified Visium HD mapping.
  2. VLA Pipeline Throughput & Gradient Optimization (Error-Recovery Benchmark & VLA Training) — Increased GPU utilization from 10% to >60% via BatchedVLAServer and resolved training oscillations by scaling task_pos_weight and protecting primary action gradients.
  3. TokenMonitor Cross-Platform Migration (TokenMonitor & Desktop UI) — Successfully migrated to a Tauri/Svelte architecture, stripping all macOS-specific dependencies and enabling stable Windows/Linux builds.
  4. Gadget Pipeline Modularization & Debugging Skill (Gadget Research Toolkit) — Decomposed a 3000-line monolithic module into a schema-driven micro-package architecture and implemented the ‘cchypothesis’ structured debugging workflow.
  5. Autonomous Butler/Life Copilot Deployment (Personal Butler System (CalendarPro)) — Transitioned CalendarPro to a session-aware autonomous system with a robust MultiIntent routing architecture and 360+ passing tests.
  6. Large-scale Codebase Refactoring (MIHD Repository & Data Restructuring) — Executed massive technical debt elimination, including 2400+ LoC cleanup and the removal of 65+ legacy symlinks across the MIHD repository.
  7. Infrastructure & Secure HPC Workflow (Gadget DevOps & Reporting) — Established secure, idempotent data movement and remote metadata extraction protocols to bypass enterprise network/DNS constraints.

Recurring Problems

1. Silent failure under misconfiguration/schema mismatch (17 occurrences)

Dates:Root Cause: ML frameworks (JAX/PyTorch) and OS APIs defaulting to fallback/zero-success modes when attributes, entitlements, or observation keys (HDF5/robomimic) are missing. Status: ✅ Resolved

2. Resource contention and VRAM starvation (4 occurrences)

Dates:Root Cause: Eager device backend initialization and zombie processes in shared HPC/cluster environments monopolizing memory. Status: ✅ Resolved

3. Hardware Visibility & Access Constraints (8 occurrences)

Dates:Root Cause: Kubernetes/Container PID isolation blocking telemetry and macOS TCC/Sandbox entitlements restricting media and hardware access. Status: 🔧 Workaround

4. Data/Schema misalignment in multi-stage pipelines (11 occurrences)

Dates:Root Cause: Discrepancies in tensor shapes, coordinate scales, or implicit vectorization in data loaders between training and evaluation phases. Status: 🔄 Ongoing

5. IPC and UI lifecycle desynchronization (5 occurrences)

Dates:Root Cause: Race conditions between asynchronous Win32/macOS IPC calls and subsequent DOM reflow/CSS rendering cycles. Status: 🔧 Workaround

6. AI Context/Token Overflow (2 occurrences)

Dates:Root Cause: Massive git diffs and large configuration files exceeding the token window during automated code reviews. Status: 🔧 Workaround

  • Human-initiated insights: 63 items
  • AI limitation patterns: Convention bias toward standard HPC/JAX scheduling/SDK heuristics
  • AI limitation patterns: Inability to anticipate platform-specific native constraints (macOS/Docker/Win32)
  • AI limitation patterns: Lack of proactive memory/resource profiling and environmental auditing
  • AI limitation patterns: Tendency to hallucinate academic/bibliographic lineages without explicit anchors
  • AI limitation patterns: Failure to detect silent numerical/class attribute degradation
  • AI limitation patterns: Difficulty synthesizing large-scale structural migrations without explicit dependency graph instructions
  • AI limitation patterns: Defaulting to sequential execution instead of parallel/data-driven workflows
  • Improvement areas: Proactive infrastructure/filesystem discovery for remote/containerized environments
  • Improvement areas: Mathematical verification of gradient/loss scales and environmental/simulation constraints
  • Improvement areas: Context-aware long-horizon planning and explicit dependency/provenance tracing
  • Improvement areas: Requirement for strict JSON/byte-stream schema validation in generative pipelines
  • Improvement areas: Enhanced awareness of shared resource (GPU/Memory) states and external API rate-tables
  • Improvement areas: Mandatory inclusion of grounding/disambiguation keys in research prompts

Monthly Learnings Digest

Architecture (architecture)

  • Decoupling high-latency feature extraction into cached staging areas is mandatory for controlling multi-modal compute costs. (Source: 2026-03-03)
  • Analytical resilience in probabilistic pipelines demands multi-model cascading fallbacks rather than simple retry loops. (Source: 2026-03-13)
  • True cross-platform robustness requires the complete elimination of native OS bindings (e.g., stripping objc2) rather than using conditional compilation. (Source: 2026-03-25)
  • Enforcing strict hierarchical scoping and modularity in prompt-driven tools is necessary to prevent attention degradation as codebases scale. (Source: 2026-03-25)
  • Remote preprocessing and metadata-first extraction fundamentally transform cross-host synchronization feasibility by reducing payload volume. (Source: 2026-03-29, 2026-03-31)

Debugging (debugging)

  • Physics engine memory layouts and model attributes are not synchronized by state-setting functions; explicit state clearing is required. (Source: 2026-03-03)
  • Modern ML frameworks require explicit TOML-level dependency overrides in isolated environments to ensure C-extension coupling. (Source: 2026-03-11)
  • Distributed workflow state management should rely on temporal expiration checks rather than strict boolean gates to prevent sync stalls. (Source: 2026-03-15, 2026-03-19)
  • Hypothesis-driven workflows with explicit triage and mandatory evidence logging significantly reduce confirmation bias in complex systems. (Source: 2026-03-27, 2026-03-28)
  • Silent ML framework defaults frequently override intended hardware routing; explicit environment flags (JAX_PLATFORMS) are mandatory. (Source: 2026-03-29)

Domain Knowledge (domain)

  • Vision-only models capture visual homogeneity; multi-modal constraints are required to bypass texture bias in biological mappings. (Source: 2026-03-01)
  • In multi-objective training, auxiliary loss head weighting must be explicitly scaled against primary head gradients to prevent signal dilution. (Source: 2026-03-10, 2026-03-14)
  • Cross-modal/cross-sectional retrieval necessitates joint latent space alignment; post-hoc batch correction cannot recover lost dimensional semantics. (Source: 2026-03-16, 2026-03-17, 2026-03-20, 2026-03-23, 2026-03-31)
  • In highly imbalanced classification, direct sample weighting is significantly more stable than complex sampling or focal loss. (Source: 2026-03-19, 2026-03-20)

Tools (tools)

  • Large-VLA models require explicit model parallelism (FSDP/ZeRO-2) and local artifact governance to bypass cluster network constraints. (Source: 2026-03-05)
  • Virtualized GPU environments require direct /proc filesystem inspection for accurate process-to-accelerator mapping. (Source: 2026-03-11)
  • Stabilizing hybrid desktop (Tauri/Svelte) UIs requires absolute viewport anchoring and JS-side pre-layout synchronization. (Source: 2026-03-27, 2026-03-28)
  • SRI integrity in static site generators requires explicit .gitattributes configuration to prevent normalization corruption. (Source: 2026-03-31)

Token Usage Statistics

AI Usage · 2026-03 Claude Code + Codex
Total cost
$1,010.08
Total tokens
1.77B
Output tokens
10M
Cache read
92.5%
Cost split Claude Code $807 · Codex $203
Token character Cache reads 92.5% · Active 7.5%

Most token volume came from cache reads.

Peak Day: 2026-03-07 — $123.56 / 219.1M tokens

Daily Average: $33.67