Agent systems
PIF
65 of 83 tasks passed the verifier first time (no baseline yet)
- AI
Evidence
- Data
- PIF task logs from 6 projects · 83 tasks that reached verification
- Stage
- Prototype
- Stack
- Python, Claude Code subagents, Markdown task briefs
Evaluation
What was measured, against what, and how.
| Result | 65 of 83 tasks |
|---|---|
| 95% interval | 68% to 86% (Wilson score interval on 65 of 83) |
| Baseline | None: the same tasks in a single session without PIF were never run (needs measuring) |
| Sample size | 83 tasks that reached verification, across 6 projects |
| Timeframe | PIF logs as read on 2026-10-17 |
| Method | Strict rule: a task counts as first-pass when its first transition out of verification goes to done; tasks never verified are excluded |
| Source | .pif/log.jsonl in 6 projects (counting rule in the AI build-stats deep dive, section 5a) |
| Confidence | Verified |
Limits
- No baseline, so this says how often the verifier accepted first time, not that PIF beats a single session
- Of the 18 not first-pass, 17 were verifier rejections and 1 an expired login; excluding that infra case gives 66 of 83
- The logs have grown since 2026-10-17; the figure is not refreshed
Outcome: Too early. Goal: tasks that pass an independent verifier the first time. 65 of 83 did across 6 projects, but there's no single-session baseline yet to compare against.
What was built
The deliverables this project covers.
- PIF (Project Infrastructure Framework) framework Prototype
- Claude Project Organizer tool Prototype
- Agent coordination plugin and dashboard plugin Prototype
- Claude skills, styles and plugin setup report Prototype
- Claude app usage notes report Prototype
- AI image provenance research report Prototype
- Media identification searches report Prototype
- Product-intelligence skill tool Prototype
How I used AI
What the tools did, and what was checked.
Built with: Jett with Claude Code sessions; PIF lead sessions (lead / worker / verifier); Jett, by hand with Claude Code.
Note to be written: tool and model, what the assistant produced, what was rejected or rewritten, and which check caught a problem.