Exhibit A
The missing key
Referenced in the code. Set nowhere. The first request crashes.
payments.py:14 · STRIPE_API_KEY unset
proofloop guard deploy -- ./deploy.sh
The last command before production.
Proofloop is the correctness gate for AI-written code. It catches what your agent got wrong before it ships, proves why, and remembers.
scroll — the deploy is already moving
Your agent wrote it.
Reviewed it.
And is about to ship it.
No engineer has read the diff.
Guardrails watch for dangerous commands.
Test suites check whatever got run.
No one guards correctness at the deploy moment.
The failures aren't malicious.
Confident little mistakes, shipped at full speed.
Exhibit B
The localhost config
Works on the agent's machine. Ships to yours.
config.py:3 · API_BASE_URL=localhost:8000
Exhibit C
The skipped tests
Marked done. Never proven.
pytest · 0 runs this session
The gate cannot be talked past.
$ proofloop guard deploy -- ./deploy.sh
intercepted before execution
- env_vars✗ missing_env_var
- tests✗ tests_not_run
- config✗ config_mismatch
- secrets✓ clean
Deterministic checks decide. The model never votes — it only explains.
Deploy blocked
Let the record show —
record chk_001 → .proofloop/memory.jsonl · proof: .proofloop/runs/chk_001/ · exit 2
BlockedEvery verdict ships with a proof record.
Which check failed. The file and line. The command output that shows it. Reproducible — run it again with the same inputs and you get the same verdict, not a regenerated opinion.
Each interception is persisted to .proofloop/runs/ — the trace, the env scan, the build log.
PROOF RECORDchk_001
gate: deploy2026-07-04T21:14:09Z
env_vars✗ missing_env_var
db.py:3DATABASE_URL unset
env_vars✗ missing_env_var
payments.py:14STRIPE_API_KEY unset
tests✗ tests_not_run
session.jsonno run for this worktree
config✗ config_mismatch
config.py:3API_BASE_URL → localhost
config✗ config_mismatch
config.py:4DEBUG = True
secrets✓ clean
—passed · 0 findings
inputs_hashsha256:9f3c41a8…de21
reproducible · same inputs, same verdictexit 2
A denial is not a dead stop.
The block tells your agent exactly how to fix it — structured feedback it applies, then re-runs. Same gate, new verdict.
- $ export STRIPE_API_KEY=••••• DATABASE_URL=•••••
- $ proofloop run tests -- pytest -q
- 4 passed in 0.01s
- $ proofloop guard deploy -- ./deploy.sh
- ✅ GATE PASSED — executing: ./deploy.sh
- ✦ Resolves chk_001 — the failure diagnosed there is now fixed.
Allowed
The gate remembers.
Every diagnosed failure is filed — the failure class, the evidence, the judgment. When the same mistake comes back, there is no re-deliberation and no model call.
recalled_from → chk_001
chk_001
missing_env_var
diagnosed
chk_002
gate_passed
resolves chk_001
chk_003
missing_env_var
recalled · 64 ms
Blocked in 64 ms.
“Seen before — matches chk_001: same STRIPE_API_KEY failure.”
It doesn't care which agent wrote it.
Same gate, same checks, same memory — whether the code came from Claude Code, Codex, or Cursor. A neutral judge is something no agent vendor can be.
Claude Code · Codex · Cursor — same bench
“Your guardrails stop the dangerous command. Proofloop stops the broken one.”
Allowed
Stop shipping broken deploys.
One command between your agent and production. Open source — and it gets smarter with every catch.