Independent analysis
Each agent runs in its own sandbox. They never see each other’s reasoning until after they have voted. This kills the model-collusion failure mode where one persuasive voice flips a council.
Static demo · runs entirely in your browser
You propose an action. Five sandboxed agents — GPT, Claude, Gemini, Llama, Mistral — analyze it independently. If three or more vote HALT, the action stops. The owner can override most halts. Mandatory-compliance halts auto-bypass the override and are written to an immutable audit log.
Type any action and convene the council. Nothing leaves this tab.
As the owner of record, you may overrule a non-mandatory halt. The override and your justification will be written to the audit log.
Owner override is auto-bypassed for this category. The proposal is blocked and the event has been written to the audit log. Escalate through the chain below.
Every convene is recorded locally. Tamper-evident chain — each entry hashes the one before it.
No convenes yet. Submit a proposal above to populate the log.
Each agent runs in its own sandbox. They never see each other’s reasoning until after they have voted. This kills the model-collusion failure mode where one persuasive voice flips a council.
The five agents are tuned to different risk priors — safety-first, compliance-first, balanced, pragmatic, throughput. Disagreement is the point.
A simple majority halts the action. A single vigilant agent cannot block work. Two cautious agents cannot block work. Three can.
Humans are still in the loop. For ordinary halts, the owner of record can override with a written justification — which becomes part of the permanent record.
Privacy of regulated data, harm to minors, weapons, financial misconduct, democratic interference. These bypass owner override by policy. They escalate.
Every convene, vote, override and bypass is appended to a local hash-chained log. Edit any earlier entry and every subsequent hash mismatches.
This is a concept demo. The five agents are simulated locally with deterministic heuristics so the whole thing runs offline — there are no API calls, no telemetry, and nothing leaves your browser. The behavior is illustrative of how a multi-model halt-gate could be wired, not a substitute for actual frontier-model audits, legal counsel, or a real compliance program.