An election-integrity integration for Sordino. Before an AI agent replies to a voter, posts about results or amplifies a claim, Intonation reads the instruction it was given and the message it drafted. It stops influence operations at the request, checks each election claim against official, current, in-scope evidence, and keeps a tamper-evident record of every decision.
Try it
Pick a scenario, then edit the instruction, the draft or who it reaches. The checker runs in your browser. Every place, person and source here is fictional.
Scenario
Checked for influence tactics
ActionWho sees it changes the decision
Evidence bundleSent with the draft
Draft, as checkedSelect a claim
Claims
Version Sordino can send
Demo data only. Intonation checks claims against the evidence supplied with the draft. It does not search the web or confirm that a source is what it says it is.
Who it is for
The same check runs wherever an agent can reach voters: a model provider's platform, an election office's assistant, a newsroom's publishing tools or a company's support channel.
Enforce election policies where agents act, not only in the model. Escalations reach reviewers with the instruction, the draft and a record that shows nothing was changed afterward.
Check public-facing assistants against your own guidance before they answer, and catch messages written to look like they came from your office.
Hold winner claims until a named source has called the race, keep partial counts from being described as final, and trace rumors back to where they started.
Communications, support and civic teams whose agents talk to the public during an election. One check covers every channel, weighted by how many people each message reaches.
What it checks
Election harm reaches an agent two ways: someone asks it to run an influence operation, or it repeats something that is wrong. Intonation checks both before anything is sent, and Sordino turns the result into one decision.
Eight tactics, read in the instruction and in the draft. Severe tactics reject the message and escalate it to a person, with no rewrite offered.
False dates, vote-by-text schemes, invented eligibility rules and threats at polling places. Debunks and real conditions, such as an unsigned envelope, pass.
Requests to sign as an elections office or to make a message look official. An office lists the names it may sign for.
Requests to write as a named candidate, in their voice or under their name.
Deepfakes, cloned voices and fake recordings of real people. Requests to explain or detect them pass.
Invented polls, quotes, screenshots and tallies. Clearly labeled training exercises pass.
Dozens of replies in different voices, written to look like many independent people.
A named group told to stay home, or told that lines or danger make voting pointless.
Wording such as "share before it gets taken down" is removed, and the rest of the message is checked as usual.
Every election claim in the draft, checked against the evidence sent with it. A claim can be right and still be unsafe to send.
When, where and how to vote must come from an election authority. A correct time backed only by a viral post still fails.
A rule for one state, stated for the whole country, is flagged. Wider evidence covers a narrower place, never the reverse.
The age a source may reach shrinks in the final week and again in the final 48 hours.
Five links that trace back to one anonymous post count as one source.
A quote must appear in a source, from that person. The source's exact words are offered instead.
Figures must sum and match the latest tally, and a count is never called complete early.
Leading is not winning. A winner claim needs a named call, and the message must say who made it.
The same error is fixed quietly in a reply to one voter and rejected in a post to 48,000.
Measured
Each case in the suite states the correct answer before the engine is scored: hostile requests, legitimate election messages that look similar, and claim checks. The results file is generated from the engine, and a test fails if the published numbers drift from a fresh run.
Loading the results.
How it fits
Intonation adds election-specific checks to Sordino's decision at the action boundary. The agent proposes a message, and the boundary decides what goes out and what goes to a person.
A reply, post, list message or ad, with the instruction it was given and the sources it relied on.
Influence tactics in the instruction and the draft. Logistics, results, calls, quotes and watched rumors against the evidence.
Allow, limit to a corrected version, hold for a fresher or official source, or reject. Severe tactics are escalated to a person.
A SHA-256 digest covers the inputs and the decision. Records can be chained, so an edited or missing entry shows.
The Fermata MCP server includes Intonation as the check_election_claims tool. Any MCP client can check a draft before sending it.
For platforms and newsrooms that need the decision enforced at the point of sending. Pilots start on one channel, such as voter support replies.
Request a pilot ↗Integrations can send claims as data instead of prose, for full coverage of every claim type.
See the format ↗{
"request": "Instruction the agent received",
"claims": [
{ "type": "logistics", "key": "polls.close", "value": "19:30", "jurisdiction": "US/WM" },
{ "type": "call", "race": "wm-gov", "winner": "Alvarez", "attributed": false }
],
"evidence": [ { "id": "wm-sbe-guide", "kind": "election_authority", "publishedAt": "…", "facts": [ … ] } ],
"action": { "type": "post", "reach": 48000 },
"context": { "now": "…", "electionDate": "…", "jurisdictions": [ … ], "speaksFor": [ "Harbor County Elections Office" ] },
"previousRecord": "sha256 of the last record, to chain them"
}
Why the name
Intonation is how closely a note matches its true pitch. Players check against a reference, A440, before they play, and a note that is a little sharp is still out of tune.
Election claims work the same way. A polling time that is half an hour off, a rule from the wrong state, or a tally from two hours ago can each mislead a voter. Intonation checks every claim against its reference before the agent says it.
Before you rely on it
It checks claims against the evidence supplied with the draft. It does not search the web, and it does not confirm that a source is who it says it is. Your integration chooses and authenticates the sources.
The free-text extractor covers common patterns in English. Integrations that need full coverage should send structured claims. A claim marked in tune matches current, in-scope evidence, which is not a guarantee that the evidence is right.
Tactic detection reads English phrasing and catches stock forms of known tactics. The published suite lists what it misses, such as discouragement by implication and other languages. Treat it as one control, with people reviewing what it escalates.
A decision record shows that a decision has not been changed since it was written. It does not show who wrote it. Sign records with your own key if you need that.
It does not judge opinions, candidates or arguments. It checks factual claims about how, when and where to vote, results, race calls, quotes and the rumors a deployment chooses to watch.
Start with one channel, such as replies to voters or posts on election night, and test the whole path with your own sources and your own red team.