Ref Bridge

Ref Bridge

ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. ref-bridge v1.0 — recent: initial version. Then proceed.

Why reference data lives in the brain and not in a spreadsheet. A CRM stores clients; a second brain can tell you which of your insights a client is a live instance of. ref-bridge mounts a reference sub-scope alongside core and finds the bridges from an entity to the user's cognitive notes ("Acme Corp, the lighthouse deal ↔ your note on land-and-expand / reference-customer leverage").

It is read-only sense. It surfaces candidates; it never writes, and — critically — it never promotes a record into an endorsed insight. A record informing a thought is fine; a record becoming the user's originated/endorsed thought is the one line the reference kind exists to hold, and it stays the human encountered → endorsed act.

Read first: resources/layered-brains/COMPANY-BRAIN-SCHEMA.md → "The playbook suite" (ref-bridge row) + "Automation boundary"; resources/layered-brains/REFERENCE-SCOPE-SCHEMA.md.

Composes

• /find-connections — the connection-discovery engine; this skill runs it with the reference scope mounted and frames the output as entity→cognitive-note bridges. (Do not re-implement connection logic.)

State Dependencies

Process

Step 1 — Pick the anchor + read its themes

Resolve the entity note (or take a whole sub-scope) and read it to extract its distinctive themes — the economic/strategic concepts it embodies, not its proper nouns (e.g. Acme Corp's engagement = "one lighthouse deal deposits a reusable blueprint + license ARR + a reference case" → themes: optionality / one-bet-many-payoffs, reusable-asset leverage, reference-proof credibility). These themes are the query surface — the bridge is to the insight a fact instantiates, so search by concept.

Step 2 — Theme-seeded CORE search (run BOTH modes; judge relevance)

A reference note's own precomputed edges cluster tightly with its siblings (all engagement notes look alike), and mounting core,Company/<sub> lets those siblings crowd out core in the results. So the method that actually surfaces entity→insight bridges is a theme-seeded search of core alone, using KB-aligned vocabulary — run in both modes, because they fail differently:

# the bridge: which of the user's insights is this entity a live instance of?
BRAIN_READ_SCOPE=core resources/local-brain-search/run_search.sh "<entity themes, KB vocabulary>" --mode static    --threshold 0.35 --limit 6 --json
BRAIN_READ_SCOPE=core resources/local-brain-search/run_search.sh "<entity themes, KB vocabulary>" --mode spreading --threshold 0.35 --limit 6 --json

static is the primary finder for cross-cluster bridges; spreading drifts to the KB's densest hub. Spreading activation flows toward high-centrality nodes — in this KB the Identity / Dopamine / Eight-Circuit cluster. For an entity whose true theme lives outside that cluster (agent architecture, orchestration, GTM — i.e. most Company products, agents, and engagements), spreading returns plausible-but-spurious Identity/Dopamine bridges while static returns the right ones (validated 2026-07-08: Trinity Cloud, themed "agent orchestration / harness is the constraint / architecture over models" → "Most AI agents are chatbots with fancy wrappers" · MOC - AI and Agents under static; → Identity/Dopamine notes under spreading). Only when the entity's theme is the dense decision-science/consciousness core do the two agree (Acme Corp's decision-science theme → Decision Making · Prospect Theory · Loss Aversion under both). Read the candidates and judge whether each insight actually relates to the entity's theme — never take top-k on trust.

Threshold is a floor, not a filter to game. Reference entity bodies are terse; --threshold 0.35 is a safe floor that keeps genuinely-related notes a higher cut would drop. Depending on the query, a 0.5 cut may return the same bridges or an empty set (seed_count: 0) — it is query-dependent, not a universal failure. So if a theme returns nothing, re-phrase toward the KB's actual concepts rather than assuming the threshold is the problem.