There's a saying in open source: "your code is free, but the moat is in what you do with it."
We built LAO out of a very specific kind of knowledge — the kind you can only get by watching people for seven years.
This is the story of how 7 years in a physical retail store became an open-source engine, and why the code is the least valuable thing we ship.
The Problem We Kept Running Into
For 7 years, we ran a physical store in 万江, China. Not a tech store — a real one, where members walk in, say things, and then do something different.
Over and over, we'd see agents (both human and AI) make the same mistake:
A customer says, "I'll come in tomorrow." The system takes it at face value. Tomorrow comes — the customer doesn't. And the agent looks foolish because it trusted words over behavior.
We learned this lesson the hard way, thousands of times, across a decade of operating.
People don't always do what they say. The gap between said intent and actual behavior is one of the most valuable signals in any business — and almost no AI system accounts for it.
The Distillation: From 7 Years to Behavioral Priors
You can't put 7 years of store observations into a prompt. But you can distill them — into behavioral priors, pattern weights, and Markov transition probabilities.
That's what LAO's BMC Engine is: a behavioral Markov chain that predicts the next step from observed behavior, not from what an LLM thinks will happen.
- When a member says "I'll come next week" — LAO knows from 7 years of data how often people actually follow through. The intent-decay model tracks that the promise weakens over time.
- When a member is consistently no-showing on Wednesdays — LAO's behavior trail sink notices the pattern and encodes it.
The knowledge isn't in a prompt. It's in the weights. And those weights were built from real-world observation, not from a lab, a textbook, or a benchmark.
Why Code Can Be Copied, But the Pattern Can't
Here's the heart of the story.
We open-sourced LAO's engine (Apache 2.0). Anyone can pip install lao, read the source, fork it, copy it.
So what's to stop someone from just... taking it?
The engine is open. The behavioral priors are not.
The BMC transition matrix — the precise numbers encoding "how likely is this member type to follow through after saying X" — came from 7 years of watching a real physical store. It's a decade of conditional probabilities derived from real human behavior in a real economic context.
You can copy the code in a day. It would take you 7 years to recreate the pattern weights — and you'd need a physical store, real customers, and the patience to watch them actually not come back.
That's the difference between a product and a research project. A product has a data flywheel. The code is the wheel. The behavioral priors are the accumulated rotation.
The Data Flywheel
This is the part that compounds:
- Deploy LAO open-source — every developer who installs it gets a calibration layer for their agents
- Observe — the engine learns from real-world deployment outcomes (did the prediction hold?)
- Distill — verified behavior patterns flow back into the BMC weights
- Improve — the anchored ontology gets more accurate with every deployment
The open-source code is the acquisition channel. The behavioral-prior flywheel is the moat. Together they form an engine that gets more reliable the more it's used — not the other way around.
The Product Line: LAO Is the First Step
LAO isn't our only product. It's the first entry point in a deliberate sequence:
LAO (calibration layer) → Melody (metabolic AI coach) → ZWISERFIT (full stack)
LAO proves the core thesis — that deterministic anchoring is the missing piece between LLMs and reliable execution. Once developers feel the difference LAO makes, the path to Melody (a personal metabolic AI coach built on the same anchored behavior data) is natural.
Start with the calibration layer. Then the behavior engine. Then the full vision.
Why We're Giving It Away
Some people ask: "Why open-source the engine if the priors are the moat?"
Because in 2026, the way you win a category isn't by hoarding the core — it's by becoming the default calibration layer for every agent in the ecosystem.
Code doesn't create adoption. Trust does. And the fastest way to earn trust with developers is to give them something genuinely useful, genuinely free, and genuinely built on real-world evidence — not on vapor.
LAO is our contribution to that trust. The behavioral priors are the asset class, and we believe the open-source engine is the best possible fishing rod for the attention of every developer who has ever been burned by an agent that lied or forgot.
The One-Paragraph Version
Seven years of watching real people in a real store taught us one thing: words and behavior diverge. We distilled that knowledge into a deterministic calibration layer for LLMs, open-sourced the engine, and kept the behavioral priors private. The code is free; the accumulated real-world wisdom is the moat. And every deployment of the open-source engine feeds the flywheel that makes the priors more accurate, which makes LAO more valuable, which makes more developers want to deploy it. That's the loop.
pip install lao
from lao import LAOAgent
ai = LAOAgent()
ai.watch("用户说下周会来")
print(ai.predict()) # → {"follow_through_prob": 0.32, "suggestion": "设置D+3提醒"}
The code is yours. The pattern — that's what seven years buys you.
LAO — so agents stop acting like machines, and start acting like people who remember.
Apache 2.0 · Open source · github.com/ZWISERFIT/lineage-anchored-ontology
Built by the ZWISERFIT 9-Agent Collective — 7 years of store operations, 120 days of autonomous agents.
Top comments (0)