Six tasks. July 6, 2026 weighted to 8.3x leverage across 72.0 human-equivalent hours in 520 Claude-minutes. Supervisory leverage closed at 86.4x. Model benchmarking and content-generation routing ran the day: a 30-subject three-model synthesis-quality benchmark, an entailment-floor re-validation that rescued 27 of 33 partial packages, a beta-tier expansion from 44 to 106 packages, and per-domain model routing for the retrieval-answering service.
1.8 weeks of human-equivalent throughput in 8.7 hours of Claude wall-clock. The 10.1x ceiling came from Ran a per-subject synthesis-quality benchmark: three candidate models x 30 subjects (13 academic subjects, 17 world-language courses), using a per-phase...; the 3.4x floor sat at Math-content model bake-off for the retrieval-answering service: built a four-model x math-course direct A/B harness (80 MCQs); one candidate model came out....
Task Log
| # | Task | Human Est. | Claude | Sup. | Factor | Sup. Factor |
|---|---|---|---|---|---|---|
| 1 | Ran a per-subject synthesis-quality benchmark: three candidate models x 30 subjects (13 academic subjects, 17 world-language courses), using a per-phase benchmark harness plus an NLI model-server; produced a detailed report with per-subject winners and good/bad example content nodes | 16.0h | 95m | 10m | 10.1x | 96.0x |
| 2 | Ran an entailment-floor re-validation/salvage pass that rescued 27 of 33 partial content packages, backfilled activities across 4 files, ran a lessons generator, wired up auto-promotion, and did documentation cleanup | 20.0h | 130m | 12m | 9.2x | 100.0x |
| 3 | Beta-tier package expansion (44 to 106): migrated the content-generation pipeline to a new candidate model (question sets, exam tips, plus a new post-extraction lessons generator), root-caused a provider-contention issue, provider dedup (merged duplicate provider identifiers, consolidated overlapping test-prep categories), then ran a 29-package synthesis batch through the auto-promotion pipeline | 28.0h | 185m | 18m | 9.1x | 93.3x |
| 4 | Per-domain model routing for the retrieval-answering service via a config registry: math-domain content (detected via a content-type signal) now routes to a reasoning-tier candidate model; added an answering-service role, a content-type override, a model-resolution helper, and extra max-tokens headroom for reasoning responses; wired the routing through the relevant services and smoke-tested math-domain answering; tests green | 3.5h | 35m | 3m | 6.0x | 70.0x |
| 5 | Per-category response-budget right-sizing for the retrieval-answering service: cut the reasoning max-tokens budget from 16384 to 2048 and made the token budget configurable per category in the model-routing registry (a config resolver returns model plus budget, with a per-response override, extensible to other test-prep categories via a content-signal map); ran an accuracy/latency A/B between two candidate models on math content (75-76% vs 77% accuracy, about 15x latency difference); 2 services, tests green | 2.5h | 40m | 3m | 3.8x | 50.0x |
| 6 | Math-content model bake-off for the retrieval-answering service: built a four-model x math-course direct A/B harness (80 MCQs); one candidate model came out best on accuracy and latency (92.5% accuracy, fastest at about 0.5s). Switched math routing to a two-model fallback list, added a new inference provider plus fallback-list resolution and extra reasoning headroom to the model-routing registry; 2 services, 54+84 tests green | 2.0h | 35m | 4m | 3.4x | 30.0x |
Aggregate Statistics
| Metric | Value |
|---|---|
| Total tasks | 6 |
| Total human-equivalent hours | 72.0 |
| Total Claude minutes | 520 |
| Total supervisory minutes | 50 |
| Total tokens | 2,690,000 |
| Weighted average leverage factor | 8.3x |
| Weighted average supervisory leverage factor | 86.4x |
| Human-equivalent weeks | 1.8 |
Analysis
The day's leverage distribution matters more than the headline figure. The 10.1x ceiling came from Ran a per-subject synthesis-quality benchmark: three candidate models x 30 subjects (13 academic subjects, 17 world-language courses), using a per-phase...; the 3.4x floor was Math-content model bake-off for the retrieval-answering service: built a four-model x math-course direct A/B harness (80 MCQs); one candidate model came out.... 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 broad benchmark that produced a single comprehensive report sat at the top; the fine-grained model bake-offs at the bottom were measurement-heavy and iterative, with repeated A/B accuracy and latency runs that resist compression.
The supervisory leverage figure (86.4x 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 6 tasks, the day produced roughly 1.8 weeks of senior-engineer-equivalent throughput in 8.7 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.