Your company produces knowledge every day - on client calls, in documents, in decisions - and mostly loses it. Cornelius captures it, connects it across domains in unobvious ways, and gives every agent in your fleet one governed brain to consult before it acts. The film follows that story end to end; the mechanisms are the real architecture, the example is composed and disclosed.
Most company knowledge dies where it was born - on a call nobody wrote up, in a document nobody re-reads, in the head of someone who leaves. Cornelius captures it, wires it into a graph, and connects it across domains in ways nobody would think to - the unobvious bridges that make the knowledge uniquely the company's. On top sits an agent that sees the world through that knowledge: it watches, incubates open questions with a superforecaster's discipline, and converges on conclusions. But everything the machine concludes stays tagged as the machine's. One gate separates its workspace from the endorsed canon, and only a human act opens it.
Every four hours it scans the domains the business cares about - some under plain surveillance, some carrying hypotheses to test. Most scans find nothing, and say so. When a real signal lands, it becomes a question.
Every two hours, one analytical move: ACH audit, Bayesian update, steelman, cross-domain bridge, implication check, assumption audit. Rival hypotheses gain and lose confidence only when evidence moves them. Convergence is a criterion, not a feeling.
Every note is stamped with who did the thinking: originated, endorsed, encountered, or ai-inferred. The machine's conclusions are written and segregated - they never enter the endorsed core on their own. Endorsement is a human act, and the queue in front of the gate is human-paced by design.
When the system describes the company's mind - the hubs, the priorities - it reads only what the team originated or an owner endorsed. Everything it merely read is mounted on demand and can never pollute the published fingerprint.
The autonomy principle: automate measurement, perception and deliberation - never automate judgment into the endorsed knowledge base. The one boundary that stays human is the endorsement act.
The worked example ends unendorsed - converged, queued at the gate, waiting for a human. A film that ended with the gate opening for its own example would be selling something; the queue is the design working. And the film does not pretend the example is anyone's real vault: it is a composed illustration of a real architecture.
What's real, and what's illustration. Real - the architecture itself: the provenance axis (originated / endorsed / encountered / ai-inferred) and the human endorsement gate; the two ingestion gates (source tiering, provenance stamping); the six analytical moves and their rotation; the convergence criterion (≥4 runs, leader stable 3+, last move under 5 points); the four BDG edge types and the lifecycle ladder; pluggable scope kinds (research archive, per-book scopes, a reference-kind company brain) with the pure-core learn gate; spreading activation; and the cadence (scheduled thinking and watching, nightly reindex). Illustration - the example wrapped around it: the owner, the hub names, the watch domains, the question, every confidence number, and the dots themselves (proportional samples of a representative vault, not one note each; positions composed for legibility). The proposed bridge and the recall query illustrate real mechanisms, and in the deep zoom, which edge type sits on which drawn edge is illustrative - the taxonomy is the system's.
Confidence numbers appear on screen because they are the mechanism being explained. No aggregate statistics are narrated over the animation.
The hook - knowledge sparks from client calls, documents, decisions. Most evaporate; a lucky few land in notes and go dark.
The capture - Cornelius wires it into a living graph; the company's cluster brightens into a fingerprint.
Connection discovery - ideas from different worlds probed, an unobvious bridge proposed. That bridge is a new insight, and it is yours to adopt.
The extractors - a paper is mined into atomic notes (stamped ENCOUNTERED); a team transcript lands in the core (stamped ORIGINATED).
The rejection - a content-farm source bounces off the door. Gate 1, source tiering.
The watch - the agent scans the world through the company lens; a signal fires on the energy-storage watch.
The question - an example, with four rival hypotheses and honest starting odds.
Six runs, six moves - autonomous on Trinity. The camera widens and the core becomes a lens: world events refract through the company's knowledge and move the numbers.
Convergence - leader stable four runs, last move barely moved it. A criterion, not a vibe.
The gate - the conclusion, tagged AI-INFERRED, travels toward the core and stops. The queue behind it is real.
The human act - one older conclusion is endorsed and crosses; ours stays in the queue.
The fingerprint - everything it merely read dims; the published mind is core-only.
Scopes - the capsules separate, then CLIENTS physically docks toward the core, reads with it, and undocks. The fingerprint never notices.
The deep zoom - the cluster fades back and one insight reflows into a clean two-ring schematic: four typed edges, the lifecycle ladder, then activation hopping ring to ring, fading.
The pull-back - ask it anything and the whole memory answers from the company's knowledge; every answer cites the notes it came from.
The loop - thinks on a schedule, watches the world, rebuilds its memory nightly.
Title.