Start here¶
Verel is the framework where nothing is "done" until a grader returns a verdict. An agent (or a human) does some work; a grader checks it and emits a signed
pass/warn/fail; only verified work is allowed to compound. This page gets you from zero to your first verdict in about two minutes.
What problem it solves¶
An LLM agent that grades its own work is both author and reviewer — so "done" means nothing. Verel makes
"done" a verdict from a real check, not a claim: deterministic graders (tests, types, lint, security,
telecom config/KPIs, IaC, …) each return an agentsensory.Report = verdict + grounded issues + a signed
receipt. A report that says pass with no receipt of a grader actually running fails the gate — you
cannot claim success without evidence.
Install + first verdict (no API key needed)¶
The graders are deterministic and offline — grading needs no LLM and no key. Install the dev graders and check a repo:
pip install "verel[dev]" # core + the pytest / ruff / mypy graders
verel doctor # sanity-check your environment (what's installed, what's missing)
verel-ci check --repo . # grade this repo → one signed verdict
verel-ci check runs the graders over your repo and prints a pass / warn / fail with grounded
issues. Each grader's result is bound into a signed receipt (see the trust model) — the
gate rejects a "hollow" pass that has no receipt behind it.
The one idea¶
- Precise graders gate. Tests / types / lint / security / telecom-config / KPI / IaC failures are
deterministic evidence → they can
failthe build. - Advisory graders never fail. Open-ended model judgment (vision, LLM-as-judge) is clamped to
warn— it can inform, never block. - A verdict without a receipt is not a pass. Missing grader, or a grader with no signed receipt → the gate fails closed.
That's the whole contract. Everything else — memory that only compounds verified work, a fleet, the senses (eyes/ears), the telecom and IaC grader tracks, the act-then-verify actuators — is built on it.
Where to go next¶
| You want to… | Go to |
|---|---|
| A hands-on 5-minute walkthrough | 5-minute tutorial |
| Run it against your own repo | Try it yourself |
| Pick the right install for your use case | Install & extras |
| Understand what's cryptographically guaranteed | Trust model |
| Give an agent trusted, compounding memory | Memory in 5 minutes |
| Gate 5G RAN/Core changes | Telecom RAN / 5G Core |
| Gate Terraform / cloud-IAM / K8s | SRE / Platform / Cloud |
Agentic features that write code (e.g.
verel heal --repo ., which fixes a red build and re-grades until green) need an LLM configured — Ollama by default, orVEREL_LLM_PROVIDER=openai. Grading itself never does.