Pact economics: why non-competitor cohorts beat scale in federated enterprise AI
A Pact of 8 aligned, non-competing organizations outperforms a Pact of 30 misaligned ones. We walk through the math behind structural likeness, why competitor overlap kills compounding value, and what makes federated enterprise AI actually work in practice.
Key takeaways
- Pact value scales with structural likeness, not member count. Adding a misaligned member dilutes signal and generates false-positive Atoms.
- Three aligned nodes outperform thirty misaligned ones. Verified across every Pact CSSI has launched.
- Non-competition is the unlock: the highest-value Atoms (e.g., "this outreach converts at 18%") flow freely between non-competing peers but are hoarded between competitors.
- Real example: a Ukrainian-focused credit union in Toronto and a general-member credit union in Kitchener share operating structure, regulatory surface, and member-service patterns — but have zero customer overlap. Atoms transfer on day one.
What is a Pact?
A Pact is a vetted cohort of 3–10 non-competing organizations with structural likeness. They pool their successful AI outcomes (Atoms) through a federated learning mechanism. When five members produce similar Atoms, the system fuses them into a Compound — a durable skill every member inherits.
Not a conference. Not a buyer club. A production-grade federated learning cohort with a governance board.
Isn't bigger better? Why small Pacts?
Intuition says more members = more data = better models. Federated learning research says otherwise. Recent work on federated learning cohort sizing (see Google's federated analytics papers, 2024–2026) consistently finds that cohort quality dominates cohort size past a low threshold (~4-6 high-signal participants).
CSSI's observation is specific to enterprise federated AI:
- Structural likeness (shared operating shape, regulatory surface, workflow patterns) → Atoms translate between members on day one.
- Structural mismatch → Atoms look similar at the pattern level but map to different underlying dynamics → false positives → model degradation.
Three aligned nodes outperform thirty misaligned ones. Every Pact we launch, we launch small on purpose.
Why does non-competition matter?
The moment two members compete for the same customers, they stop sharing the Atoms that would most help each other.
The Atom that says "this outreach sequence converts at 18%" is the exact Atom a competitor will never contribute.
So the governance question isn't "how big can we grow the Pact?" It's "who can we add without breaking the trust that makes Atoms flow?"
What do non-competing peers actually look like?
Two real patterns:
Credit union example A Ukrainian-focused credit union in Toronto and a general-member credit union in Kitchener:
- Same Canadian OSFI regulatory surface
- Same core banking stack (both run Celero)
- Same member-service patterns (branch, mobile, member calls)
- Zero overlapping customers (one serves Toronto Ukrainian diaspora, the other serves general Kitchener-area members)
An Atom about "Friday morning SMS fraud alerts get 3× the member response vs afternoon" translates perfectly between them. Their customers never hear of each other.
D2C e-commerce example A pilates-socks D2C brand and a winter-jacket D2C brand:
- Both Shopify + Klaviyo
- Same paid-acquisition dynamics (Meta + TikTok)
- Same churn curves, return policies, review-request cadences
- Zero overlapping customers (different product categories entirely)
"Email at 3pm outperforms 9am by 50%" is a transferable Atom. "Our $80 winter jacket churn curve is bathtub-shaped" is a transferable Atom. No customer ever learns of the other's existence.
What does compounding actually look like year-over-year?
Pact year 1: ~40 Compounds accumulate. Pact year 2: ~180 Compounds. Pact year 3: 400+ Compounds.
A late joiner in year 3 inherits the full 400-Compound history on entry. An outside competitor starts from zero. That gap is the moat.
How do you vet new Pact members?
Three axes:
- Structural likeness — operating structure, regulatory surface, workflow patterns
- Non-competitiveness — verified zero overlap with existing members
- Board-level trust — unanimous governance-board approval, or referral from an existing member
No sales funnel. No onboarding. Just integration, led by a CSSI engineer embedded with the new member during integration.
Where to start
- Read Why we built CSSI for the sovereignty thesis.
- See the Three Pillars of a Pact (Security · Compounding · Vetting).
- Apply for a Pact.
Likeness without competition. That is the math that makes a Pact compound.