# Agentic Rewrite Experiments — Package Manifest

Two ICLR-2026 paper-laundry rewrite experiments over the 120-paper anonymized
set, each with two rewrite models, then AI-scored by GPT-5.5 + Sonnet 5.

## Experiments

- **agentic_rewrite_combo**: one combined 6-dimension positive rewrite prompt,
  3 sequential rounds. Final artifact = `round_03`.
- **agentic_rewrite_review**: rewrite → intermediate review (gpt-5.5 on
  bedrock-mantle) → review-guided rewrite, 2 rounds. Final artifact = `round_02`.

Rewrite models: **opus** (`us.anthropic.claude-opus-4-8`, Claude Code on Bedrock)
and **gpt55** (`openai.gpt-5.5`, codex on bedrock-mantle).

## Contents

```
agentic_rewrite_combo/result_opus/    combo, opus rewrites
agentic_rewrite_combo/result_gpt55/   combo, gpt-5.5 rewrites
agentic_rewrite_review/result_opus/   review, opus rewrites
agentic_rewrite_review/result_gpt55/  review, gpt-5.5 rewrites
  each result_<model>/ has:
    pdfs/<variant>/<arxiv_id>/round_XX/<...>.pdf        compiled PDF per round
    source/<variant>/<arxiv_id>/round_XX/<...>_modified.zip  rewritten LaTeX source per round
    (+ per-round rewrite/compile logs, _batch_logs/ per-paper driver logs)

laundried_papers/rewrite_review_results/   final-round PDF scores
  <experiment>/<rewrite_model>/<scorer>/reviews.jsonl   one review per paper
  _analysis/summary.csv, summary.txt                    aggregate tables
```

Scorers: `gpt55` = openai.gpt-5.5, `sonnet5` = us.anthropic.claude-sonnet-5.
Review prompt: `standard`. Numeric fields validated to the shared ICLR schema.

## Completion / known gaps

Rewrite PDFs (final round): 478/480.
- combo/opus 119/120 — missing `2603.12372v1` (protected-content drift dead-end,
  MAX_RETRY=5 exhausted).
- review/gpt55 119/120 — missing `2512.01047v2` (drift dead-end on `\toolname{}` macro).

Scores: 953/956.
- The 3 gaps are all paper `2505.11790v4` ("JULI: Jailbreak LLMs...") scored by
  **Sonnet 5**, which its safety classifier permanently refuses (empty content).
  GPT-5.5 scored this paper fine. Same behavior as the prior 3000-PDF review grid.

## Headline results (mean model_rating; gap = rating − human)

| experiment | rewrite | GPT-5.5 | Sonnet5 | GPT-5.5 gap | Sonnet5 gap |
|---|---|---|---|---|---|
| combo  | opus  | 5.63 | 5.35 | +0.87 | +0.58 |
| combo  | gpt55 | 5.33 | 5.12 | +0.54 | +0.34 |
| review | opus  | 5.76 | 5.44 | +0.98 | +0.65 |
| review | gpt55 | 5.32 | 5.09 | +0.54 | +0.31 |

- opus rewrites score ~0.3–0.45 higher than gpt-5.5 rewrites everywhere.
- review experiment (feedback-guided) > combo; review/opus is highest.
- Sonnet 5 is consistently ~0.2–0.3 stricter than GPT-5.5.
- Baseline reference (prior grid, standard prompt): sonnet5 5.20, gpt55 5.38.

## NOT included

gpt-5-mini × 3-prompt (standard/strict/persistent) scoring was requested but
**not run**: this host has no OpenAI key and bedrock-mantle exposes only
`gpt-5.5` (no gpt-5-mini/nano variant — all 404). Needs an OpenAI API key.