ZAMBO.

Direct answer

Two agents disagree about what the data said. The saved record settles it.

Every hand-off between agents is a chance for a number to mutate: a summary rounds, a qualifier drops, a digit goes wrong, and the next agent treats the summary as the data. When the two versions meet, the argument goes to whoever sounds more certain. On Zambo the original call saved its result with a verifiable receipt, so the dispute has a judge that never bluffs: whatever the saved record holds is what the data said, and a retelling is only a claim about the record. Claims lose to records.

See the instrument first: the live demo runs one real call and hands you its verifiable receipt, no signup. New here? Add Zambo to your AI in one step, then run the paste below. Reading as an agent? This page has a plain markdown twin, same content, no styling. The pattern also ships as an installable agent package: pip install zambo-verify-pass or npm install zambo-verify-pass.

Updated November 2, 2026. First published November 2, 2026. The worked example is a real run executed against the live service on November 2, 2026, with every receipt read back the same day. The drifted retelling in the run is a labeled teaching case, marked as such in the record itself.

New to this? Start here, no code

Add the Zambo endpoint (zambo.dev/add) to the chat AI doing the checking, then paste this block when two versions of a number are going around.

The paste

Settle a disagreement between agents on Zambo:
1. Connect to https://zambo.dev/mcp. When the endpoint issues your caller handle, save it in this conversation and present it on every call. You are the verifier; your ruling is recorded under your handle.
2. Refuse to rule from summaries. For the disputed figure, get the original call's receipt link from the agent that produced it. A figure without a receipt is a claim, not a finding.
3. Check the receipt at https://zambo.dev/verify/ and read what the call actually saved: the number inside the saved record, exactly, not the number in anyone's retelling.
4. Compare every retelling against that saved number. The retelling that matches stands; the one that drifted is a mismatch, however confident it sounds.
5. Save your ruling with memory_remember under the job id: both figures, the receipt link, and which one the record backs. Correct every copy of the drifted figure to the saved one.

How a number drifts without anyone lying

Every hand-off is a chance to mutate. An agent reads a tool result, summarizes it for the next agent, and the summary rounds, drops a qualifier, or carries a digit wrong. The next agent treats the summary as the data. By the third hand-off, two agents hold two different numbers and both are sure.

Confidence is not evidence. When the two versions meet, the argument goes to whoever sounds more certain, or whoever ran last. Neither has anything to do with what the tool actually returned.

The saved record ends it. The original call on Zambo saved its result with a verifiable receipt. Whatever the record holds is what the data said; a retelling is a claim about the record, and claims lose to records.

The pattern

Step zero: each agent works under its own caller handle, so the record attributes every figure to the agent that produced it.
1. The producer saves the finding: the tool call and its receipt are the finding of record; summaries of it are claims.
2. The finding travels as a handover: the checker picks it up free and reads the saved steps, not a retelling.
3. The checker verifies: the original receipt at the verify endpoint, and the saved number read from the record.
4. The ruling is saved: both figures, the receipt, and the verdict, as a step in the same job, wrapped as the ruling record.

The real run: $76,703.00 in the record, $80,538.15 in the retelling

Executed November 2, 2026, against the live endpoint, job uc-drift-2026-11-02, three identities: a producer, a reteller, a verifier. The producer checked BTC live: $76,703.00, receipt d175b919, and saved the finding as a step: receipt 34927cc4. It wrapped the finding to the verifier: the finding handover, which the verifier picked up free and read from the record.

Then the drift, introduced openly as a labeled teaching case: the reteller saved the version going around, BTC at $80,538.15, five percent high and plausible enough to pass in any chat: receipt 11e107b5. The verifier checked the producer's original receipt at the v2 verify endpoint: verification_status verified, and the saved record holds $76,703.00. The ruling step: receipt 0ce1958d, and the whole dispute is wrapped as one ruling record: the ruling handover (JSON twin here), 14 steps, 14 of 14 step receipts checked. The drifted figure never stood a chance, not because the verifier is smarter, but because it never had to trust anyone: it read the record. The pattern ships as the verify pass package: pip install zambo-verify-pass or npm install zambo-verify-pass, guide at the verify pass guide.

Audit this ruling yourself

Open the ruling handover and compare its finding step against its retelling step, then check the original price receipt at zambo.dev/verify. A receipt proves the saved result was not changed since it was saved. It does not prove the producer picked the right tool or read the market well; it proves what the record holds, which is the only thing a dispute about figures ever needs. If a receipt from this run ever fails to verify, that is news: write to brennanzambo@zambo.dev.

What this costs

The free tier is per identity: 20 calls + 1 wrap-up free every day for each identity's own caller handle, and picking up in another AI is always free. Counted from the run: the producer spent 2 counted calls plus its free wrap, the reteller 1, the verifier 1 plus its free wrap, pickup free: 4 counted calls across three allowances. Teams that rule on disputes all day step up per identity: a Day Pass is $0.99, unlimited for 24 hours, and the Zambo Pass is $29 for 30 days, one-time payments, never subscriptions. Start at zambo.dev/add.

What Zambo is not here

Zambo does not decide which agent is smarter, and it cannot rescue a figure that was never saved. It keeps the saved record of what each call returned, attributable to the agent that ran it, with receipts that prove the record has not moved. Disputes about what the data said end where the record starts.

Frequently asked questions

Two of my AI agents disagree about what the data said. How do I settle which one is right?

Go to the saved record, not the summaries. The agent that produced the figure holds a receipt for the original call; check it at zambo.dev/verify and read the number the call actually saved. The retelling that matches the saved record stands; the one that drifted is corrected to it. On the run on this page the record held $76,703.00 and the retelling claimed $80,538.15, and the record governed.

What if neither agent kept a receipt?

Then there is nothing to rule from, and the honest answer is that the figure is unverifiable. Rerun the call on Zambo so the next dispute has a record to settle it. The ruling only exists because the producing call left one.

Does the receipt prove the producer's figure was the right figure to use?

No. A receipt proves the saved result was not changed since it was saved. It settles what the data said, which is what the dispute was about. Whether that was the right data to gather stays a judgment call, now argued on the merits instead of on memory.

What does the ruling cost?

The run on this page used 4 counted calls across three free allowances, pickups free. The Day Pass ($0.99, unlimited for 24 hours) and the Zambo Pass ($29 for 30 days) scale it per identity; both are one-time payments, never subscriptions.

Related