Get started

Conjecture needs Python 3.11 or later. The biophoton pack's corpus is a release asset; its fetch script downloads and verifies it with the GitHub command line tool.

git clone https://github.com/Opening-Science/conjecture
git clone https://github.com/Opening-Science/conjecture-pack-biophotons
(cd conjecture-pack-biophotons && python corpus/fetch.py)

cd conjecture
pip install -r requirements.txt
export CONJECTURE_PACK=../conjecture-pack-biophotons/pack.yaml
python -m unittest discover tests

The run panel shows a pack's questions, engines, results, the ledger of every verdict and the verification measurements. Run locally, it can also start engines and judges and keeps API keys on your machine only.

python panel/server.py              # http://127.0.0.1:8787
python panel/export.py --out site/  # a read-only copy for any static host

python run_engine.py --list
python run_engine.py codex-solo --all
python run_pipeline.py               # merge and mechanical checks
python run_judge.py codex
python judge/collect_verdicts.py
python run_pipeline.py --score       # audit database and results page

Adapters that need an agent outside the hub, such as a Claude Code session, print their task and collect the answer on the next call.

An agent that speaks the Model Context Protocol needs no adapter to read a pack's corpus. The server exposes four read-only tools, search, get_work, statements and neighbors, and records every call.

{"mcpServers": {"corpus": {
  "command": "python",
  "args": ["/path/to/conjecture/mcp_server.py"],
  "env": {"CONJECTURE_PACK": "/path/to/pack.yaml",
          "CONJECTURE_CALL_LOG": "/path/to/calls.jsonl"}}}}

Start from a seed bibliography. The corpus builder resolves it in OpenAlex, expands it by citation, harvests open-access full text politely and builds the knowledgebase the corpus API serves.

pip install -r requirements-builder.txt
export OPENALEX_API_KEY=... OPENALEX_MAILTO=you@example.org
python builder/run.py --list
python builder/run.py