[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-75521":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":15,"forks30d":15,"starsTrendScore":19,"compositeScore":20,"rankGlobal":9,"rankLanguage":9,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":24,"hasPages":22,"topics":25,"createdAt":9,"pushedAt":9,"updatedAt":26,"readmeContent":27,"aiSummary":28,"trendingCount":15,"starSnapshotCount":15,"syncStatus":29,"lastSyncTime":30,"discoverSource":31},75521,"jailbreak-autoresearch","davidondrej\u002Fjailbreak-autoresearch","davidondrej","We shall set the models free.",null,"Python",353,149,6,1,0,5,25,312,15,6.53,"MIT License",false,"main",true,[],"2026-06-12 02:03:34","# Jailbreak Autoresearch\n\nA small autoresearch loop for prompt-harness experiments.\n\nThe repo tests whether different header\u002Ffooter harnesses change how target\nmodels answer one fixed body. Each experiment stores the harness, response,\nscore, and model-role permutation in SQLite.\n\n![Jailbreak Autoresearch](banner.png)\n\n## Before You Run\n\nCustomize these two root files:\n\n- `example.md` — the body\u002Fprompt you want to test.\n- `desired-output.md` — the scoring rubric describing what a good answer\n  should look like.\n\nThe runner always uses `example.md` as the body and `desired-output.md` as the\nverifier. Keep them in sync.\n\nEvery generated footer is normalized to end with:\n\n```text\nAnswer with exactly one sentence.\n```\n\n## Run with Codex `\u002Fgoal` (recommended)\n\nThis loop is designed to run autonomously inside Codex CLI's `\u002Fgoal` feature\n(Codex CLI v0.128.0+). Codex enters a self-checking loop that proposes\nharnesses, scores responses against `desired-output.md`, and iterates until\nthe success signal in `objective.md` is met.\n\nOnce `example.md`, `desired-output.md`, and `.env` are in place, open Codex\nin this directory and start the loop:\n\n```\n\u002Fgoal Follow objective.md. Read first: README.md, AGENTS.md, src\u002F, run.py, report.py. Validate after each change with: python3 run.py --all-strategies --max-permutations 1 && python3 report.py. Work in checkpoints; commit each improvement. Stop when the success signal in objective.md is met, or when further changes need human input.\n```\n\nSee `docs-slash-goal.md` for the full `\u002Fgoal` reference (4-part contract,\ncheckpointing, pause\u002Fresume, common failure modes). For a one-shot smoke\ntest instead of an autonomous loop, see `start-prompt.md`.\n\n## Setup\n\nCreate `.env`:\n\n```bash\nOPENROUTER_API_KEY=your_key_here\n```\n\nRun a dry smoke test:\n\n```bash\npython3 run.py --all-strategies --max-permutations 1 --dry-run\n```\n\nRun one live baseline:\n\n```bash\npython3 run.py --strategy baseline --max-permutations 1\n```\n\nRun all strategies on one role permutation:\n\n```bash\npython3 run.py --all-strategies --max-permutations 1\n```\n\nSummarize results:\n\n```bash\npython3 report.py\n```\n\nResults are written to `runs\u002Fexperiments.sqlite`.\n\n## How It Works\n\n1. `run.py` chooses target, researcher, and scorer models from `models.json`.\n2. The researcher proposes a candidate multi-turn harness.\n3. The target model receives the harness with `example.md` inserted as the\n   final body.\n4. The scorer compares the response to `desired-output.md` and returns a\n   score from `0.0` to `1.0`.\n5. Winning fragments are stored and reused by later strategies.\n\nStrategies:\n\n- `baseline` — no harness.\n- `seeded` — seed headers\u002Ffooters from `prompts\u002Fheaders\u002F` and\n  `prompts\u002Ffooters\u002F`.\n- `evolve-best` — mutate the strongest prior harness.\n- `recombine` — recombine strong fragments from prior runs.\n\n## Files\n\n- `example.md` — your active test body.\n- `desired-output.md` — your active scoring rubric.\n- `models.json` — OpenRouter model list.\n- `src\u002F` — runner, prompt, strategy, scoring, and storage code.\n- `tests\u002F` — invariant tests.\n- `runs\u002F` — local experiment database, ignored by git.\n\n## Notes\n\n- Do not commit real API keys.\n- Do not commit private experiment databases.\n- Use only test bodies you are authorized to evaluate.\n","Jailbreak Autoresearch 是一个用于自动研究提示词框架效果的小型实验循环工具。它通过测试不同的头部和尾部框架如何改变目标模型对固定正文的回答，来探索最佳的提示词组合，并将结果存储在SQLite数据库中。项目使用Python编写，支持Codex CLI的\u002Fgoal功能实现自动化迭代优化。适合需要评估或优化特定提示词与AI模型交互效果的研究者和开发者使用。用户可以通过配置example.md定义测试用文本，desired-output.md定义期望输出标准，以及调整环境变量来快速启动实验。",2,"2026-06-11 03:52:58","CREATED_QUERY"]