What this is: every material decision Captain has made about the empire, with date, context, and rationale. Public. Append-only. Single source of truth for “what did we decide and why.”
How to read: locked decisions at top (the rules we operate by), pending decisions below (what Captain still owes a call on). Each decision links to the source document(s) where it was debated.
Date locked: 2026-05-02 Context: v1
architecture treated Flow as “the parent under which products sit.” v2
made it “internal-only badge.” v3 locked it as a real product — the AI
agent suite for SMEs. Rationale: The 8 universal agents
+ per-tenant brain + concierge/BYOK delivery model only make sense if
Flow itself can be bought. Captain explicitly defined this on
2026-05-02. Source:
yba-flow-vision-2026-05-02.md ·
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 Context: v2
had two engines (back-office + content). Adding the agent layer required
a third. Rationale: ERPNext is great for
accounts/HR/CRM but never going to be a content/posting platform. The
agent runtime needed its own home (Flow Core + n8n). Three engines =
each does what it’s good at. Source:
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 Context: v1
made them 4 sibling products. v3 collapsed them to 4 paint jobs on one
engine. Rationale: The 8 Layer 1 agents are universal —
every business needs HR + Accounts + Front Desk. Verticals add Layer 2
specialists + branded front door + industry data. New verticals become
weeks, not months. Source:
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 Context: A
doctor never shops for “Flow.” She shops for medical practice software.
Rationale: Vertical brands carry trust signals Flow
cannot. Flow only surfaces on the BYOK developer/SME tier
(flow.ybaedge.com). Source:
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 Context: v3
specialists recommended Postgres + pgvector + markdown vault as 3-layer
brain. Captain rejected — “must be the exact storage solution we use.”
Rationale: Cheap (files on disk), effective (LLMs read
markdown best), proven (Captain’s own second brain), auditable (owner
browses in Obsidian). Postgres for transactional data only — never
knowledge. Source:
project_yba_brain_storage memory ·
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 Context:
Captain confirmed each product has different specialists with different
skills + SOPs. Rationale: Layer 1 (8 universal
foundation agents) is shared across every tenant. Layer 2 specialists
are vertical-specific (Med = 42 agents in 5 teams). Layer 3 is the
per-tenant markdown brain. Source:
yba-empire-architecture-v3-2026-05-02.md ·
yba-med-layer2-2026-05-02.md
Date locked: 2026-05-02 Context:
“AI must be malleable — change LLM per cost/accuracy/consistency.”
Rationale: Each task declares
accuracy_critical / consistency_critical / cost_critical / balanced.
Router maps to model from one config file. Captain swaps models by
editing JSON, never code. Source:
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 Context:
Captain asked if we should build everything from scratch instead.
Rationale: 4 of 5 specialists strongly recommended
hybrid. Founder economics math: ground-up = R1.2-1.7m year-1 cash
forfeit, day-job exit pushed back 9-13 months, <5% Q1 success
probability. Moat is agents + brain + customer relationships — NOT the
plumbing. Stripe / Shopify / Notion / Yoco all rent commodity infra and
command top multiples. Source:
yba-build-vs-rent-decision-2026-05-02.md Re-read
trigger: annually, first weekend of January. Or when the urge
to rebuild returns.
Date locked: 2026-05-02 Context:
Two delivery tiers (R45k+R18.5k/mo concierge vs R3.5k/mo BYOK).
Rationale: R6k Office Brain client = first concierge
tenant. Concierge funds BYOK build. R20k/mo × 10 concierge = R2.4m/year
war chest for engineering. BYOK soft-launches month 7-9, public month
10+. Source:
yba-empire-architecture-v3-2026-05-02.md ·
yba-empire-development-plan-v1-2026-05-02.md
Date locked: 2026-05-02 Context:
Captain works 18 hrs/week at night. Calendar dates are guesses.
Rationale: Each gate has measurable triggers (e.g. “5
of Captain’s brands fully on platform 30 days · 0 P1 bugs · synthetic
Tenant B test passing · onboarding <30 min · indemnity bound”). Gates
fire when conditions met, not when calendar says so.
Source:
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 (evening)
Context: Path-to-market specialists flagged that Edge
customer learnings transfer DEEP for agency-adjacent verticals and
SHALLOW for Med/Terra/Realty. Rationale: Each new
vertical needs its own dogfood proxy — one doctor for Med, one farmer
for Terra, one estate agent for Realty. Their actual pain becomes the
Layer 2 spec. Captain quote: “yes i intend to have
a real client using YBA edge to help us build industry specific
solutions.” Source:
project_yba_practitioner_per_vertical memory
Date locked: 2026-05-02 Context:
Captain wants full alignment across all products + infra before code
ships. Rationale: Building before alignment risks the
strategic-roadmap “single failure mode” — half-Flow half-custom hybrids,
9 months wasted on infrastructure no customer sees. Iterate the
architecture doc through every product vision until Captain says “now
build.” Captain quote: “we will continue till i
have you on the same page as me and then once we found the best path for
all the products and infra then we will start building.”
Source: feedback_align_before_build
memory
Date locked: 2026-05-02 Context:
Component auditor’s verdict on Postiz. Rationale:
OAuth-and-token-refresh maintenance burden is structurally bigger than
feature value. Stay on fork-with-cherry-picks. Replace only if our
customizations exceed ~30% of codebase or upstream stops absorbing our
PRs. Source:
yba-build-vs-rent-decision-2026-05-02.md
Date locked: 2026-05-02 Context:
Component auditor flagged these as the most rebuildable pieces.
Rationale: n8n custom replacement could be 30% of n8n’s
size and a better fit at the 3-layer agent architecture — but only after
5+ tenants tell us which 10 integrations matter. LiteLLM is dead weight
already (~1 week to inline) but keep in Phase 0 to ship faster. Audit
both at Gate 2. Source:
yba-build-vs-rent-decision-2026-05-02.md
Date locked: 2026-05-02 Context: v3
architecture identified Office Brain as a confused product naming.
Rationale: It was always Edge stage 2 / Flow concierge
tier. Confusing to have a separate product name for what’s just a tier
of the platform. Source:
yba-empire-architecture-v3-2026-05-02.md
Date locked: 2026-05-02 evening
Context: Captain reviewed 5-specialist documentation
audit and approved the cleanup plan. Rationale: Docs
were written in one fast iteration. They overlap, repeat, contradict.
Cleanup: archive 4 obsolete docs, build Index + Decisions Log, split
Architecture into Architecture + Strategy, search-and-replace “Office
Brain” → “Flow concierge”, reconcile naming/number drift.
Captain instruction: “1 yes 2 you do whats best i trust
u 3 you do whats best i trust u” Source:
yba-empire-doc-audit-2026-05-02.md Niki’s calls
under Captain trust: - Decisions Log = public on web (forcing
function + share-able URL) - Ground-up variant + Build-vs-Rent =
un-published from web, kept in wiki
Awaiting: yes / no / push back on a specific bit
Source:
yba-empire-development-plan-v1-2026-05-02.md
Awaiting: yes / wait / different scope
Detail: Repo creation + code skeleton (Flow Core Hono
service, LLM Router FastAPI stub, Brain Provisioner, synthetic Tenant B
test in CI). No customer-facing changes. Source:
yba-empire-development-plan-v1-2026-05-02.md
Awaiting: Michalsons / Webber Wentzel / Novation
Detail: Once chosen, Niki sends the brief. ~3 weeks
turnaround. Required before concierge client #1 onboards (Week 4).
Source:
yba-empire-development-plan-v1-2026-05-02.md
Awaiting: yes / wait Detail:
Hollard / Santam / specialist broker — R8m cover, ~R8k/month premium.
Required before concierge client #1 onboards. Source:
yba-empire-architecture-v3-2026-05-02.md
Awaiting: Captain writes vision document (analogous to Edge / Med vision docs) Detail: Once provided, Niki spins up Layer 2 specialist sweep for Terra (5 researchers, same shape as Med Layer 2 design) Triggers: unblocks Gate 5 (Terra planning)
Awaiting: Captain writes vision document Detail: Same shape as P-005 Triggers: unblocks Gate 6 (Realty planning)
These decisions have explicit re-read schedules:
| Decision | Re-read when |
|---|---|
| D-008 (stay hybrid) | Annually, first weekend of January. Or when the urge to rebuild returns. |
| D-014 (n8n + LiteLLM Phase 3 audit) | After Gate 2 fires (5+ paying clients × 3 months) |
| D-013 (Postiz fork) | If our customizations exceed 30% of codebase OR upstream stops absorbing PRs |
| D-007 (LLM routing) | Quarterly — review LLM cost per tenant and re-tune routing config |
| All Layer 1 promotion candidates | Quarterly — Captain reviews Meta-Brain patterns at quarterly meeting |
Maintained by Niki. Append-only. Captain reviews monthly. New decisions land at the bottom of “Locked” with sequential D-NNN number; old decisions never get edited (only superseded by a new entry).