MasteryMade · Product Specification · April 2026

The Knowledge Flywheel

A new product category: software that gets smarter with every user.
What it is, how it works, why it's defensible, and what we're building first.

1. The Problem Nobody's Solving

Every SaaS product in the world sits on a pile of user experiments it throws away.

When 500 businesses use the same CRM, each one independently discovers what email subject lines work, what follow-up cadence converts, what lead scoring thresholds matter. They each pay the same tuition — weeks of trial and error — and the platform learns nothing. User #500 starts from the same blank slate as User #1.

The entire AI industry has converged on RAG (Retrieval-Augmented Generation) as the solution: chunk documents, embed them as vectors, retrieve fragments when asked. It works for search. But it doesn't learn. Ask a subtle question requiring synthesis across five sources, and the system re-derives the answer from scratch every time. Nothing accumulates. Day 100 is exactly as smart as Day 1.

Karpathy's insight: knowledge management is a compiler problem, not a search problem. Don't retrieve and re-derive. Compile — synthesize knowledge once, keep it current, and make it available as a persistent, interlinked, self-correcting artifact.

We took that insight and asked a different question: what happens when you compile knowledge not just for one user, but across an entire community of users — weighted by how relevant each person's experience is to your specific situation?

2. The Knowledge Flywheel

The Knowledge Flywheel is a system where every user's experience is compiled into the collective knowledge base in real time, weighted by contextual relevance, and returned to all users as statistical distributions — not expert opinions. The more users contribute, the better everyone's outcomes.

It operates at three layers:

Bridge-Weighting: The Key Mechanism

Not all community knowledge is equally relevant to you. Bridge-weighting scores each compiled insight by its contextual relevance to YOUR specific situation — your industry, your business model, your team size, your goals. The result isn't "here are 50 implementations, good luck reading them." It's:

This is the difference between a library (here are all the books) and a research assistant who read every book and tells you which three pages answer YOUR question.

3. Why This Works (And Where It Doesn't)

What's Validated

What Needs Proving

Known Failure Modes (And How We Handle Them)

These are not optional nice-to-haves. They are structural requirements. Without them, the Flywheel degrades into a confidence machine that scales misinformation.

4. What We're Building First: The Agent Offer

The first application of the Knowledge Flywheel is an AI agent setup service — the same surface the market recognizes (5 agents, 4 pricing tiers), but with fundamentally different economics underneath.

The 5-Agent Stack

Pricing

The Economics Shift

Why Clients Stay

5. The Infrastructure: The Loom

The Knowledge Flywheel runs on infrastructure we call The Loom — a compilation pipeline built on top of the Knowledge Fabric Service (KFS), our entity graph + temporal reasoning engine.

How Compilation Works

Raw sources (sessions, agent outputs, outcomes, decisions)
  → Normalized by source adapters (automatic, zero manual effort)
    → Compiled by LLM into structured articles (concepts, connections, Q&A)
      → Dual-written: filesystem (human-browsable) + KFS (API-queryable)
        → Bridge-weighted by context (industry, model, size, goals)
          → Served to users as distributions with confidence scores
            → Good answers filed back as new articles (compounding loop)
              → Daily lint checks for quality (self-correcting)

Cost per client: ~$25-50/month (Supabase share + LLM calls). At $500-5,000/month revenue, margin is 90-99%.

6. The Standalone Service Vision

The Loom and KFS are designed as independent services, not features embedded in a single product. This means:

Phase 1 serves Forge internally. Phase 2 extracts KFS as a standalone repo with multi-tenant API keys. Phase 3 adds MCP (Model Context Protocol) so any LLM session on any platform can read from and write to KFS automatically. Phase 4 opens KFS to external products.

The endgame: Knowledge Infrastructure as a Service (KIaaS). Not SaaS (software that does things). Not AI-as-a-Service (models that generate things). KIaaS: infrastructure that learns from everything everyone does, and makes everyone better at doing things.

7. What Makes This Defensible

A competitor can copy the code in a week. They cannot copy the compiled knowledge of 50 clients across 20 verticals, bridge-weighted and confidence-scored over 12 months of real outcomes. That's the moat.

8. Go-To-Market

9. Honest Caveats

• Day-one value is template-grade. The Flywheel doesn't help until ~15+ clients. Early clients need a different value prop: personal attention, custom configuration, founder involvement.
• Competitive verticals limit sharing. Two competing plumbers in the same city shouldn't share compiled intelligence. Geographic + vertical isolation is required.
• "Agents that learn" is a hypothesis until proven. We need 3-6 months of real deployment data showing measurable improvement over time. The first 10 clients are the experiment, not the proof.
• Compilation quality depends on LLM quality. If the compiler produces subtly wrong patterns, it scales misinformation. The anti-fragility mechanisms are structural requirements, not optional features.
• The biological evolution metaphor was oversold in early thinking. What we're actually building is real-time collaborative filtering applied to knowledge — the same proven mechanism behind Netflix and Spotify recommendations, applied to business approaches instead of content. It doesn't need a biological metaphor to stand on its own.

10. The One-Line Pitch