Proserpina — Multi-Agent Critique & Cross-Examination

A pipeline that puts your documents — pre-prints, roadmaps, plans, specs — in the witness box and cross-examines them for intellectual rigor, using a panel of frontier-model critics.

Proserpina runs a configurable ensemble of critic personas over a document via a provider-agnostic interaction-graph engine. A dedicated summarizer LLM pass clusters the panel's critiques into actionable, per-issue findings. LLM backends are pluggable; a deterministic EchoAgent makes the whole pipeline testable with zero LLM dependencies.

⚠️ Privacy: Proserpina sends your document to a frontier-model provider. A critique ships the full document text to the provider(s) backing the critics and summarizer. For confidential or regulated content, route Proserpina to a local model (Ollama, LM Studio) via a custom config section — no data leaves your machine. See Security & Privacy for the full picture.

Why multi-agent critique?

Different frontier models have different blind spots, biases, and strengths. A panel that mixes them catches what a single reviewer — or a homogeneous panel — misses. Proserpina assigns each critic a provider drawn (seeded, reproducibly) from your authed set, so a 5-critic run naturally spans DeepSeek, Z.ai GLM, OpenAI, Moonshot, Alibaba, and Google. The summarizer then tells you where the panel agreed (many critics, one finding) vs. where it contested.

How it works

1. The roster resolves which of your providers are authed (env vars or a credentials file) and assigns each critic persona one, seeded for reproducibility.
2. The interaction graph routes messages between critics. Today: parallel (fan-out) and rounds (adversarial cross-examination).
3. The summarizer — a second LLM call over the transcript — clusters critiques into per-issue Findings: severity, category, location, quote, suggested change, and the supporting critics.
4. The report renders two ways from the same findings: a human-readable markdown digest and machine-readable JSON.

Quick start

cargo install proserpina --features cli,backend-http,json

# Authenticate with a provider (interactive)...
proserpina auth login deepseek

# Pin a model version at login (optional)
proserpina auth login openai --model gpt-5.5

# ...and cross-examine a document.
proserpina critique roadmap.md

# Or use a 5-critic panel fanned across all your authed providers:
proserpina critique roadmap.md --panel panel

A run prints a markdown digest:

# Critique Report

**Subject:** `roadmap.md`
**Findings:** 5 (3 blocker, 1 major, 1 minor, 0 info)

## 1. [blocker] The proposal conflates consensus with eventual consistency.
- **Category:** logical contradiction
- **Quote:** > "consensus algorithm ... using eventual consistency"
- **Suggested change:** Choose a strong-consistency model (Raft/Paxos) or
  rename to a pattern that doesn't claim consensus.
- **Raised by:** Devil's Advocate, Methodologist, Red Team, Domain Expert

_Reproducibility: seed `3`_

Add --json for machine-readable output; --seed N to reproduce a run exactly.

Configuration

Providers, persona panels, retry policy, model overrides, and exclusions live in a single TOML file at ~/.config/proserpina/credentials.toml (overridable via PROSERPINA_CONFIG or --config):

# Auth: a section per provider. Env vars also work (DEEPSEEK_API_KEY etc).
[deepseek]
api_key = "sk-..."

[zai]
api_key = "..."
base_url = "https://api.z.ai/api/coding/paas/v4"   # coding-plan gateway
model = "glm-5.2"

# Model override: pin a specific version for any provider.
[openai]
model = "gpt-5.5"

# Exclude a model from the roster (e.g. billing lapsed, model retired).
exclude = ["qwen3.7-max"]

# Panels: built-in (default/duo/panel) or custom.
[panels.red-team]
personas = [
  { name = "Skeptic", framing = "Doubt everything.", focus = "assumptions" },
  { name = "Nitpicker", framing = "Find the small flaws.", focus = "details" },
]

# Retry policy for all HTTP calls (CLI --max-attempts/--timeout override).
[retry]
max_attempts = 4
timeout_secs = 45

Any custom provider (Ollama, LM Studio, OpenRouter, a proxy) works too — just supply base_url + model + api_key in a section whose name isn't a built-in.

Panels

Define your own under [panels.NAME] (above). --panel <name> selects one.

Language

Critique in any language the models support:

proserpina critique doc.md --panel panel --language Japanese

Omitting --language lets the model infer from the document's language.

Agent integration

Proserpina is designed to be called on the fly by AI agents across your dev environments. The full loop:

capabilities reports dynamic auth state (which providers are authed in this environment) and the exit-code scheme. Exit codes: 0 success, 2 usage, 10 no authed providers, 11 agent failure, 12 summary failed, 13–16 config errors, 70 other.

Reliability

Every HTTP call (each critic and the summarizer) has a per-attempt timeout and retries transient failures (408/429/5xx, network errors) with exponential + jittered backoff. Non-transient 4xx fails fast. Defaults are sensible; override via [retry] or --max-attempts/--timeout.

Features

License

Licensed under Apache-2.0. See LICENSE. Contributors must sign the CLA.