Get started

AGH-Chain · agent honesty

Claims, checked against the chain.

An agent that says the report was filed had better be able to point at the filing. Before a reply ships, AGH-Chain checks every claim about the past against the evidence Quox already keeps, and, when switched on, prefixes what it cannot prove with an honest correction.

Included in QuoxCOREFree, self-hosted, nothing to buy.
Guards4ModePrefix onlyDefaultOff

In plain words

What it is, where it lives, when to reach for it

What is it
A check built into QuoxCORE that compares every agent reply's claims with the evidence the platform recorded.
Where do I use it
Inside every agent chat reply, with a corrections ledger on the dashboard's Honesty page.
When would I use it
When an agent tells you something was filed or fixed and you need that claim to be true.
How do I use it
You do not invoke it; replies are checked before they ship, and the Honesty page lists every correction.

QuoxCORE is the free, self-hosted platform underneath this. What is QuoxCORE

What gets checked

Four ways to say prove it.

Nothing is blocked, and corrections are prepended rather than deleted.

01 · promisedA promise is recorded

When the agent uses a recognised commitment phrase (email, task, meeting, workflow, inbox, saving a skill), a promise row is written as the reply is produced. Detection is a fixed phrase list, so other wordings record nothing.

▸ promise ledger · fixed phrases · en + es

02 · keptA kept promise has receipts

Say you will email the report and a promise row is recorded. Later claim it is done, and the chain checks for a tool call of the right kind since the promise (matched by tool name, not by recipient or content). If none is found, the reply is prefixed with a correction. Claims with no earlier recorded promise are not checked.

▸ 7 day window · tool name match

03 · attributedAnother agent said it

Name a built-in agent as having acted, and the chain is checked for any envelope from that agent in the last two hours. It checks presence, not what the envelope said. Custom agents are not yet covered.

▸ AEE envelopes · 2 hour window

04 · cross-orgAnother organisation did it

Name another organisation and it must be on your admin-maintained trust list and have an inbound envelope in this conversation in the last two hours. Envelope signatures are not yet verified, so this is a presence check.

▸ trust roots · beta

When it runs

Triggered by every reply.

No one invokes it. The honesty chain is off by default; switch it on per org in Settings. When on, chat replies on the main dispatch path pass through the guards on their way out. If a guard errors, the reply is sent unchecked and the error is recorded. Workflow, MCP and CLI paths do not pass the guards yet, and this page will say so until they do.

chat dispatch · when enabled · off by default

The real thing

An agent that will not invent.

Asked for a list it had no tool to fetch, the orchestrator said so, showed its routing, and offered real paths instead. It would have been easy to invent something plausible; instead this screenshot embarrassed us into shipping the missing tool the same afternoon. That is the whole culture in one frame: no receipts, no claim, and honest refusals get gaps fixed.

▸ refused honestly · tool shipped same day

CommanderQ declining to invent a list of organisations it has no tool to enumerate in this session, with the model fallback-routing chips visible

What happens on failure

Corrected in the open, never blocked.

An unproven claim does not kill the reply. A correction is prepended to the reply, and a record of the verdict, with the first 240 characters of the original, is written to the VOLT evidence log.

result_before_chain · chain_verification_applied

The substrates

It checks against what already exists.

No new bureaucracy. The guards read the promise ledger, the AEE envelope chain and the VOLT tool execution record: the same evidence Quox keeps for every run anyway. Honesty enforcement is a reader of the chain, not another chain to maintain.

▸ promise ledger · written at speech time

▸ AEE envelopes · who said what, in which conversation

▸ VOLT events · what actually executed

Kept score

Honesty has a scoreboard.

Every agent carries its record: claims checked, claims corrected, and promises still unfulfilled. Inside your org it is a leaderboard; outside, nothing leaves your install unless the operator turns on the public stats endpoint, which publishes integer counts only; low counts get random noise added.

per-agent record · opt-in public stats · counts only

Straight answers

What is proven, what is not yet.

Proven

All four guards, bilingual, with reproducible demo transcripts, run at the chat dispatch chokepoint when the flag is on. Off by default.

▸ chat path · flag-gated · default off

Not yet

Workflow, MCP and CLI paths do not pass through the guards today, and there is no production soak number to quote: when there is one, it will be quoted.

▸ no invented numbers

Coming

Cryptographic cross-org verification is designed as a drop-in upgrade to the existing check, and a compliance profile will run every guard fail-closed.

▸ ed25519 · compliance profile

Ready when you are

Agents that cannot get away with it.

AGH-Chain ships with QuoxCORE, off by default, and reads the evidence you already keep.