---
license: cc-by-4.0
tags:
  - agents
  - llm
  - agent-skills
  - evaluation
---

# AIP-SkillBench — 16-task combined cohort (Sonnet, AIP v0.3a2)

Raw evaluation-run data.

These are the **first-version (v0.3a2)** runs of cohorts A and B. They surfaced two AIP-compilation
failure modes that motivated a spec bump to v0.3a3; the corrected re-runs are in the companion
[24-task v0.3a3 dataset](https://huggingface.co/datasets/neo4j/aip-skillbench-24task-sonnet-aipv0_3a3).

## What this is

A head-to-head evaluation of two skill formats on the same tasks:

- **`human-curated`** — the task's original human-authored skill (prose).
- **`aip-from-curated`** — that same human skill compiled into AIP (a schema-validated
  execution-graph representation), authored against AIP spec **v0.3a2**.

Two balanced 8-task cohorts (**A, B**), each task × mode run for **5 independent trials**.

| field | value |
|---|---|
| tasks | 16 (cohorts A & B, 8 each) |
| modes | `human-curated`, `aip-from-curated` |
| trials | 5 per task × mode |
| total runs | 16 × 2 × 5 = **160** |
| solver agent | `claude-agent-acp` |
| solver model | `claude-sonnet-4-6` |
| AIP authoring | `claude-opus` against AIP spec **v0.3a2** |
| sandbox | docker |
| benchmark | extends [SkillsBench](https://www.skillsbench.ai/) |

## Why these runs matter — the v0.3a2 → v0.3a3 bump

These runs exposed two ways the v0.3a2 AIP compilation could produce a worse skill than the human
prose it was built from, both stemming from the authored script being **frozen at compile time**:

- **`offer-letter-generator`** (cohort A) — the converted skill's template-filling script encoded a
  conditional-handling bug (it resolved an `{{IF_…}}` block against the wrong data key), so a required
  section was dropped on **every** trial — a deterministic regression below the human skill, which
  reasons over the document instead.
- **`bike-rebalance`** (cohort B) — the conversion produced an over-engineered optimizer (~1,100+
  lines) heavy enough to exceed the per-trial time budget on most runs, where the leaner
  human-written approach finished.

These two patterns — a frozen *incorrect* script, and a frozen *over-engineered/slow* script —
motivated AIP spec changes in **v0.3a3** (a script-vs-prose decision rule, leanness guidance, and a
functional-correctness check during authoring). After those changes the cohorts were re-authored and
re-run; see the v0.3a3 dataset linked above.

## Layout

```
cohort-a/   cohort-b/                  # one folder per cohort, each:
  campaign.json     # the run matrix (tasks, modes, trials, model, agent)
  status.json       # run totals (done / pass / fail / error)
  summary.csv       # one row per trial — the primary table
  summary.jsonl     # same, JSON Lines
  cells/            # per-trial working dirs: rewards, timing, result.json, agent trajectory
  logs/             # per-trial solver logs
```

`summary.csv` columns: `task, model, mode, trial, status, reward, n_tool_calls, wall_clock, error,
jobs_dir, trial_dir, started_at, finished_at, subprocess_rc`.

## Source code, skills, and how to reproduce

The benchmark harness, run configs, and the **AIP-compiled (v0.3a2) skills themselves** live in the
GitHub repo. Check out the matching tag to see the exact skills used for these runs:

- **Repo:** https://github.com/zach-blumenfeld/aip-skillbench
- **Tag:** `sonnet-aipv0.3a2`
- **AIP-compiled skills:** `generated-skills/<task>/aip-from-curated/…`
- **Human skills:** `vendor/skillsbench/tasks/<task>/environment/skills/…`
- **Run configs:** `configs/eval-cohort-{a,b}-sonnet.yaml`

```bash
git clone https://github.com/zach-blumenfeld/aip-skillbench
cd aip-skillbench && git checkout sonnet-aipv0.3a2
```

## Reading the data

```python
import pandas as pd
a = pd.read_csv("hf://datasets/neo4j/aip-skillbench-cohort-ab-sonnet-aipv0_3a2/cohort-a/summary.csv")
```