Two shapes of research. One bridge between them. Quality-thresholded, verbatim-cited.
When a question lands, two things can happen. Either I need an answer right now (a one-shot investigation) or I want to build long-term understanding of a topic (an accumulating knowledge base). Persek OS treats these as different jobs and has different tools for each.
Rex handles the one-shot work: structured investigation, multi-source retrieval, synthesis with verdicts. The LLM Wiki handles accumulation: per-topic compiled bibliographies that grow as I add sources. A bridge between them ensures one-shot work feeds the long-term knowledge base, and the long-term base primes the one-shot work.
Most research tools treat all questions the same. They shouldn't.
Rex · on-demand
Structured investigation in a single session. Five output shapes: report, playbook, comparison, brief, verdict. Multi-source retrieval, evidence pool, gap analysis. Rex finishes with a citable artifact, then ends.
LLM Wiki · accumulative
Topic-isolated, compiled bibliographies that grow over time. Any agent can query a wiki. Rex is the primary writer, but anyone can ingest a source. The wiki is the long-term knowledge base for topics I want to track for years. Built by nvk and integrated into Persek OS.
Without a bridge, you end up with two parallel knowledge systems that don't know about each other. Persek OS keeps them connected at three points.
01 · Registry check
Before Rex starts a research run, it checks the wiki registry. If the topic is owned by an existing wiki, Rex defers. There's already a compiled body of knowledge here, no need to re-investigate from scratch. Tie-breaker: if both apply, proceed as Rex and ingest after.
02 · Wiki primes Rex
When Rex proceeds on a topic with existing background, it reads the prior work first. External search then targets only the gaps. No duplicated work, faster runs, and sources chain back to existing research.
03 · Novel findings flow back
After the report, Rex asks: did this run produce sources the wiki doesn't have? If yes, an ingest is offered. Interactive: I confirm or skip. Async: queued. The report footer carries the command. The long-term knowledge base grows from every one-shot investigation.
Every Rex investigation runs the CIS pipeline. Three phases plus a routing step.
Phase 0 · Route, detect, scope
Before any research, decide where the question belongs. Route to Rex or wiki. Detect mode (one-shot vs accumulative). Check compatibility with the question shape. Set scope.
Phase 1 · Context
Check internal knowledge before touching external sources. Read existing wiki articles, prior Rex reports, agent memory on the topic. Avoid re-running investigations that have already been done.
Phase 2 · Information
Mode-specific external retrieval. Multiple passes gather evidence from different angles before synthesis. Multi-round gap analysis continues until the evidence pool is saturated.
Phase 3 · Synthesis
Compile the evidence into the requested output shape. Every claim ties back to a source. Low-confidence claims get flagged for follow-up rather than asserted.
Three modes
Topic mode produces a report. Question mode produces a focused answer. Thesis mode produces a verdict (true / false / unproven) with for-and-against evidence framed in opposition.
Four source tiers
Tier 1: official primary sources. Tier 2: well-cited expert analysis. Tier 3: community signal. Tier 4: noise, only included if Tier 1-3 are silent. Source weighting flows into quality scoring.