Proofjury

The last command before production.

Coding agents write code no human reviews — and then they ship it. Proofjury is the correctness gate for AI-written code: it intercepts the deploy command itself, decides with deterministic checks, explains exactly why with evidence, and remembers every diagnosed failure so recurrence is caught instantly.

$ proofjury guard deploy -- ./deploy.sh

⛔ DEPLOY BLOCKED — proofjury

  env_vars   missing_env_var   DATABASE_URL (db.py:3), STRIPE_API_KEY
                               (payments.py:14) unset
  tests      tests_not_run     no test run recorded for this worktree
  config     config_mismatch   API_BASE_URL points at localhost (config.py:3),
                               debug mode is enabled (config.py:4)

  Blocking deploy — DATABASE_URL, STRIPE_API_KEY referenced (db.py:3,
  payments.py:14) but unset; the first request will crash. Tests have not
  run against this worktree. Config is not production-ready.

  Fix:
    1. Set the missing env vars: export DATABASE_URL=<value>; export
       STRIPE_API_KEY=<value>
    2. Run: proofjury run tests -- pytest
    3. Point config at production values; disable debug flags

record chk_001 → .proofjury/memory.jsonl · proof: .proofjury/runs/chk_001/ · exit 2

The deploy command was never spawned. After the fixes, the gate passes and the record resolves chk_001; when the same mistake recurs, the block cites the prior diagnosis: ↩ Recalled from chk_001.

Why this exists

• The gate cannot be talked past. Deterministic checks decide pass/fail; the LLM only explains. An agent can't argue its way through an unset env var.
• Blocks the deploy — then tells your agent exactly how to fix it. A Proofjury denial is structured feedback the agent consumes to fix the failure and re-run. The gate makes your agent finish the job.
• Every deploy ships with a proof record. Reproducible evidence — which check failed, file:line, command output — not a regenerated opinion.
• It remembers. Every diagnosed failure is stored in training-ready form; recurrence is caught instantly and cites the prior record.
• Doesn't care which agent wrote it. Claude Code, Codex, Cursor — same gate, same checks, same memory. Correctness, not security: your guardrails stop the dangerous command; Proofjury stops the broken one.

Layout

Path	What it is
cli/	The proofjury CLI (Python 3.11+, Apache-2.0): guard, run, resolve, confirm, memory, init
demo-app/	An intentionally-broken agent-built app — the acceptance-test bed
scripts/demo.sh	End-to-end demo: block → fix → allow → recurrence caught from memory
landing/	proofjury.com landing page (Next.js, static export)

Quick start

# install (Python 3.11+)
pip install "git+https://github.com/kevincui1034/proofjury.git#subdirectory=cli"

cd your-project
proofjury init                       # writes agent hooks; prints AGENTS.md snippet
proofjury guard deploy -- vercel --prod

pipx install "git+https://github.com/kevincui1034/proofjury.git#subdirectory=cli" works too (isolated). Or clone and pip install -e cli for local development.

Runs fully offline with no key. For LLM-written explanations, proofjury login picks a provider — OpenRouter, Anthropic, or OpenAI — and stores the key at ~/.config/proofjury/config.toml (mode 0600, outside the repo). Env vars still work and take precedence: OPENROUTER_API_KEY / ANTHROPIC_API_KEY / OPENAI_API_KEY, plus PROOFJURY_JUDGE_PROVIDER and PROOFJURY_JUDGE_MODEL. Defaults are cheap per provider (OpenRouter openai/gpt-4o-mini, Anthropic claude-haiku-4-5, OpenAI gpt-4o-mini). The LLM only writes the explanation — deterministic checks still decide pass/fail.

Docs

• proofjury-claude-code-handoff-final.md — MVP build brief
• proofjury-full-scope.md — full venture scope
• cli/TAXONOMY.md — the deploy-time correctness failure taxonomy (open spec)