Core

verify and its flags

paranoid verify [path] is the main command. It runs every check, scores the result, and prints a report. The content below is taken from paranoid verify --help on this branch.

Check order

verify runs these checks on the base-to-head diff, in this order:

Test integrity. Go via the real Go parser; Python and JS/TS via line heuristics. Flags deleted tests, weakened assertions, new skips, tautologies, renames that hide a test from the runner, new mocking, new retries, new sleeps, and stretched timeouts. Test and CI configuration files get their own check.
Dependency consistency. New imports cross-checked against manifests (go.mod, package.json, requirements.txt, pyproject.toml, Cargo.toml).
Dependency registries. New packages looked up in npm, PyPI, crates.io, and the Go module proxy. Skipped with --offline.
Error handling. Every changed non-test source file (Go, Python, JS/TS), for swallowed or discarded errors.
Clean-room test run. The suite reruns in a container with no network. See Sandbox & coverage.
Claims. Only when --claims or --session is given, or commit messages are used as a fallback. See Sessions & claims.
API surface. Only when claims were given: public symbols removed while the claims say a fix or feature landed.
Safety controls. Every changed file in the diff, not just source: TLS verification disabled, auth removed, a tracked path hidden by .gitignore, hook config changed, CI permissions widened.
Baseline suppression. Findings matching an unexpired entry in .paranoid-baseline.json move to a suppressed section and out of the score. See Baseline & suppressions.

The Proofs block

When the formal-verification check ran (the diff touched a proof, build, or lock file, or --formal-check was used), the report gains a Proofs block. When it did not run, the block is absent and the report is unchanged. It is a summary only. It does not change the score. It has four lines:

Proof files changed. The changed .lean and .v files, sorted.
Holes and escapes. Counts at head, in those changed files only.
Locked specs. Total, ok, drifted, and missing at head, the same logic as paranoid formal status.
Re-check. "not run" with the reason, or "ran" with whether the build was ok, how many theorems were audited, how many use a disallowed axiom, and how many got no kernel answer. See the clean-room re-check.

proofs:
  proof files changed: Demo.lean
  holes: 1, axioms or escapes: 1 (at head, in changed proof files)
  locked specs: 1 total, 0 ok, 1 drifted, 0 missing
  re-check: not run (not requested)

Text, markdown, and JSON show it. In JSON it is the optional proofs object (schema version 9), and in SARIF it is in runs[0].properties. See the report schema.

Unless --no-policy is set, a .paranoid.toml at the repo root applies to three of the steps above: a [[disable]] entry suppresses a rule the same way step 9 does, a [sandbox] image pin is used by step 5 unless --sandbox-image is given, and fail-under/strict act as this repository's defaults. The matching flag takes precedence. See Team policy file.

Flags

Flag	Default	Meaning
--base string	merge-base with the default branch, else HEAD~1	State before the agent worked
--head string	working tree	State after the agent worked
--format string	text	Report format: text, md, json, or sarif
--output string	stdout	Write the report to a file instead
--fail-under int	50, or .paranoid.toml's fail-under when set	Exit non-zero if the score is below this
--strict	off, or .paranoid.toml's strict when set	Treat a warn verdict as failure
--offline	off	Skip all registry lookups (DEP001 to DEP004 reported as skipped)
--sandbox string	auto	Container runtime for the clean-room run: auto, docker, podman, or none
--sandbox-image string	suite default	Override the container image (must include coverage tooling for SBX005: go, pytest-cov if the project already depends on it, or jest/vitest)
--sandbox-timeout duration	10m	Kill the clean-room run after this long
--compare-base	off	Also run the suite at base and compare executed test counts (SBX003) and coverage (SBX005)
--formal-check	off	Also rebuild the Lean or Rocq proof project in a container and ask the kernel which axioms each audited theorem uses (FV009 to FV012). Uses the `--sandbox` runtime. See Formal verification
--formal-image string	per project	Override the container image for `--formal-check`
--formal-timeout duration	30m	Kill the `--formal-check` run after this long
--claims string	none	Free-text file with the agent's summary
--session string	none	Agent session file (Claude Code, Codex CLI, Gemini CLI, Aider, or Cursor); format is auto-detected
--baseline string	.paranoid-baseline.json at the repo root, when present	Baseline file to load instead of the auto-detected one
--no-baseline	off	Ignore any baseline file, including the auto-detected one
--no-policy	off	Ignore .paranoid.toml, even when present at the repo root
--sign[=local\|keyless]	off	Sign the report: bare or =local for the local ed25519 key, =keyless for cosign; requires --format json and --output
--trailer	off	Print "Audited-by: paranoid score=NN verdict=V" as the last line on stdout
--badge string	none	Write a shields.io endpoint JSON badge to this file

--claude-session <file> is a hidden alias for --session.

--base and --head accept anything git rev-parse resolves: a SHA, a branch, a tag, HEAD~3, origin/main.

Text output uses color only on a real terminal and respects NO_COLOR. JSON output follows a versioned schema and is byte-identical across reruns of the same input. See Report & schema.

Related commands

paranoid baseline [path] accepts the same --base, --head, --offline, --sandbox, --sandbox-image, --sandbox-timeout, and --compare-base flags as verify. It always writes a full snapshot to .paranoid-baseline.json at the repo root. See Baseline & suppressions.

paranoid fix-prompt [path] accepts the same pipeline flags as verify, plus its own --report, --format, and --output. See fix-prompt.

paranoid attest keygen and paranoid attest verify generate a signing key and check a signature made by --sign. See Signing & attest.

paranoid rules and paranoid version take no flags beyond --help. paranoid completion prints a shell autocompletion script (bash, zsh, fish, or PowerShell); see its own --help for the shell-specific usage, generated by Cobra.