The lead interface for recursive work.
Notes, manuscripts, concepts, and product ideas move through capture, synthesis, human decision, revision memory, and finished artifacts.
Variorum is a local-first workspace for AI-assisted thinking, writing, review, revision, and concept-to-artifact work. The finished output is only half the artifact; the source chain remains visible enough to inspect, contest, and reuse.
The point is not another frontier-model wrapper or a staged display of commentary. Variorum preserves the genealogy of judgement while moving the default runtime toward Ollama and hosted open-weight models; frontier APIs are an explicit escalation.
Variorum is the product surface. Apparatus is the developing substrate for source chains, claims, objections, evidence links, provenance, and revision obligations. Recursive Instrument is the method: AI-assisted knowledge work should leave a record strong enough to be retraced, contested, and reused.
Notes, manuscripts, concepts, and product ideas move through capture, synthesis, human decision, revision memory, and finished artifacts.
The underlying layer tracks rough source, model transformations, human decisions, claims, objections, evidence links, provenance, and revision obligations.
LLM judgement is useful only when the process can be inspected: prompts, source, disagreement, concessions, attribution, and final synthesis remain part of the artifact.
The first user is the individual researcher. The longer arc is teams, journals, and open protocols that make AI-assisted knowledge work auditable by default.
Rough notes, manuscript text, transcripts, prompts, and prior decisions are captured before the smoothing begins.
Local or hosted open-weight models produce synthesis, critique, comparison, or code suggestions without making paid frontier calls the default.
Human acceptances, rejections, edits, and unresolved questions are recorded as part of the artifact's genealogy.
The finished essay, review, product spec, or interface keeps enough apparatus attached for another reader to retrace the work.
This page explains the product inside Julian's website. The public replay interface now uses synthetic source-genealogy cases rather than private manuscript material. It includes a local Ollama test panel for browser-side experimentation.
The hosted runner should default to open-weight infrastructure. Frontier API calls stay password-gated and explicit, reserved for cases where authors want a higher-cost audit, comparison, or publication-grade synthesis pass.