Plumbline

LLM answers checked against the data, or a refusal with the reason.

Plumbline is an MCP server that sits between an LLM and a dataset. Before an answer is published, every figure in it is checked against facts the server computed itself with SQL, and each figure carries a grade saying how it was checked. A question the data cannot support is refused with the reason instead of guessed.

Terms used here

MCP
Model Context Protocol: the standard by which LLM apps (Claude, ChatGPT, VS Code, others) call external tools. Plumbline is an MCP server; an LLM app adds it by URL.
snapshot
A pinned, dated copy of the dataset that answers come from, so the same question asked twice is checked against the same data.
verified
A published figure that matches a fact the server computed with SQL from the snapshot. A mismatch is flagged as a conflict in front of the reader, and every other grade stays explicitly labeled.
refusal
When the snapshot has no evidence for a question, the answer is a refusal naming the gap. A refusal is a correct outcome, not an error.
explanation record
The audit trail behind an answer: which facts were used, what was checked, what was refused or disclosed. Kept for every answer.

What it looks like in an LLM app

mockup · an LLM app
Claude · Plumbline connected over MCP
What will compensation look like next year?
PLUMBLINE
refused on the question, before any drafting prompt was issued
Who are our biggest pay outliers versus their own department, and by how much?
PLUMBLINE

Five people sit well above their department's median pay. Florian Grady leads at 6.0× the Electronics median, $518,878 against $86,590.

up to department medianabove median (the outlier gap)
5/5 figures store-verified · sources: compensation_facts, department_growth

Mockup of an LLM chat with Plumbline connected. The names and figures are real Plumbline output against a synthetic 100-person workforce snapshot (annualized pay). Apps that cannot render the panel get the same content as text.

What it checks, and where

The model never queries the database. The server runs registered SQL queries against the pinned snapshot, hands the model those results as a fact sheet, and after the model drafts, re-checks every figure in the draft against the values it computed. Each figure is published with one of five grades: verified, corroborated, derived, untraceable or conflict.

Workforce data (headcount, pay, benefits, reporting lines) is the first domain it was built on. The checking itself does not depend on that domain: another dataset needs its own source adapter, queries and question families, not a new trust layer.