Elenchus

A standalone system for dialectical knowledge base construction, implementing the Elenchus protocol (Allen 2026) with a DuckDB material base backend.

The respondent develops a bilateral position [C : D] through natural language dialogue with an LLM opponent. Accepted tensions become material implications in a NMMS material base satisfying Containment.

Requirements

• Python 3.11+
• An LLM API key (Anthropic, OpenRouter, or any OpenAI-compatible provider)

Installation

pip install elenchus
export ELENCHUS_API_KEY=sk-ant-...    # or ANTHROPIC_API_KEY

For development:

git clone https://github.com/bradleypallen/elenchus-server.git
cd elenchus-server
pip install -e ".[dev]"

User Guide — full documentation covering concepts, the web interface, CLI, configuration, and a worked example.

Usage

First-time setup (multi-user mode)

Elenchus is a multi-user platform. Before the first sign-in you create an admin actor; everyone else joins by accepting an invite that admin issues.

# 1. Bootstrap the admin.
elenchus admin create --email admin@local --name "Admin"
# (prompts for a password, or reads ELENCHUS_ADMIN_PASSWORD)

# 2. Start the server.
elenchus

# 3. Open the browser and sign in as the admin.
# 4. In the admin dashboard, issue an invite to each user.
#    The dashboard renders the full ?token=... URL — share that link.
# 5. Users click the link, choose a display name and password, and land
#    in the app.

Upgrading from a single-user install? Run elenchus migrate-legacy once to register every existing ./dialectics/*.duckdb under the admin's account and relocate it into the per-actor directory layout. Idempotent.

Web interface

elenchus

Options: --port, --model, --api-key, --base-url, --protocol, --data-dir (see elenchus --help).

Open the URL shown in the terminal (default http://localhost:8741). The web interface provides:

• Creating and resuming dialectics scoped to the current user
• Natural language dialogue with the LLM opponent
• Live bilateral state display [C : D]
• Tension resolution (accept/contest)
• Material implications accumulating in I
• Derivability queries against the material base
• An admin-only dashboard for issuing invites and managing users

Command line

# Interactive session (in-memory)
elenchus-cli --name "My Inquiry"

# Persistent session (saved to DuckDB file)
elenchus-cli --db my_inquiry.duckdb --name "My Inquiry"

# Resume a saved session
elenchus-cli --db my_inquiry.duckdb

API

All /api/sessions/* and /api/admin/* routes require an authenticated session. The flow is: POST to /api/auth/login to get a session cookie, then send subsequent requests with that cookie. A dialectic is addressed by the numeric session_id returned on create.

# 1. Log in (saves the cookie to ./cookies.txt).
curl -c cookies.txt -X POST http://localhost:8741/api/auth/login \
  -H "Content-Type: application/json" \
  -d '{"email": "you@example.com", "password": "..."}'

# 2. Create a dialectic — the response includes "session_id" (e.g. 1).
curl -b cookies.txt -X POST http://localhost:8741/api/sessions \
  -H "Content-Type: application/json" \
  -d '{"name": "prov-o", "topic": "PROV-O Starting Point Terms"}'

# 3. Send a message.
curl -b cookies.txt -X POST http://localhost:8741/api/sessions/1/message \
  -H "Content-Type: application/json" \
  -d '{"message": "Entity is a thing with fixed aspects."}'

# 4. Get state.
curl -b cookies.txt http://localhost:8741/api/sessions/1

# 5. Accept a tension.
curl -b cookies.txt -X POST http://localhost:8741/api/sessions/1/tensions/1 \
  -H "Content-Type: application/json" \
  -d '{"action": "accept"}'

# 6. Check derivability.
curl -b cookies.txt -X POST http://localhost:8741/api/sessions/1/derive \
  -H "Content-Type: application/json" \
  -d '{"gamma": ["entity_fixed_aspects"], "delta": ["individuation"]}'

# 7. List your sessions (admins see every base in the platform).
curl -b cookies.txt http://localhost:8741/api/sessions

The legacy name-keyed routes (/api/dialectics/{name}/...) are retained as a compatibility alias — the study-participant flow still uses them — but new clients should use /api/sessions/{id}.

Admin-only routes: /api/admin/invites, /api/admin/users, /api/admin/users/{id}/{deactivate,reactivate}, /api/admin/backup, /api/admin/audit. Non-admins get a 403.

Architecture

src/elenchus/
├── server.py ──→ opponent.py ──→ LLM API (Anthropic / OpenAI-compatible)
│       ↓
│   dialectical_state.py
│       ↓
│   material_base.py
│       ↓
│   dialectics/*.duckdb
├── static/index.html
├── cli.py
└── pdf_report.py

• server.py: FastAPI app serving the API and static frontend
• opponent.py: LLM oracle — sends state to LLM API (Anthropic or OpenAI-compatible), parses structured responses, applies state transitions
• dialectical_state.py: Definition 4 — S = ⟨[C : D], T, I⟩ backed by DuckDB
• material_base.py: Definition 5 — B = ⟨L_B, |∼_B⟩ with pyNMMS-based derivability
• dialectics/*.duckdb: Persistent state files (one per dialectic)

Persistence

The data directory ($ELENCHUS_DATA, default ./dialectics/) holds:

• platform.duckdb — actors, auth sessions, invites, magic links, base ownership, platform settings. Held open by the server for its lifetime.
• bases/{actor_id}/{base_name}.duckdb — one file per dialectic, owned by actor_id. Each file contains the atomic language L_B, all assessments (the base consequence relation |∼_B), the bilateral position [C : D], open and resolved tensions, and conversation history.
• backups/elenchus-*.tar.gz — output of scripts/backup.py / POST /api/admin/backup (see Operations below).

Schema versions live in each file's meta.schema_version. The migration runner brings every file forward at open time; see src/elenchus/migrations/README.md for how to add a new migration.

Operations

Full deployment guide — systemd unit, Nginx + TLS, backup cron, log rotation, and a pre-pilot checklist — is in docs/OPERATIONS.md. See also the task-oriented guides on the documentation site: Administration (roles, dashboard, invites, users, cost, audit, backups, alerting), Running a Study (conditions, participant flow, questionnaires, blinded judging, export), and Deployment (local / production VM / cloud). Quick reference:

# Liveness + readiness probe (unauthenticated; for uptime monitors).
curl -sf http://localhost:8741/healthz
# {"status":"ok","schema_version":7,"phase_b_enabled":false,
#  "llm_configured":true,"checks":{"platform_db":"ok","data_dir":"ok"}}

# Audit drift between platform.duckdb and the filesystem.
elenchus audit

# Run a one-shot backup (everything → backups/elenchus-YYYYmmdd-HHMMSS.tar.gz).
# Server must be running; this hits POST /api/admin/backup.
python scripts/backup.py --email admin@local --password ...

# Cron entry (3 AM daily, keep latest 14 archives):
# 0 3 * * * ELENCHUS_BACKUP_EMAIL=admin@local ELENCHUS_BACKUP_PASSWORD=... \
#           python /opt/elenchus/scripts/backup.py >> /var/log/elenchus-backup.log 2>&1

# Agent-driven pilot simulation — robustness check before a live study.
# Drives researcher + participants (both conditions) + judges + export
# end-to-end against the real API. 'scripted' is free + deterministic
# (CI gate); 'llm' uses real personas for a pre-pilot dress rehearsal.
elenchus sim                       # 4 participants, 2 judges, scripted
elenchus sim --driver llm          # real LLM personas (needs an API key)

The simulation validates the platform (every role, every transition, blinding mechanics, cost, error handling under the real request stack) — not the science. LLM participants prove the machinery is robust; whether the two conditions produce different-quality outputs is what the human pilot measures.

# Per-persona MP4 walkthroughs — watch the system being used through
# the real UI, one video per perspective (participant in each
# condition, judge, researcher). Drives the served frontend in a
# headless browser; needs the [e2e] extra + a one-time chromium install.
pip install -e ".[e2e]" && python -m playwright install chromium
python scripts/record_demo.py --out demo-videos            # scripted (free)
python scripts/record_demo.py --driver llm                 # real dialogue

# Watch a dialectic unfold between two LLMs from a positum: a respondent
# (a domain expert defending a posited paragraph) vs the real Elenchus
# opponent. Prints the transcript + the resulting bilateral position
# [C : D], accepted material implications, and open tensions.
python scripts/run_dialectic.py                            # default sci-ontology positum
python scripts/run_dialectic.py --positum @paragraph.txt \
    --domain "the Gene Ontology" --turns 6 \
    --respondent-model claude-sonnet-4-6 --out dialectic.json

DuckDB is a single-writer-per-file store, so the production server runs as one process; horizontal scaling means migrating the platform DB to Postgres (the db/registry.py boundary is shaped to absorb that swap without touching route handlers). See design-notes/architecture-vision.md for the larger framing.

Configuration

Environment variables:

• ELENCHUS_API_KEY: Required. Your LLM API key (also accepts ANTHROPIC_API_KEY).
• ELENCHUS_MODEL: LLM model for the oracle (default: claude-opus-4-6)
• ELENCHUS_BASE_URL: API base URL for OpenAI-compatible providers (e.g. https://openrouter.ai/api/v1)
• ELENCHUS_PROTOCOL: API protocol — anthropic or openai (auto-detected from base URL)
• ELENCHUS_DATA: Directory for .duckdb files (default: ./dialectics)
• PORT: Server port (default: 8741)
• SESSION_COOKIE_SECURE: Set to true in production behind HTTPS so the auth cookie carries the Secure flag.
• BCRYPT_ROUNDS: bcrypt cost factor for password hashing (default 12). Tests override this to 4 to keep the suite fast — do not lower in production.
• ELENCHUS_ENABLE_PHASE_B: opt-in flag for the theory-articulation speech acts (ASSERT_IMPLICATION, INTRODUCE_BEARER, RETRACT_IMPLICATION). Off by default: the live message route exposes only {COMMIT, DENY, ACCEPT_TENSION, CONTEST_TENSION, RETRACT, REFINE} to match the planned study's Elenchus-condition vocabulary. Set to 1/true outside study contexts.

Using OpenRouter

export ELENCHUS_API_KEY=sk-or-...
export ELENCHUS_BASE_URL=https://openrouter.ai/api/v1
export ELENCHUS_MODEL=anthropic/claude-opus-4-6
elenchus

Any OpenAI-compatible endpoint works the same way (Together, Groq, etc.).

Acknowledgements

This work was supported by a grant from the Alfred P. Sloan Foundation to the Alliance for Data Science and AI (G-2026-79650).

(Alfred P. Sloan Foundation · Alliance for Data Science and AI)

License

MIT — Copyright © 2026 University of Amsterdam. Author: Bradley P. Allen.