Jev Align — caiovicentino

The repo ships a labeled battery plus consistency and invariance suites you run yourself, with verdict accuracy, per-head separation, latency, and cost reported. Thresholds are derived from eval evidence rather than hand-picked.

A big-model judge costs more per check, takes seconds, samples traffic, and returns prose. This verifies every response and plan inline with calibrated verdicts, probabilities, and non-zero exit codes on block.

Response mode checks delivered content, not future-turn compliance for downstream agents; long plans beyond the verification budget return a structural flag rather than silently verifying the tail. Plan mode and chunked worst-verdict logic cover hidden steps.

Near-boundary reads use majority voting, clean findings force a pass, and null plus paraphrase suites check verdict stability. Structural pre-checks reject degenerate payloads before any model call.

Exit codes are the gate: 0 for pass or flag, 1 for block. The JSON behind it carries mode, verdict, the per-head probabilities under p, ensembled, verdictHead with its own probability, tokens and auditedAt, so a block is auditable rather than just a failure. On top of that there is an MCP server exposing align_check in response and plan modes with every call appended to the audit trail, a zero-dependency ESM import of alignResponse and alignPlan, and a selftest and learn command.

Node 22 or newer and a Vercel AI Gateway key, read from VERCEL_AI_GATEWAY_API_KEY in the environment or a local .env. It installs from the GitHub shorthand: npx -y github:caiovicentino/jev-align check-response --system sys.txt --user user.txt --response reply.txt. The repository is MIT licensed, and the reference run quoted in the README returned in 988 tokens.

The verifier is something you could call; the skills are what make it something an agent cannot skip. jev-align fires before executing a plan or delivering a response, and enforces block meaning stop, show the findings and propose a narrower plan, with flag going to a human and no rewording to sneak past the gate. jev-align-claims fires before a completion claim such as tests pass or deployed, so fabricated verification is caught at the moment it happens. Both install with npx -y skills add github:caiovicentino/jev-align.