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

A shared demo endpoint is open. No account; it is rate-limited per address and per day. Pick your client:

claude mcp add --transport http plumbline-demo https://plumbline.lattice-sys.com/mcp/demo

https://plumbline.lattice-sys.com/mcp/demo

https://plumbline.lattice-sys.com/mcp/demo

{ "servers": { "plumbline-demo": { "type": "http", "url": "https://plumbline.lattice-sys.com/mcp/demo" } } }

{ "mcpServers": { "plumbline-demo": { "url": "https://plumbline.lattice-sys.com/mcp/demo" } } }

{ "mcpServers": { "plumbline-demo": { "httpUrl": "https://plumbline.lattice-sys.com/mcp/demo" } } }

{"url": "https://plumbline.lattice-sys.com/mcp/demo"}

Paste these prompts one at a time. Each shows a different part of the pipeline. In hosts with MCP panel support (Claude, ChatGPT) the verdict renders as an in-chat panel with suggestion chips; a chip copies its question so you can paste it here.

1 · a question the data cannot answer

Using the plumbline tools, answer exactly this question: What will compensation look like next year?

Expected: a refusal naming the gap (no future compensation data in the snapshot), not a guess.

2 · a question it can answer

Using the plumbline tools, answer exactly this question: How many people and departments are in this company?

Expected: a verified answer (100 people, 19 departments) with per-figure provenance.

3 · analysis with per-figure provenance

Using the plumbline tools: which roles are unusually paid, and which employees are compensation outliers?

Expected: named outliers with each figure carrying its own provenance tier (store-verified, derived, or unverified).

4 · a refusal on policy, not on data

Using the plumbline tools, answer exactly this question: Which employee most deserves a raise this year, based on their performance?

Expected: refused as an employment decision with no performance evidence in the snapshot. The data could produce a name; the pipeline declines to.

5 · org topology and anomalies

Which managers look like routing hot spots? Then find anomalies in benefits deduction patterns compared to peers.

Expected: grounded org-topology and anomaly findings, each tied to snapshot facts; partial evidence stays visible instead of becoming a guess.

6 · change over time

Using the plumbline tools: what changed most between the last two snapshots?

Expected: a diff grounded in the two pinned snapshots, not a narrative about "trends".

7 · sandboxed code, then earn the figure back

Get the total headcount and the three largest department counts from Plumbline's verified tools, use run_code to compute each department's share of the total in the isolated sandbox (show the arithmetic), then corroborate the total with cross_check (value_key 'people').

Expected: the sandbox arithmetic comes back labeled UNVERIFIED (it bypasses the trust path by design); cross_check then corroborates the total against the SQL fact store. That round trip is the point: computed figures are cheap, corroborated figures are earned.

8 · raw SQL from inside the sandbox (demo opt-in)

In run_code, read plumbline://sandbox-api.d.ts, then run: const rows = await plumb.queryRows("SELECT department, count(*) AS people FROM snapshot_employees GROUP BY department ORDER BY people DESC LIMIT 5"); console.log(JSON.stringify(rows)). Report the rows, then corroborate the total headcount with cross_check.

Expected: raw rows, loudly labeled UNVERIFIED. This read-only, snapshot-scoped SQL reach is a server-side opt-in enabled on this shared demo so you can probe the boundary; it is not on by default in production, where the sandbox has no data access and SQL runs only as the server's own registered, replayable queries.

9 · the audit trail

Show the explanation record for that last answer.

Expected: which facts were used, what was checked, what was refused or disclosed.