Scope Mount

ℹ️ First, set expectations: print one line with this skill's version and its most recent change - the top of metadata.changelog - e.g. scope-mount v1.0 — recent: initial four-role contract. Then proceed. (Skip the banner when applied silently inside a consuming skill.)

Purpose

The scope primitive (ARCHITECTURE.md → "Scope") was built for hygiene: keep the core fingerprint and q-learning clean of encountered material. Its engine default is fail-closed - unset BRAIN_READ_SCOPE → core. That default is correct for the engine and wrong as a retrieval policy, and until 2026-09-04 it was the retrieval policy for every everyday read: /recall, /advise, /decide and eight other skills carried no mount and so read 1,187 of ~6,800 notes. Everything the user has read (Books, Document Insights) and every reference record was invisible to advice and recall. Nobody chose that.

This file is the policy of record for what a read mounts (ARCHITECTURE.md → "Scope Mount Policy"). The engine keeps failing closed; the mount is declared in the playbook, per read, by role.

Contract rule (the reasoning-checks / SOURCE-AUTHORITY.md pattern): this file is the single source of truth. Consuming skills reference it and put the mount inline on their search commands; they never copy the tables. Change a mount here, not in consumers.

State Dependencies

Never writes. Never changes the engine default.

The four read roles

Lookup semantics. "What do I think about X" is answered by the core group; "what do I know about X" by all groups. The user narrows on demand: --scope core (or "only my notes") pins the read to core; --scope <token> mounts exactly that.

/talk is reasoning, not voice (the user, 2026-09-04): the thinking partner speaks from the whole knowledge base, including what the user has read; it still labels what it cites.

Conditional reference mounts (reasoning + lookup only; never standing)

Detection is a cheap title match, run once before the searches (no LLM call, ~50 ms) by the resolver shipped with this skill:

.claude/skills/scope-mount/resolve_mount.sh <voice|reasoning|lookup> "<the question>"
# BRAIN_READ_SCOPE=core,Books,document-insights,company      <- use verbatim on the wide pass
# # trigger company: Trinity
# # extra pass: BRAIN_READ_SCOPE=core,Books/black-swan-taleb  <- a book hit adds a THIRD, precision pass

Rules (whole-word matching; a false positive costs one slightly wider read, nothing else): a Company record title appears whole in the question (single-word titles such as Trinity match as words - agent in a question never mounts BDR agent); any Thinkers name token of 5+ characters appears (the surname is the natural trigger); a Books slug has one distinctive token of 6+ characters or two tokens present, with a stoplist that keeps generic words (thinking, guide, science, agent, ...) from mounting anything. voice always answers core and runs no detector.

A hit adds the token to the mount for that read only (a book hit adds a pass instead). No hit, no mount. A question that is about the org's direction leaves /advise for /canon-advise (dispatch rule in CLAUDE.md).

Never in a reasoning or lookup mount

These stay perception-only (domain-watch, auto-discovery, integrate-recent-notes already mount them for that purpose).

Retrieval shape: core-first, wide-second

Document Insights alone is 2.3x core, so one wide search with a small --limit can push core out of the top-k. Run two passes in the same parallel batch - core as the spine, wide as the evidence layer:

# one query, two populations, one parallel batch
BRAIN_READ_SCOPE=core                          resources/local-brain-search/run_search.sh "<term>" --limit 3 --json
BRAIN_READ_SCOPE=core,Books,document-insights  resources/local-brain-search/run_search.sh "<term>" --limit 5 --json

Neighbourhood calls follow the same mount as the search that found the anchor (run_connections.sh "<note>" at the reasoning mount reaches Books/DI neighbours). --hubs / --bridges / --stats stay core always - they are the fingerprint, not a read.

Autonomous loops keep --no-track (learning hygiene is about who is reading, not what is mounted).

Output: group by scope, label by provenance

Scope is the note_id folder prefix - no engine change:

Load-bearing sources in a conclusion additionally pass reasoning-checks Check 3 - Provenance Weighting, already wired into /advise, /decide and both crystallization paths. Under the reasoning mount that check is the load-bearing voice guard: it is what keeps a wider read from becoming a wider voice.

What the mount does NOT change

• Learning. Since Phase 10a (2026-09-04) the learn-gate is per note (learning.py → trains()): the core hits of a mixed read train q-values and get the q-boost, non-core hits never do. Mounting never switches learning off. (Before 10a any mixed read disabled learning - the "landmine" that kept every read at core.)
• The fingerprint. Hubs, bridges, lifecycle, redundancy: core, by construction.
• The engine default. Unset scope → core, never "all". The policy lives here and in the playbooks.
• The endorsement boundary. Reading wider changes what the agent can see, never what it may endorse. encountered → endorsed remains the human act.
• Thresholds are population-bound. The reasoning mount is its own population; its row lives in SIMILARITY-CALIBRATION.md (Space 2, scoped). Do not carry a core band to the mount or back.

Direct invocation

/scope-mount role=<voice|reasoning|lookup> <question> → runs resolve_mount.sh with those arguments and prints its output: the mount line to use verbatim, the triggers that fired, and any extra precision pass. Useful when writing or debugging a consumer.

Consumers (who declares which role)

/advise · /decide · /think-about-it · /dialectic · /abstraction-ladder · /synthesize-insights · /talk (via thinking-partner) · incubation-loop → reasoning. /recall · /search-vault · /quick-search · /find-connections (+ connection-finder when called from them) → lookup. /get-perspective-on · /create-article → voice (explicit core). Perception, ingestion and maintenance skills keep their existing declared mounts. A consumer with a bare (unmounted) search command is a defect: diagnostics check D3.6 lints for it.