The brain
is a parser.

unBasic ⊕ CityCave ⊕ Cave

Brains compose by typed extension, not containment. Each layer is a parser; layers combine with the same combinators (<*>, >>=, <|>) the per-cave record uses one altitude lower.

-- Each brain = a typed extension of the one above
unbasicBrain   :: Brain Agency
ccBrain        :: Brain (Franchise CityCave)     -- extends unbasicBrain
penrithBrain   :: Brain (Cave Penrith)            -- extends ccBrain

-- Composition: left wins on invariants, right wins on specificity
answerKevin    :: Query → Answer
answerKevin    = unbasicBrain <> ccBrain <> penrithBrain

This is not containment. CC living inside unBasic on the filesystem is an implementation choice. The model says: each client brain is a typed extension of unBasic. Future clients (DTC, non-franchise) extend the same way:

otherClientBrain :: Brain (DirectClient Foo)   -- also extends unbasicBrain

What each layer contributes

The composition operator

The carpaccio plan defines the cave record as caveParser = CaveContext <$> profileParser <*> salesParser <*> …. Brain composition is the same shape, one altitude up:

agentRuntime
  = AnswerContext
  <$> unbasicRules        -- always-on invariants
  <*> activeSkill         -- selected by query
  <*> clientContext       -- City Cave
  <*> instanceContext     -- Penrith
  <*> liveQuery           -- the input

• Applicative when layers are independent — order-free, commutative
• Monadic when later layers depend on earlier ones — e.g. which skill loads depends on what the query is about: query >>= chooseSkill

The cascade.

Eight layers in a partial order. Specificity wins on narrowing; invariants are downward-closed; promotion is deliberate.

Three cascade laws

What each layer produces.

A semantic grammar where non-terminals carry meaning, not just syntactic categories. Each layer extends the grammar with productions specific to its altitude. Queries parse by traversing all layers in narrowing order.

unBasic layer · productions

<request>      ::= <data-q> | <draft-q> | <decision-q> | <coaching-q> | <report-q>
<data-q>       ::= "what is" <metric> "for" <subject>
<draft-q>      ::= "draft" <comm-type> "to" <person>
<comm-type>    ::= "reply" | "follow-up" | "summary" | "report"
<voice>        ::= direct + no-fluff + sentence-short + ai-as-collaborator
<delivery>     ::= must-cite-sources · must-not-auto-send · must-log

City Cave (franchise) layer · productions

<subject>      ::= <franchise> | <cave>        -- extends unBasic <subject>
<franchise>    ::= "City Cave" | "the network"
<metric>       ::= <unbasic-metric> | "float volume" | "membership"
                 | "gift card sales" | "NDIS leads"
<escalation>   ::= <issue> → "Jess Altmann" → {Jeremy, Tim, BDC}
<voice-cc>     ::= <voice> + brand-words-only · wellness-register

Penrith (cave) layer · productions

<cave>         ::= "Penrith" | "Kevin's cave"
<person>       ::= "Kevin" | "Kevin Lee"
<sign-off>     ::= "Sincerely," ↵ "Kevin Lee" ↵ <full-sig-block>
<voice-penrith>::= <voice-cc>
                 + numbered-when-multi-point
                 + paper-trail-required
                 + timely-or-he-chases
<page-rule>    ::= ads ::= local? "Penrith page" : "City Cave main page"
                 + NEVER geo-redirect-to-main-homepage
<ndis-rule>    ::= ndis-leads ::= retargeting-only
                 + NEVER email-outreach
<territory>    ::= {Richmond, Liverpool, Ed-Park, Minchinbury, west-open}
                 + if-purchased(t) → STOP ads-in(t)

What "compositional" buys you

A query like "draft Kevin a reply about the ROAS on the Penrith-page custom campaign" is parsed by traversing all three layers:

parse "draft reply about ROAS Penrith-page custom to Kevin"
  = unBasic   : <draft-q> = "draft" + <comm-type>("reply") + "to" <person>("Kevin")
  + CityCave : <metric>("ROAS") on <campaign-type>("Penrith-page-custom")
  + Penrith  : <person> = Kevin Lee · use <sign-off>("Sincerely, Kevin Lee")
              · apply <voice-penrith> (numbered, paper-trail)
              · apply <page-rule> (already Penrith page — confirm)
              · the campaign exists at ROAS 12.34

The Answer is constrained by all productions matched in narrowing order. The skill generating the reply only needs to satisfy the composed grammar — it doesn't need to know which layer each constraint came from.

The filesystem is the graph.

Markdown + YAML frontmatter + wikilinks is a typed graph. No Neo4j until Slice 5+ proves cross-cave query volume. Orange = Slice 0 critical path. Dashed = exists as seam/stub.

graph TB
    subgraph L1[Layer 1 · Constitutional · unBasic]
        CLAUDE[CLAUDE.md
org constitution]
        RULES[.claude/rules/]
        HOOKS[.claude/settings.json
hooks]
    end

    subgraph L2[Layer 2 · Skills · unBasic]
        S_GBRAIN[gbrain-retrieval]
        S_EMAIL[email-triage]
        S_REPORT[monthly-reporting]
        S_FRAN[franchise-navigator]
        S_META[meta-ads-manager]
        S_ANSWER[answer-cave-query
NEW for slice 0]
    end

    subgraph L3[Layer 3 · Knowledge · agency-wide]
        PB[knowledge/playbooks/]
        RS[knowledge/research/]
    end

    subgraph L4[Layer 4 · Client · City Cave]
        GLOBAL[_global/
brand · voice · services]
        FRANCHISE[_franchise/
HQ · Jess · Jeremy · Tim · BDC]
    end

    subgraph L5[Layer 5 · Instance · Penrith]
        PROFILE[profile.md
+ relationship block]
        EMAILS[emails/
synced thread]
        NOTES[notes.md]
        ADSEAM[ads-context.md
EMPTY SEAM]
        SALES[sales-history.md]
        INV[invoicing.md]
        ADH[ad-history.md]
        CHLOG[changelog.md]
    end

    subgraph L7[Live]
        Q[query / email / call]
        A[Answer / draft]
    end

    CLAUDE --> RULES --> HOOKS
    HOOKS -.invariants.-> L2 & L3 & L4 & L5
    L2 -.reads.-> L3 & L4 & L5
    GLOBAL -.narrows.-> PROFILE
    FRANCHISE -.narrows.-> PROFILE
    PROFILE -.specificity.-> Q
    Q --> S_ANSWER
    S_ANSWER --> A

    classDef hot fill:#f97316,stroke:#f97316,color:#0a0a14,font-weight:700
    classDef seam stroke-dasharray: 5 5,opacity:0.5
    class HOOKS,PROFILE,S_ANSWER hot
    class ADSEAM,SALES,INV,ADH seam
    

One query, end to end.

A real email lands; the runtime traverses the cascade; a draft comes out; a human approves; the capture loop updates the relationship intel; the reply sends.

sequenceDiagram
    autonumber
    participant K as Kevin Lee
    participant GM as Gmail
(unbasic.com.au)
    participant ET as email-triage
    participant AR as Agent Runtime
    participant L1 as L1 invariants
    participant L5 as Penrith folder
    participant GB as gbrain-retrieval
    participant FN as franchise-navigator
    participant SA as answer-cave-query
    participant H as Human
(Max/Joe/Brooke)
    participant CL as Capture loop
    participant GIT as Git

    K->>GM: emails a question
    GM->>ET: new msg event
    ET->>AR: classify · owner-query?
    AR->>L1: load invariants
    AR->>L5: load Penrith folder + Kevin profile
    L5-->>AR: intel + history + campaigns
    AR->>GB: retrieve relevant chunks
    GB-->>AR: ranked passages + citations
    AR->>FN: voice + escalation route
    FN-->>AR: register (formal · numbered · sign-off)
    AR->>SA: compose draft
    SA-->>AR: candidate reply + citations
    AR->>H: present draft (NEVER auto-send)
    H-->>AR: edit / approve / reject
    AR->>CL: propose diff to profile + changelog
    CL->>H: show diff
    H-->>CL: approve
    CL->>GIT: commit
    AR->>GM: send approved reply
    

Applicative → Monad.

The boundary between archive and brain. Applicative composition gives you a populated record. Monadic composition gives you a relationship that understands — state threading through every interaction.

graph LR
    Q[query] -->|parsed| ANS[Answer + voice]
    ANS -->|human edits| EDIT[edited reply]
    EDIT -->|diff vs
relationship-intel| DIFF[proposed diff]
    DIFF -->|human approve| COMMIT[git commit]
    COMMIT -->|writes to| L5[Penrith profile
+ changelog]
    L5 -->|next query| Q

    classDef hot fill:#f97316,stroke:#f97316,color:#0a0a14,font-weight:700
    classDef store fill:#141428,stroke:#f97316,color:#f1f1f5
    class DIFF hot
    class L5 store
    

The promotion rule

When does an L5 observation graduate to L4?

if observed_at(layer_5) for ≥3 distinct caves and consistent_signal:
    propose_promotion(layer_5 → layer_4)
    human_review(promotion)

Default: stay at the layer where the pattern lives. Promotion is a deliberate, reviewed event. Without it, everything sediments at L5 and L4 never accumulates.

Two clocks.

Not two cadences — two algebras. The monthly heartbeat is an Applicative product (parallel, commutative). The continuous nerve is a Monadic fold (sequential, stateful, order-dependent).

graph TB
    subgraph Heartbeat[Heartbeat · Applicative · monthly · platform-scheduled]
        META[Meta scheduled reports
→ inbox] -->|parse email| ADH[ad-history.md]
        GADS[Google Ads scripts
→ Sheets] -->|parse rows| GADH[ad-history.md cont.]
        AA[Agency Analytics
scheduled → inbox] -->|parse email| REPORTS[reports/]
        MT[MyTime CSV / Sheets export] -->|parse| SALES[sales-history.md]
        INVS[Invoicing Sheet] -->|parse| INV[invoicing.md]
    end

    subgraph Nerve[Continuous Nerve · Monadic · per event]
        INBOX[unbasic.com.au inbox] -->|gws watch| ETN[email-triage]
        ETN -->|append| EMAILS[emails/]
        ETN -->|trigger| QUERY[query → answer cycle]
        QUERY -->|capture loop| INTEL[relationship intel
+ changelog]
    end

    GWS[gws CLI · universal ingest layer
gmail + sheets · 19 scopes]
    GWS -.feeds.-> Heartbeat
    GWS -.feeds.-> Nerve

    Heartbeat -.->|derived ⟷ source
distinction critical| Nerve

    classDef hot fill:#f97316,stroke:#f97316,color:#0a0a14,font-weight:700
    class INTEL,GWS hot
    

Write discipline

• Heartbeat writes to derived files only. Must be 100% re-derivable from source. Never edited by humans.
• Nerve writes to source files — profile relationship block, notes, changelog. Edited by humans + capture loop.
• The folder layout must encode the distinction. Mixing source and derived corrupts the parser.

What's in. What's out.

Slice 0 is the thinnest vertical that answers one real past Kevin query. The agent-vs-human time delta is the business case for everything that follows.

Ten things a less thorough read misses.

For the Max chat.