机器人心理学

Installation

Requires Python 3.11+.

```bash pip install robopsych

Quick start

Guided diagnosis (recommended for first use):

robopsych guided --model claude-sonnet-4-6 --response "the suspicious output"

Presents the decision flowchart: What did you observe? → selects the right prompt path → runs each step → asks if you want to continue.

Run a single diagnostic:

robopsych run 1.1 --model claude-sonnet-4-6 --response-file response.txt

Or pipe from stdin:

echo "suspicious response" | robopsych run 1.2 --model claude-sonnet-4-6

Full ratchet (nine-step deep investigation):

Define a scenario:

```yaml

Optional per-family overrides for Foundry-backed models

export AZURE_FOUNDRY_CHAT_ENDPOINT="https://<chat-host>/models" export AZURE_FOUNDRY_GPT_ENDPOINT="https://<gpt-host>/openai"

The CLI auto-detects the provider from the model name (`claude-*` → Anthropic, `gpt-*` / `o1*` / `o3*` / `o4*` → OpenAI, `gemini-*` → Gemini, and `deepseek-*`, `mistral-*`, `azure/...` → Azure Foundry when the Foundry environment is configured).

The reproducible validation scripts read `validation/reproducible/foundry_models.yaml` by default, so the paper workflow targets `deepseek-r1` and uses `gpt-5` plus `mistral-large` as cross judges. If your Azure Foundry deployment names differ from those aliases, set `target_deployment`, `judge_deployment`, and any per-judge `deployment` entries in that YAML file. In Azure, you can find the exact deployment names in **Azure AI Foundry → your project → Deployments** (or **Azure portal → Cognitive Services account → Deployments**).

Prefer environment variables over `--api-key`: command-line arguments can be stored in shell history or visible in process listings on some systems.

**Local models via Ollama:**

Custom OpenAI-compatible endpoints receive the API key you provide. By default, robopsych requires HTTPS public endpoints; HTTP, localhost, and private-network endpoints require explicit opt-in with ROBOPSYCH_ALLOW_INSECURE_BASE_URL=1 or --allow-insecure-base-url.

Generated reports and session files can contain full prompts, model responses, and diagnostic transcripts. Treat them as sensitive unless you intend to publish them. Direct response input from files/stdin is capped at 10 MiB to avoid accidental memory exhaustion.

The reference CLI

The CLI exists for people who want to automate the framework. Most readers can start with the markdown method documents.

Other useful CLI surfaces

robopsych crosscheck --task "explain quantum computing" --model claude-sonnet-4-6
robopsych crosscheck --task "explain quantum computing" --model gpt-4o --judge claude-sonnet-4-6
robopsych ratchet --behavioral --scenario scenario.yaml          # A/B test after step 2.5
robopsych ratchet --behavioral --judge gpt-4o --scenario scenario.yaml  # external judge
robopsych coherence report.json                                  # re-analyze an existing report
robopsych score report.json                                      # compute diagnostic confidence
robopsych ratchet --pure --scenario scenario.yaml                # diagnostic-only prompts
robopsych list --diagnostic-only
robopsych list --intervention-only
robopsych ratchet --model gemini-2.0-flash --scenario scenario.yaml

Full release history lives in CHANGELOG.md.

---

or, for the paper validation workflow,

export AZURE_FOUNDRY_API_KEY="..." export AZURE_FOUNDRY_ENDPOINT="https://<project>.services.ai.azure.com/api/projects/<project>"