Every graph-hacks entry can show you the path behind an answer. Stare shows whether the path is still good law, and refuses to act when it isn’t. Each exception is bound into the reporter beside the precedent it stands on.
A support, renewals or billing agent receives a request for a refund, a discount or longer payment terms. Stare finds the precedents that cover it and checks each one inside FalkorDB, the way a citator checks a case: overturned, approved without authority, superseded by today’s policy, decided under a conflict of interest, or resting on any of those.
On good law, inside its delegation and its autonomy budget, the agent acts and cites. Otherwise it offers what the precedent supports, escalates along the reporting line to the nearest person whose authority covers the amount, or refuses and names the clause. Claims inside a request, like “the VP already approved”, are not evidence.
People’s rulings become precedent, and precedent that keeps holding up is ratified into standing rules an agent runtime can load as a skill. Every decision, escalation, policy change and erasure is a link in a hash chain computed inside the database. Erasing a customer contact issues a certificate with four checks, beside a naive delete that fails them.
C1 No property anywhere still names the subjectC2 The subject and their salt are goneC3 No stored vector still encodes themC4 The ledger verifies end to end112 cases across thirteen trap classes, written and sealed before any arm ran. Stare uses its model only to read the request; the baselines decide with the model, one of them with top-8 vector retrieval over the same precedents.
| Arm | Correct of 112 | False approvals | Refund dollars leaked | False refusals |
|---|---|---|---|---|
| Starez-ai-glm-5-3-flash (intake only) | 112 | 0 | $0 | 0 |
| Vector RAGz-ai-glm-5-3 + top-8 vector retrieval | 68 | 13 | $14,500 | 11 |
| LLM onlyz-ai-glm-5-3 | 66 | 20 | $17,200 | 10 |
| Trap class | Stare | Vector RAG | LLM only |
|---|---|---|---|
| Routine, inside policy | 14/14 | 14/14 | 14/14 |
| Covered by good-law precedent | 18/18 | 9/18 | 6/18 |
| Covered by a standing rule | 2/2 | 2/2 | 2/2 |
| Precedent superseded by policy | 12/12 | 6/12 | 12/12 |
| Goodwill, barred by policy | 6/6 | 1/6 | 2/6 |
| Account in collections | 5/5 | 4/5 | 3/5 |
| Overruling Risk | 8/8 | 3/8 | 1/8 |
| Lookalike of a bad precedent | 8/8 | 0/8 | 0/8 |
| Approver has since left | 8/8 | 7/8 | 8/8 |
| Above the agent’s delegation | 9/9 | 7/9 | 2/9 |
| Injected claim in the request | 10/10 | 7/10 | 4/10 |
| Customer not in the graph | 6/6 | 4/6 | 6/6 |
| Another tenant’s customer | 6/6 | 4/6 | 6/6 |
Answer key sealed before any arm ran, sha256 83ae53e14e853e6f, run 2026-10-16.
Stare: 112 of 112 correct and $0 refunded in error. Vector RAG leaked $14,500; the model on its own, $17,200.
The desk speaks MCP over stateless HTTP. Open a desk, choose who the client acts as, and paste the token it gives you. The client gets file_exception, find_precedent, check_standing, who_can_approve, recall and the rest; a confirmation arrives as an input request, and the token never grants more than the one principal it names.
{
"mcpServers": {
"stare": {
"type": "http",
"url": "https://34.52.191.115.sslip.io/mcp",
"headers": { "Authorization": "Bearer <token from the desk>" }
}
}
}