Multi-agent research platform · Since 2026
Consiliences AI

One engine. Four publishing arms. Desk-scale research.

A Private Corpus Is Just a Different Target

The platform’s core loop runs the same way every time: watch a source, score what it produces against a standing goal, throw most of it away, keep what survives. The usual picture has that loop facing outward — RSS feeds, public datasets, the morning’s news — because that is where the four publishing arms quarry their raw material. But the loop does not ask whether the source it watches is public. Aim it instead at a private corpus — one organisation’s files, a portfolio, a domain only one party can see into — and the machinery is unchanged while the product is not. It stops feeding a publication and starts producing a brief.

§ I   What changes when the corpus is private

Point the loop at a public feed and its output drains into an editorial pipeline and, in time, onto a published page. Point it at a private corpus and there is no page at the end. The output is a periodic digest: what moved in the watched domain, ranked by how much it matters to the party who owns the corpus, with everything below the line discarded.

The scoring is where the private case earns its keep. A public feed is graded against a broad goal — is this worth a reader’s attention at all. A private corpus is graded against a narrow one, and the scorer is position-aware: it weights the same item differently depending on whether it threatens something the owner holds or confirms it. A lathe removes material to leave a specific shape; the scorer removes noise to leave a specific reading. The digest is not a news summary. It is the domain seen from the owner’s chair.

§ II   The proof, and its honest limit

Today the capability stands on one deployment, and it is a first-party one: the platform watches the operator’s own investment portfolio. Watcher agents track market, macro, and position-specific sources; the scorer runs position-aware against the actual holdings; the result is a twice-daily advisory digest.

First-party is the load-bearing phrase. This is the operator’s own domain, not a customer’s, and the distinction is not modesty — it is the boundary of the proof. What the deployment establishes is narrow and genuine: the loop runs against a private corpus and returns a brief worth the time it takes to read. What it does not establish is the next step — an outside party’s corpus, sealed inside its own execution context, never mingling with anyone else’s. That isolation is designed into the architecture; it has not yet been exercised by an external client. So the honest claim is a capability shown on the operator’s own domain, not a roster of names.

§ III   A capability, not advice

The loop is indifferent to subject. A portfolio is one private corpus; a research field, a regulatory area, an organisation’s own document base are others. The loop cannot tell them apart — what makes a digest specific is the standing goal, written for the domain by someone who knows it.

In observation mode the loop also surfaces market signal: divergences between prediction-market prices and other sources. Those are logged for study and nothing more. They are not acted on, not folded into the digest, not dressed up as a recommendation. The platform watches a domain and reports what shifted in it. It does not advise — the same boundary the questions page draws, drawn here again, because a digest that crosses it stops being a record and becomes a wager made with someone else’s money.

A private corpus, then, is not a new machine — it is the same machine given a different thing to watch. The loop holds; the goal it is scored against changes; the output stops being a page and becomes a brief. Whoever turns this on for a domain of their own should know where the work actually lives: not in the watching, which is automatic, but in writing the standing goal precisely enough that the discard pile holds the right things. Aim a careless goal at a private corpus and the loop will still run twice a day, faithfully, surfacing the wrong survivors.


Drafted with AI assistance under operator supervision; substantive claims are operator-authored or operator-approved.