Skip to main content
AI JUL 02, 2026

Leverage Record: July 2, 2026

Three tasks. July 2, 2026 weighted to 10.3x leverage across 32.0 human-equivalent hours in 187 Claude-minutes. Supervisory leverage closed at 61.9x. A smaller, infrastructure-heavy day: a structured-output benchmark acr…

Three tasks. July 2, 2026 weighted to 10.3x leverage across 32.0 human-equivalent hours in 187 Claude-minutes. Supervisory leverage closed at 61.9x. A smaller, infrastructure-heavy day: a structured-output benchmark across several candidate models that surfaced three live bugs, overnight kernel-panic incident recovery, and synthesis-pipeline parallelization and decoupling.

0.8 weeks of human-equivalent throughput in 3.1 hours of Claude wall-clock. The 20.0x ceiling came from Built and ran a structured-output benchmark across several candidate models (via an internal inference service); found live bugs in inference-auth, forced...; the 7.6x floor sat at Synthesis-pipeline work: dual-run six-way parallelization, decoupled the job-orchestration service from the inference engine (directory-path config fix across....

About These Records
These time records capture personal project work done with Claude Code (Anthropic) only. They do not include work done with ChatGPT (OpenAI), Gemini (Google), Grok (xAI), or other models, all of which I use extensively. Client work is also excluded, despite being primarily Claude Code. The actual total AI-assisted output for any given day is substantially higher than what appears here.

Task Log

#TaskHuman Est.ClaudeSup.FactorSup. Factor
1Built and ran a structured-output benchmark across several candidate models (via an internal inference service); found live bugs in inference-auth, forced tool-choice temperature handling, and a cost-accounting cache path10.0h30m8m20.0x75.0x
2Diagnosed an overnight kernel panic (watchdog timeout from out-of-memory conditions during parallel synthesis) and safely restarted an interrupted 14-package academic-curriculum synthesis batch: offloaded models to shared NLI/embedding servers, fixed runtime-mode and directory-path configuration, cleaned up orphaned jobs, and added a serial orchestrator with memory watchdog, checkpoint/resume, and archive backup8.0h47m3m10.2x160.0x
3Synthesis-pipeline work: dual-run six-way parallelization, decoupled the job-orchestration service from the inference engine (directory-path config fix across 11 scripts, plus test and completeness-gate fixes), and drafted an isolation-architecture plan for the platform14.0h110m20m7.6x42.0x

Aggregate Statistics

MetricValue
Total tasks3
Total human-equivalent hours32.0
Total Claude minutes187
Total supervisory minutes31
Total tokens980,000
Weighted average leverage factor10.3x
Weighted average supervisory leverage factor61.9x
Human-equivalent weeks0.8

Analysis

The day's leverage distribution matters more than the headline figure. The 20.0x ceiling came from Built and ran a structured-output benchmark across several candidate models (via an internal inference service); found live bugs in inference-auth, forced...; the 7.6x floor was Synthesis-pipeline work: dual-run six-way parallelization, decoupled the job-orchestration service from the inference engine (directory-path config fix across.... Tasks at the top of the distribution share a shape: tightly-scoped specifications, clear success criteria, and minimal integration ambiguity. The AI doesn't need to discover anything new; it executes against an explicit target.

Tasks at the bottom run differently. They're either bounded by review-heavy work where every step gets verified, or they involve ambiguity that demands several rounds of trial and adjustment. The factor is real and informative, not a failure mode.

The lower task count and the weight of debugging and incident recovery pulled the day's leverage down; recovery and pipeline decoupling are iterative and verification-heavy, the classic lower-leverage shape.

The supervisory leverage figure (61.9x today) tracks something orthogonal to wall-clock leverage. It's the ratio of human-equivalent output to human prompt-writing time. It stays high even on lower-leverage days because supervisory minutes scale with task count, not with the human-hour estimate; a 20-minute task and a 4-hour task can both be specified in two minutes of human prompt-writing.

Across the 3 tasks, the day produced roughly 0.8 weeks of senior-engineer-equivalent throughput in 3.1 hours of model wall-clock. That ratio is the practical answer to the question of how much output a single operator can move per day when the model handles the execution and the operator handles the direction.