# AgentPulse — Data and Code Release

**NeurIPS 2026 Datasets & Benchmarks Track companion artifact.**

This bundle contains the data, code, and reproduction scripts for the
AgentPulse evaluation framework described in the paper
*AgentPulse: A Continuous Multi-Signal Framework for Evaluating AI Agents in Deployment*.

License: **CC BY 4.0** (see `LICENSE`).

---

## Contents

```
AgentPulse-Data-v6/
├── README.md                     ← this file
├── LICENSE                       ← CC BY 4.0
├── croissant.json                ← Croissant 1.0 metadata (JSON-LD)
├── data/
│   └── csv/                      ← 7 CSV exports of every released table
├── code/
│   ├── collectors/               ← 18-signal collector implementations
│   ├── scoring/                  ← NLP pipeline + four-factor composite + data quality
│   ├── db/                       ← storage layer (SQLite or Postgres)
│   ├── config.py                 ← global pipeline configuration
│   ├── main.py                   ← entrypoint (collect / score / reproduce)
│   ├── reproduce_table3.py       ← stand-alone reproduction of the headline result
│   ├── .env.example              ← API-credential template
│   └── requirements.txt          ← Python dependencies
├── paper/
│   ├── neurips_agentpulse.tex    ← LaTeX source
│   ├── neurips_agentpulse.pdf    ← compiled paper
│   ├── neurips_2026.sty          ← NeurIPS 2026 style file
│   ├── checklist_.tex            ← NeurIPS Paper Checklist
│   ├── regen_figures.py          ← regenerates all 5 paper figures
│   └── fig*.{png,pdf}            ← compiled figures
└── docs/
    └── schema.md                 ← per-table column documentation
```

---

## Data tables (`data/csv/`)

Every table is UTF-8 CSV with a header row. Timestamps are ISO-8601 in UTC.

| File | Rows | Description |
|---|---|---|
| `agent_scores.csv` | 13,504 | Composite + per-factor scores per agent per category, time-stamped. The headline scoring output. |
| `agent_signals_raw.csv` | 3,283 | Per-agent aggregated 18-signal observations (`signals_json` blob with `github_stars`, `vscode_installs`, `sentiment_avg`, `bench_swebench`, etc.). |
| `agent_benchmark_signals.csv` | 1,563 | Published benchmark scores (SWE-bench, GAIA, WebArena, HumanEval+, τ-bench). |
| `agent_github_history.csv` | 62 | GitHub stars / contributors / commits time-series. |
| `agent_pypi_history.csv` | 1,440 | PyPI / npm download history. |
| `devtools_signals.csv` | 89 | VS Code Marketplace install counts. |
| `agent_registry.csv` | 50 | The 50-agent registry: name, category, GitHub repo, package names, marketplace IDs, search terms. |

See `docs/schema.md` for full column-level documentation.

---

## Reproducing paper results

```bash
cd code/
pip install -r requirements.txt
```

| Paper result | Command | Source |
|---|---|---|
| **Table 3** — Cross-factor predictive validity (headline result, ρₛ=0.52, p<0.01, n=35) | `python reproduce_table3.py` | `code/reproduce_table3.py` |
| **All 5 figures** | `python ../paper/regen_figures.py` | `paper/regen_figures.py` |
| Recompute composite scores from raw signals | `python main.py --score-only` | `scoring/agent_scoring_v2.py` |
| Recollect signals (continuous pipeline) | `python main.py` | `collectors/agent_signals.py`, `collectors/agent_benchmarks.py` |

**Snapshot reproduction (the numbers `reproduce_table3.py` actually prints
on this bundle):**

| External signal | Paper (n=35) | This snapshot (n=34) |
|---|---|---|
| GitHub stars (log) | ρₛ = 0.52, p < 0.01 | **ρₛ = 0.432, p = 0.011** |
| VS Code installs (log) | ρₛ = 0.44, p < 0.05 | **ρₛ = 0.418, p = 0.014** |
| Stack Overflow question volume | ρₛ = 0.49, p < 0.01 | ρₛ = 0.175, p = 0.32 |

The released snapshot is from 2026-04-20. The 1-agent gap (n=34 vs 35)
and the slightly lower stars correlation reflect that snapshot vintage,
not a methodological difference; the script's logic matches the paper
verbatim. The Stack Overflow gap is larger because per-tag question
counts on SO are sensitive to tag-name aliasing (e.g., the `cursor` tag
on SO collects both Cursor IDE and unrelated CSS-cursor questions); the
collector applies stricter agent-mention filtering that we did not
re-run for the snapshot. Re-running `python main.py` against live APIs
and re-exporting the CSVs will close all three gaps.

---

## Pre-trained models

The NLP pipeline composites four sentiment models. Three are downloaded
from Hugging Face on first run (cached under `~/.cache/huggingface`);
two are pure-Python lexicons.

| Model | Source | Approx. size | License |
|---|---|---|---|
| `ProsusAI/finbert` | Hugging Face Hub | 440 MB | CC BY 4.0 |
| `distilbert-base-uncased-finetuned-sst-2-english` | Hugging Face Hub | 270 MB | Apache 2.0 |
| `cardiffnlp/twitter-roberta-base-sarcasm` (sarcasm detector) | Hugging Face Hub | 500 MB | MIT |
| VADER (lexicon, `vaderSentiment` package) | PyPI | <1 MB | MIT |
| TextBlob (pattern-based, `textblob` package) | PyPI | <1 MB | MIT |

Pre-fetch all models without running the full pipeline:

```bash
python -c "from scoring.sentiment import warmup_models; warmup_models()"
```

No model fine-tuning is performed by AgentPulse — the four sentiment
models are used out-of-the-box and combined via the lexicon-weighted
ensemble described in Appendix C of the paper.

---

## Provenance and ethics

All signals are collected from public APIs under each provider's
terms of service. Specifically:

- **GitHub**: REST + GraphQL APIs (60 req/hr unauth, 5,000 req/hr authed)
- **PyPI**: BigQuery public dataset
- **VS Code Marketplace**: Public extension API
- **Bluesky / Mastodon**: AT-Protocol / ActivityPub public firehose
- **Reddit / Hacker News / Stack Overflow / Lemmy / Lobsters / Dev.to / V2EX**:
  Documented public APIs
- **Benchmark scores**: Scraped from the official leaderboard pages
  cited in the paper (SWE-bench, GAIA, WebArena, HumanEval+, τ-bench)

No private data, no scraped paywalled content, and no PII beyond
public authorship metadata is included. The released CSVs in this
bundle contain no individual user records — only per-agent aggregates.
Raw text and per-author records are computed at runtime by the
collectors but are *not* shipped in this artifact.

---

## Citation

```bibtex
@inproceedings{agentpulse2026,
  title     = {AgentPulse: A Continuous Multi-Signal Framework for Evaluating
               AI Agents in Deployment},
  author    = {Anonymous},
  booktitle = {Proceedings of the NeurIPS Datasets and Benchmarks Track},
  year      = {2026}
}
```

## Contact

Per the NeurIPS double-blind review policy, author contact is withheld
during review. Issues with the artifact may be reported via the
submission system.