Orchestration layer to fill research gaps

The spark

I would like to create an orchestration layer that can fill in the gaps for the research we're doing for the stock market.

What I think you're reaching for

Today the stock-market research is a flat linear runbook (DAILY — fetch sources, promote, ingest, prospect, rank, dispatch). It runs the same steps in the same order every morning, regardless of where the actual holes are. You want to flip the control: an orchestrator that first reads the current state of the wiki — which chains have open/partial steps, which active theses have stale evidence, which concepts have no mechanism yet, which falsifiers were never tested — and then dispatches the right verb at the right gap instead of marching through a fixed sequence. The gap-filling primitives already exist as discrete skills (explore-chain, prospect-chains, establish-chain, signal-emit); what's missing is the layer above them that decides what to fill, in what order, and knows when it's done.

Connections (held loosely)

• DAILY — the thing this sits on top of (or reshapes). It already names the four edges (find / update / refresh / kill chains) and step 1 is literally "fill chain gaps" — but as a fixed-order runbook, not a gap-driven dispatcher. The mental model is there; the orchestration is the missing layer.
• PRINCIPLES — compounding by composition. This invents no new research capability; it composes verbs you already built. The value lives in the seam — a controller that reads vault state and routes to the existing primitive.
• 2026-06-07-finish-research-queue — the sibling half. The queue is "what topics to research next" (intake); this is "what's incomplete in research already underway" (completeness). Same underlying wish: make the loop self-driving instead of hand-shepherded.
• loosely — 2026-06-07-brain-as-add-capture-layer: capture catches the idea; this closes the act / follow-through side — the chronic gap between "started researching X" and "drove X to done."

Where this could go

• A gap census (read-only). A pass that scores the stock-market wiki for incompleteness — chains with open/partial steps, active theses with stale as_of, concepts with no mechanism, untested falsifiers, open questions — and emits a ranked list of holes. That list is the orchestrator's input, and it's useful on its own ("here are the 10 highest-value gaps right now").
• A dispatcher loop. For each top gap, pick the matching verb (weak link → explore-chain, orphan concept → establish-chain, missing cross-source chain → prospect-chains, stale signal → signal-emit), run it, re-census, repeat until a token budget or a "no high-value gap left" stop. This is the difference between a runbook and an agent.
• Generalizes past stock-market. Nothing about census → dispatch is stock-specific; any context with wiki/mechanisms/ and wiki/questions/ could be driven this way. It could become the engine DAILY.md calls, not a one-off script.

A possible next move

Build-flavored, but scope it first. The cleanest first slice is the read-only gap census alone (no dispatch): a skill that reports the highest-value holes in the stock-market wiki. It's immediately useful and is exactly the input the dispatcher needs later. Say the word and we'll scope that slice; otherwise it sits here next to 2026-06-07-finish-research-queue as the other half of the self-driving-research picture.