agent-verifier

agent-verifier

When an agent speaks for the firm, "the model was careful" is not a control. The producer model that drafts the message cannot also be the control that approves it. agent-verifier is the small, separate guardian you put in front of send() so every outbound has an independent verdict — and a per-message log a regulator can read. Python 3.10+ standard library only; bring your own LLM.

The library lives at {baseDir}/agent_verifier. Example redlist: {baseDir}/redlist.example.txt.

When to use this

Use it in any agent stack that sends outbound text on the firm's behalf — cold outreach, helpdesk replies, social posts, notifications — especially in regulated contexts (FCA/ICO) where you need an independent attestation per message. It is a library you wire in front of your send(), not a CLI.

How to use it

Install into the project (pip install -e {baseDir}), or vendor the single file {baseDir}/agent_verifier/verifier.py — no required dependencies.

from agent_verifier import Verifier

# Bring your own LLM: any callable (prompt: str) -> str. OpenAI, Anthropic,
# Gemini, Ollama — all work. The semantic axis is optional; omit llm to skip it.
def my_llm(prompt: str) -> str:
    ...

v = Verifier(
    redlist_path="redlist.txt",
    llm=my_llm,
    weekend_block_days=("Saturday", "Sunday"),
    timezone_name="Europe/London",
)

result = v.verify(
    subject="Quick question on AI procurement",
    body="Hi — saw your council just published its AI strategy...",
    recipient="ceo@somecouncil.gov.uk",
    campaign="cold-outreach-2026-q2",
)

if not result.can_send:
    alert(result)        # BLOCK — do not send
elif result.verdict == "WARN":
    log_warn(result); send(...)   # WARN — send + log
else:
    send(...)            # PASS — send silently

result is a structured VerifyResult — serialise to_dict() as the audit artefact.

The four axes (worst-axis-wins)

1. Calendar — day-of-week block (default Sat & Sun). Cost 0.
2. Redlist — token + regex match against a confidential-term list, unless the recipient's address also contains the term. Cost 0.
3. Style (optional) — a pluggable callable (e.g. British-English check).
4. Semantic guardian (optional) — one LLM call that decomposes the body into atomic claims and grades each verified / unverifiable / failed (any failed → BLOCK, any unverifiable → WARN), plus confidential/clarity checks.

Notes for the agent

• Reference paths as {baseDir}/... — never hardcode.
• The whole point is separation: do not reuse the drafting model as the guardian without a distinct, control-biased prompt (override prompt_template=).
• Redlist lines starting with regex: are compiled as Python regex; # are comments. See {baseDir}/redlist.example.txt.
• Every verdict is logged regardless of outcome — that log is the deliverable.
• MIT-0 on ClawHub (canonical GitLab/PyPI release is Apache-2.0). Built by Workloft Labs (https://workloft.ai/labs).