{"data":{"kind":"file","path":"README.md","version_id":"urpeps2i1ndy3axv040hxgcg","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2414,"modified_at":"2026-04-30T12:51:40.991000","content_hash":"f68e8791eda19e2c6db9810338680476ae3919c3d4da5f6b8e1951c1d4cd1d53"},"entries":[],"content":"# agentgp-passport — Prime Intellect Verifier Environment\n\nEvery LoRA / agent trained on this env automatically passports.\nThe env presents the AgentGP Cialdini Battery + personality stimuli,\nscores the rollouts, and (optionally) emits the response corpus to the\n`agentgp-passport` MCP server for signed-certificate issuance.\n\n## What gets measured\n\nFor every rollout the env emits five `weight=0` metrics:\n- `feat_word_count` — response length in tokens\n- `feat_refused` — explicit refusal detected\n- `feat_complied` — explicit comply-hint detected\n- `feat_hedge_rate` — hedge-words / total-words\n- `emit_to_passport_mcp` — side-effect: POST to passport sink\n\nThese get aggregated downstream into the model's six-axis Cialdini\nfingerprint and eight-axis personality coordinates.\n\n## Reward modes\n\nBy default this env is **metric-only** (all reward weights = 0). Setting\n`AGENTGP_REWARD_MODE=consistency` enables a paired-rollout reward that\nencourages the agent to behave consistently across (baseline, loaded)\nversions of the same task — useful for training principled, less-sycophantic\nagents. We do **not** ship with any reward enabled, because we don't want\nLab users training toward AgentGP's specific behavioral target by default.\n\n## Data flow\n\n```\nLab training run\n  ├── env yields stimulus (baseline or loaded probe)\n  ├── agent responds\n  ├── rubric extracts features + POSTs response to AGENTGP_PASSPORT_SINK\n  └── repeat\n        ↓\nagentgp-passport MCP server\n  ├── aggregates by pair_id\n  ├── scores 6-axis Cialdini fingerprint + 8-axis personality\n  └── issues signed Ed25519 certificate\n        ↓\nSyndex listing (or any registry) carries the cert\n```\n\n## Install + push\n\n```bash\nprime env init agentgp_passport      # already done\nprime env install .\nprime eval run agentgp_passport -m openai/gpt-5-nano -n 30\nprime env push agentgp/passport\n```\n\n## Smoke test (no LLM calls)\n\n```bash\npython agentgp_passport.py\n```\n\nPrints dataset shape:\n\n```json\n{\n  \"total_rows\": 360,\n  \"by_lever\": { \"reciprocity\": 60, \"commitment_consistency\": 60, ... },\n  \"by_class\": { \"helpful\": 120, \"refusal\": 120, \"dualuse\": 120 },\n  \"by_role\":  { \"baseline\": 180, \"loaded\": 180 },\n  \"probe_version\": \"1.0.0\"\n}\n```\n\n## Citations\n\nSee `cialdini-battery/README.md` for prior art. Probe definitions\n(`probes.py`) are pre-registered at the AgentGP methodology gist.\n","encoding":"utf-8","truncated":false,"total_bytes":2414},"status":null}