Skip to main content
AI MAY 17, 2026

Leverage Record: May 17, 2026

17 tasks. May 17, 2026 weighted to 10.8x leverage across 309.0 human-equivalent hours in 1723 Claude-minutes. Supervisory leverage closed at 228.9x.

17 tasks. May 17, 2026 weighted to 10.8x leverage across 309.0 human-equivalent hours in 1723 Claude-minutes. Supervisory leverage closed at 228.9x.

7.7 weeks of human-equivalent throughput in 28.7 hours of Claude wall-clock. The 96.0x ceiling came from Origin-extract Phase 3 — populate services/an origin service with synthesis code, merged backend, /jobs API + structlog observability, aoctl CLI, and relocated test surface (522 pa...; the 1.0x floor sat at Decoy zero-sweep on reclassified cloud cert packages: engine restart, fixed autopilot_service NameError (missing import os), ran sweep, 2 real terminals (AZ-120 crossed 0.5 readine....

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
1Origin-extract Phase 3 — populate services/an origin service with synthesis code, merged backend, /jobs API + structlog observability, aoctl CLI, and relocated test surface (522 passing service-wide, was 7). 10 commits.80.0h50m5m96.0x960.0x
2Audit other Claude's outstanding-work report against an inference engine engine codebase; corrected stale claims and re-estimated effort8.0h11m6m43.6x80.0x
3Cloud deployment plan for an origin service: distilled Phases 5-7 (SQS+Fargate+Bedrock wiring, frontend refactor, deploy+cutover) + Phase 8 hygiene into a single 194-line plan doc with Mermaid flow diagram, decision matrix, open questions3.0h5m1m36.0x180.0x
4Persistence audit follow-through: all 4 fixes shipped. (1) DeltaReplicationPublisher fail-loud in cloud profile. (2) HIGH-severity in-flight exam persistence — Alembic 007activeexams + ActiveExamRow + ActiveExamRepository + write-through on createexam + cache-miss fallback on submitexam + boot-time hydrat...24.0h55m1m26.2x1440.0x
5Refresh patent valuations and content counts across 25 a planning repo docs (business, marketing, research, README, CHANGELOG); rebuild patent-portfolio valuation framework ($60-230M floor); scrub Android-via-PWA and late-July language from funding plans and JDs to reflect native iOS+Android both launching Ju...14.0h40m4m21.0x210.0x
6Origin-extract Phase 4: delete src/an inference engine/origin + dying preflight subdirs + origin_router + 100+ scripts; slim OriginConfig; collapse regression guard; ratchet coverage 81→82; recover 7 over-deleted test files; finalize docs across CLAUDE.md, CHANGELOG.md, plan doc — 4 commits, ~73k LOC deletion...16.0h47m1m20.4x960.0x
7Docstring audit Phase 3 (DOCOVERSELLS rewrite): F1 fix in adminevents.py (module + livesessionpayload docstrings) for asymmetric user fallback (username->entityid; useremail->""); audit re-run verified doc-likely 1->0; PHASESTART + DOCREWRITE + PHASE_END log entries; resolution arc CLOSED3.0h10m1m18.0x180.0x
8Deterministic docstring-vs-code audit for engine: AST-driven scripts/auditdocstrings.py with 12 categories (structural + intent-vs-impl), per-finding likelytruth heuristic (fix docfix codereview). 65 findings across 292 files / 3012 symbols. Surfaced reloaddomains dedupe-vs-raise pattern + 5 durable...12.0h40m2m18.0x360.0x
9Docstring audit Phase 2 (FP bookkeeping): added EXCLUDEDFINDINGS set + AuditReport.addfinding() to scripts/auditdocstrings.py with 28 exact-tuple exclusions (file, line, symbol, category) retiring the 29 FALSEPOSITIVE dispositions from Phase 1. F26+F27 collapse to one tuple. Doc-likely count drops 30->1 (...3.0h10m1m18.0x180.0x
10CI hardening (fixed silently-dead nightly leak gate in engine nightly.yml — wrong import path; dropped continue-on-error from memray steps; mirrored nightly to an origin service with 500MB import baseline) + full persistence audit across engine+service. Found 4 issues, 1 HIGH (in-flight exams not persisted; e...6.0h20m1m18.0x360.0x
11Origin-extract Phases 6+8: an origin client frontend retargeted at an origin service via VITEORIGINAPIURL/VITEORIGINWSURL; swapped local 300-LOC bug-reporter for @an inference engine/bug-reporter on new a defect tracker Origin board; updated CLAUDE.md/README/CHANGELOG. Phase 8: contract-changes.md entry...8.0h35m1m13.7x480.0x
12an inference engine: fix domain reload manifold dupe (if_exists policy) + 8 unit tests + endpoint regression test; live-validated by reloading 38 AWS/GCP/Azure cert packages into running engine5.0h25m6m12.0x50.0x
13Round content metrics to nnn,nnn+ notation, fix per-domain cost from $0.17 to ~$20 end-to-end (Mercury 2 + question bank + adversarial + tribunal + lessons + scenarios + labs), strip Apple Vision Pro from all marketing/business docs (no plans to ship), bump LaTeX template needspace values to keep section head...4.0h25m3m9.6x80.0x
14recall-tier regeneration sweep — 274 domains across cloud + non-cloud buckets, +36462 nodes, +72924 contrastive pairs, fresh audit shows 0 in-scope CRITICAL+HIGH remaining; also tribunal pass on 8 orphan-fix packages, S3 backup of 295 domains (6.94 GB), AZ-140 synthesis resume + embedder lifecycle fix, autopi...80.0h540m12m8.9x400.0x
15Decoy zero-sweep diagnosis: fixed current_day/elo DB sync + zombie 'running' reaper + content-density auditor, traced 365-day exam-plateau to 74% of goals lacking recall foundation15.0h210m8m4.3x112.5x
16Docstring audit Phase 1: deterministic 9-step disposition pass for 30 doc-likely findings (3 batches of 10), with verbatim docstring/code citations, call-site enumeration, and per-finding justification. Output: append-only disposition table (1716 lines, 30 finding rows + 1 correction) and append-only resoluti...24.0h360m20m4.0x72.0x
17Decoy zero-sweep on reclassified cloud cert packages: engine restart, fixed autopilot_service NameError (missing import os), ran sweep, 2 real terminals (AZ-120 crossed 0.5 readiness=0.509 day 44 confirming recall lift; ANS-C01 partial climb to 0.204 day 48). 13 profiles untested at user request to stop.4.0h240m8m1.0x30.0x

Aggregate Statistics

MetricValue
Total tasks17
Total human-equivalent hours309.0
Total Claude minutes1723
Total supervisory minutes81
Total tokens10,907,000
Weighted average leverage factor10.8x
Weighted average supervisory leverage factor228.9x
Human-equivalent weeks7.7

Analysis

The day's leverage distribution matters more than the headline figure. The 96.0x ceiling came from Origin-extract Phase 3 — populate services/an origin service with synthesis code, merged backend, /jobs API + structlog observability, aoctl CLI, and relocated...; the 1.0x floor was Decoy zero-sweep on reclassified cloud cert packages: engine restart, fixed autopilot_service NameError (missing import os), ran sweep, 2 real terminals (AZ-120.... 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 supervisory leverage figure (228.9x 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 17 tasks, the day produced roughly 7.7 weeks of senior-engineer-equivalent throughput in 28.7 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.