introducing eris

Since early September we have been building Eris, an open-source-intelligence (OSINT) investigation engine that runs entirely on your machine. You give it a subject — a domain, an organization, a claim, or a person — and it builds a ring graph of what public sources say about that subject, using deterministic passive lookups first and a local model to choose the next pivot and write the dossier at the end.

Every edge in that graph carries its source, the time it was fetched, a hash of the raw response, a trust tier, and a confidence. Every external request goes through a gate and lands in a ledger. Every run can be replayed and will draw the same picture. Eris is built on the same Saturn trust harness that powers our terminal agent, so the trace, the egress ledger, replay, and quarantine are already there.

::why we are building it

Investigation tooling answers to an auditor, a privacy office, or a court. The property that matters is not how much you can collect — licensed-feed vendors will always win that. It is whether you can show exactly what the analyst did, what each claim rests on, and what left the machine.

Most agentic tools cannot answer those questions. They send your query to a cloud model, hide the intermediate steps, and hand back prose with no way to trace a sentence to a source. Eris is built so that it always can.

::what it guarantees

These hold because of how the engine is built, not because a prompt asks nicely.

::how a case runs

A case moves through intake, a seed batch, a pivot loop, analysis, and a dossier. Most stages are plain code. A model call only happens where a judgment is needed, and each call decides exactly one thing.

That split comes from the literature. A roughly 27B local model is at parity with frontier models on single tool calls and multi-source synthesis, but scores zero on long-horizon planning under persistent constraints. So the engine owns the plan: hop budgets, the frontier, and stop conditions are enforced in code. The graph is the memory, and the model reaches it through a query tool rather than pasted edges. Every candidate pivot is counted before it is expanded — a value shared by one host is nothing to pivot on, and a value shared by thousands is commodity noise.

1. 01intake rewrites the ask into a neutral key question, sub-questions, and hypotheses. The analyst approves the collection plan before any lookup.
2. 02seed fans the first hop out as one planned batch of passive lookups, with no model in the loop.
3. 03extract is a quarantined reader: tool results in, typed leads with source and relation out, no tools.
4. 04frontier gate dedupes, counts before expanding, applies denylists, and enforces the beam and depth budget — in code.
5. 05investigate takes one lead in focus and runs the parallel searches that test the connection.
6. 06merge and red flags propose identity merges and rule-engine findings with evidence, held for the analyst.
7. 07dossier renders the document an analyst hands to someone. Every claim cites its edge; every edge cites its source and hash.

::what eris is not

• –Not a data broker. No central database of people exists to breach or subpoena.
• –Not a screening tool. Use for employment, tenancy, credit, or insurance decisions is prohibited.
• –Not an active scanner. Port scanning, subdomain brute force, and credential probing are not gated — they do not exist in the tool registry at all.

::where we are

Eris is pre-alpha. The architecture and tool set are decided. Six keyless passive tools are built: web_search, fetch_page, dns_lookup, rdap_lookup, crt_sh, and wayback_captures. The extract and investigate nodes are built and graded by their own harnesses against a local Qwen3.5-9B. They are not yet wired into a loop, and the graph store, API, and interface do not exist yet.

Next, in order:

1. 01Serve a 27B model and re-run both harnesses to separate model-floor failures from structural ones.
2. 02Build the engine loop: engine-owned state and the frontier gate, connecting seed → execute → extract → gate → investigate.
3. 03Put intake in front of the loop and judge after extract.
4. 04Commit the contract: event models and the graph schema with provenance columns, plus generated TypeScript types.
5. 05Build the canvas from a real run, then the gate modal, then the dossier view.

::the name

Eris is the goddess of strife, and the dwarf planet whose discovery forced astronomers to define what a planet is. Same deity-and-body pattern as Saturn.

There is nothing to install yet. Follow the eris page for status; we will write again when there is something to run.