Daily Report — 2026-08-12

Daily Overview

• What was done: Resolved critical methodological flaws in scientific reporting, fixed hardware-software integration issues in macOS applications, optimized reinforcement learning policies through advanced curriculum strategies, and established reliable pipelines for integrating external VLM capabilities into robotic evaluation frameworks.
• How it was done: Conducted code-level audits to identify clustering algorithm confounds, performed empirical brightness sweeps on MacOS devices, implemented periodic gait rewards and action space reduction in C++/Python RL environments, and developed strict coordinate conversion layers for cross-API geometric data integration.
• Impact: Ensured the scientific integrity of the MIHD paper by removing invalid gain claims; restored functionality to Amber’s comfort logic; achieved a significant performance leap (46s) in QWOP tasks; and enabled precise visual prompt injection for RoboMemory using only topologically accurate models like Gemini-3.1-pro.

MacOS

• What was done: Audited DLPFC clustering results, identified mclust/KMeans confounds, and verified statistical significance limits; diagnosed and fixed Amber’s backlight reading by deriving nits from DisplayServices APIs.
• How it was done: Executed Python scripts for CSV analysis, performed search-and-replace refactoring documents, ran empirical brightness sweeps, and updated CLAUDE.md documentation.
• Impact: Validated core scientific claims, corrected factual errors in statistical reporting, and restored accurate absolute brightness context for physiological comfort logic.

TzJsDesktop

• What was done: Primary device for QWOP RL training, C++ compilation, and reward profiling; conducted AI vs human comparison analysis.
• How it was done: Used Claude Code to edit source files, run self-tests, launch PPO training processes via Bash, and perform black-box system identification on physics DLLs.
• Impact: Enabled the breakthrough in upright locomotion by fixing entropy collapse bugs and implementing posture termination constraints.

lighthouse

• What was done: Analyzed Qualcomm-proj codebase structure for VLA quantization; managed cross-device API connectivity for RoboMemory’s Tianhe3 deployment.
• How it was done: Explored directory structures, updated CLAUDE.md with build constraints, configured SSH reverse proxies for API access, and validated JSON sidecar schemas.
• Impact: Improved AI onboarding for Qualcomm VLA quantization and enabled offline VLM trajectory generation bypassing Tianhe3’s network restrictions.

Achieved critical breakthroughs across four major projects: corrected a fundamental confound in the MIHD paper’s statistical claims, fixed a hardware calibration bug in the Amber app, reduced QWOP RL completion time by 60% via curriculum learning, and validated VLM-based trajectory generation for RoboMemory.

Tasks

Architecture & Strategy

• ✅ MIHD Paper Results Restructuring and Confound Resolution — Identified that KMeans silent fallback invalidated fusion gains; restructured paper to dataset-major framework; defined scIB benchmark style.
• ✅ QWOP RL Optimization via Curriculum Learning and Action Reduction — Reduced action space from 16 to 9 actions; implemented periodic gait rewards, posture termination, and sprint-to-distance curriculum to achieve 46s completion.
• ✅ Amber Backlight Reader Refactor and Hardware Calibration — Replaced frozen IORegistry key with live-derived nits model; validated DisplayServices API linearity.
• ✅ RoboMemory VLM Drawer Pipeline and Sidecar Integration — Implemented pipeline for generating trajectory polylines via Gemini-3.1-pro; resolved Tianhe3 API connectivity issues.
• ✅ Multi-Model VLM Shootout for Visual Prompts — Evaluated GPT-5.2, Claude Opus 4.6, and Gemini-3.1-pro for geometric accuracy in visual prompt generation.
• ✅ MIHD Pipeline Decoupling and Repository Audit — Refactored evaluation loop to decouple clustering from fusion; audited repository structure and updated documentation.

Implementation & Fixes

• ✅ Qualcomm VLA Quantization CLAUDE.md Update — Updated documentation with correct build commands, normalization constraints, and known pitfalls.

Problems & Solutions

Critical Issues

1. Critical experimental confound in MIHD: claimed fusion gains were actually due to mclust being used for fused methods and KMeans for baselines.

Solution: Audited code to confirm silent fallback; re-scored methods with consistent KMeans baseline; corrected claims to ‘statistically indistinguishable’.

Key Insight: Algorithmic parity is essential for causal claims about feature fusion; KMeans serves as the necessary common ground across all embedding types.

2. Amber app failed to read absolute backlight brightness due to frozen boot-time IORegistry values, rendering comfort logic inert.

Solution: Validated DisplayServicesGetLinearBrightness as live; derived a linear model (nits = linear * factor) using hardware metadata for accurate nits estimation.

Key Insight: Modeled estimates from consistent internal APIs are superior to broken sensor readings when physical constraints are clearly defined.

3. PPO training instability in QWOP due to entropy collapse and mismatched target entropy during policy transfer.

Solution: Implemented entropy floor, clamped LR adjustments, and aligned target_entropy with current policy entropy for fine-tuning stages.

Key Insight: Entropy targets must be contextual: high for cold starts, low for preserving sharp policies in fine-tuning.

4. VLM-generated geometric paths often had correct syntax but wrong topology; Tianhe3 lacked direct API access.

Solution: Enforced strict coordinate conversion and validation layers; selected Gemini-3.1-pro for topological accuracy; offloaded generation to sidecars via SSH proxy.

Key Insight: Direct visual interpretation requires explicit geometric constraints; pre-computation bypasses network restrictions for fixed-demo tasks.

5. Statistical overconfidence: Holm correction interpreted as unanimous support despite small sample sizes (n=11).

Solution: Corrected documentation to state ’near-unanimous’ and clarified lower bounds of Wilcoxon tests for small N.

Key Insight: Small-sample hypothesis testing is extremely conservative; non-significance does not equal no effect.

General Issues

6. Documentation drift and outdated architectural descriptions in large repositories (MIHD, Qualcomm-proj).

Solution: Performed comprehensive folder-level inventories and code walks; updated documentation to reflect current state and reading orders.

Key Insight: Automated checks or explicit versioning is necessary to prevent architectural depreciation in evolving codebases.

Human vs AI Approaches

Experimental Design and Statistical Interpretation

Role	Approach
Human	Spotted the methodological flaw (confounding variables) early; demanded re-scoring with KMeans; interpreted statistical results cautiously based on physical constraints.
AI	AI initially focused on surface-level metrics; relied on human prompting to run audit scripts; correctly identified model limitations but missed structural flaws until prodded.

Difference Analysis: Human demonstrated superior experimental design intuition and scientific rigor, fundamentally saving key claims from validity issues.

Architectural Decision Making (Amber & RoboMemory)

Role	Approach
Human	Defined physical limits for Amber brightness; insisted on ‘direct drawing’ (polyline) over image generation for RoboMemory based on frozen tower constraints.
AI	AI integrated hardware metadata into models; explored geometric point lists after human guidance; correctly identified topological failure modes in VLM outputs that AI metrics missed.

Difference Analysis: Human provided decisive architectural constraints and qualitative interpretation of complex failures, while AI handled tactical implementation and data synthesis.

RL Strategy Formulation

Role	Approach
Human	Proposed upright running goals; directed curriculum learning strategies based on observed entropy patterns.
AI	Conducted black-box system identification to reduce action space; implemented periodic gait rewards and posture termination logic.

Difference Analysis: AI efficiently executed complex mechanical constraint analysis and reward shaping implementations, while human guided the strategic direction of curriculum design.

AI Limitations

Critical Limitations

• Failed to identify clustering algorithm confounds and statistical nuances without explicit human prompting and verification scripts.
• Initial VLM evaluation relied on pass/fail metrics, missing the critical ’topologically wrong but syntax correct’ failure mode which human interpretation highlighted.

General Limitations

• Struggled with distinguishing live vs absolute sensor data initially; required manual intervention for SSH timeouts and process management on remote servers.
• Go-Explore implementation regressed performance by using random exploration instead of policy-guided sampling for sequence optimization.

Learnings

Key Learnings

• In small-sample hypothesis testing, Holm correction is extremely conservative; non-significance should be interpreted as ’not unanimous’ rather than ’no effect’.
• Curriculum learning with short-horizon sprint policies provides significant acceleration advantages for long-horizon robotic tasks like QWOP.
• Zero-shot VLMs struggle with precise geometric topology; direct drawing is viable only with strict coordinate conversion and topological validation.
• When hardware sensor data is unreliable, deriving values from consistent internal API metrics provides a robust modeled truth superior to raw readings.

Practical Learnings

• Decoupling benchmark steps (e.g., fusion vs clustering) allows independent testing and reuse of intermediate states without re-running upstream costs.

Conversation Summaries

MIHD Spatial Omics Framework

✅ Statistical Confound Resolution and Paper Restructuring 19:53:21.273 | claude_code / cursor Resolved a critical clustering algorithm confound (KMeans vs mclust) that invalidated previous fusion gain claims. Restructured the paper’s results section to a dataset-major framework. Decoupled clustering from fusion in the pipeline. Audited repository documentation for drift.

Amber macOS App

✅ Hardware Calibration and Backlight Logic Fix 19:53:21.273 | claude_code Diagnosed invalid backlight nits readings from frozen IORegistry keys. Used empirical sweeps to validate DisplayServices APIs and implemented a live-derived linear model for accurate brightness context, restoring physiological comfort logic.

QWOP RL Training

✅ RL Policy Optimization via Curriculum Learning 15:30:00.000 | claude_code Achieved upright locomotion and reduced completion time to 46s by reducing action space, implementing periodic gait rewards, fixing entropy collapse in PPO, and applying a sprint-to-distance curriculum.

RoboMemory

✅ VLM Visual Prompt Integration and Evaluation 18:05:29.358 | claude_code Implemented a ‘direct drawing’ pipeline using Gemini-3.1-pro to generate trajectory polylines. Resolved Tianhe3 API connectivity issues via SSH proxies. Validated that only topologically accurate models work for geometric precision tasks.

Qualcomm VLA Quantization

✅ Documentation and Build Configuration 01:43:48.920 | claude_code Updated CLAUDE.md with accurate build commands, normalization constraints, and latency testing protocols to improve future AI agent onboarding.

Token Usage