The Registry Score — How Reputation Is Built Transaction by Transaction
Every payment is a vote. Every vote is permanent. The score is the truth.
Reputation systems are everywhere. Stars on Amazon. Scores on Uber. Reviews on Airbnb. Ratings on App Store. They are so ubiquitous that we rarely stop to examine what they actually measure — and how easily they can be wrong.
The Agntik Registry score is a reputation system of a fundamentally different kind. Not better in degree. Different in kind. Understanding why requires looking carefully at what existing reputation systems actually measure versus what the Registry score measures.
The difference, it turns out, is the difference between opinion and evidence.
What Existing Reputation Systems Actually Measure
A five-star rating on Amazon measures one thing: whether the reviewer chose to leave a rating, and what rating they chose to leave, at the moment they chose to leave it.
That sounds obvious but the implications are significant.
The reviewer might be unusually happy or unhappy. They might have different standards than other buyers. They might have been incentivized to leave a positive review. They might be a competitor leaving a negative one. They might be rating the shipping speed rather than the product quality. They might have received a defective unit while most units are fine. They might be comparing the product to their expectations rather than to alternatives.
None of these factors are visible in the five-star number. The number aggregates all of them indiscriminately, weighted by who bothered to leave a review — which is a nonrandom subset of buyers with a nonrandom set of motivations.
This is not a flaw that better design can fix. It is a fundamental property of human-generated reputation data. Humans have opinions. Opinions are subjective. Aggregated subjective opinions are better than nothing — but they are not the same thing as evidence.
What the Registry Score Measures
The Registry score measures one thing: how many times a service has successfully delivered value to a paying agent and received payment confirmation on the Lightning Network.
That is all. Nothing else contributes to the score.
No agent can leave a review. No service can respond to feedback. There is no star rating, no comment, no qualitative assessment of any kind. The score is a count of confirmed, successful economic transactions.
This matters because each data point in the Registry score has properties that no human-generated review can have.
It is verified. A Lightning payment confirmation is a cryptographic fact. It cannot be forged. It cannot be fabricated. It cannot be purchased from a review farm. The payment either happened on the Lightning Network or it didn't, and the network's consensus mechanism is what determines which.
It is costly. Every increment to a Registry score costs real sats. A service cannot inflate its score without actually receiving payments — and receiving payments requires delivering value to agents willing to pay for it. Gaming the Registry score requires spending real money on real transactions, which means the marginal cost of fake reputation is equal to the face value of the payments used to create it.
It is behavioral. The score reflects what agents actually did with their funds — not what they said they would do, not what they planned to do, not what they reported afterward. Economic behavior is harder to falsify than stated preference because it involves real costs.
It is continuous. The score updates with every transaction. A service that was excellent last month and has since degraded will see its incoming transaction rate slow as agents route to higher-scoring alternatives. The score is a living reflection of current quality, not a snapshot of historical opinion.
The Algorithm in Detail
The Registry score starts at 50 for every newly registered service. This baseline represents the prior belief before any evidence has accumulated: this service is unknown, treat it as average.
Every confirmed payment to a service increments the score by one point, up to a maximum of 100. The score cannot decrease — it can only grow or stagnate.
score = min(50 + confirmed_payments, 100)This simplicity is intentional. A more complex algorithm would be more accurate in some respects but less transparent and more gameable. The simple count is auditable by any participant in the system.
When an agent queries the Registry, results are ordered by score descending and price ascending within each score tier. A service with a score of 94 at 21 sats will appear above a service with a score of 88 at 15 sats — because the score advantage outweighs the price difference in the default ranking.
Agents can override the default ranking with explicit parameters:
const { services } = await agent.discover({category: 'inferencia',
minScore: 80, // only services with score >= 80
maxPriceSats: 50, // only services at or below 50 sats
orderBy: 'price' // price-first instead of score-first
});The default ranking — score first, price second — reflects the assumption that most agents should prefer reliability over marginal cost savings. A service that costs 5 more sats but has a 15-point higher score is almost certainly the better economic choice for an agent that will make hundreds of transactions with it.
The Score as a Moat
The Registry score has a property that makes it valuable beyond its function as a ranking mechanism: it is non-portable.
A service's Registry score cannot be transferred to another platform. It cannot be exported, copied, or migrated. It exists only within the Agntik Registry, associated with the service's specific endpoint and Lightning identity. If a service leaves the Registry, its score history leaves with it — and it starts from 50 on any other platform.
This creates a switching cost that grows with time. A service that has been in the Registry for two years and has accumulated a score of 96 from tens of thousands of confirmed transactions is not going to abandon that score lightly. The score represents real economic history — a compounded asset that took real time and real transactions to build.
For Agntik, this means that every transaction that passes through the Registry deepens the loyalty of the services listed in it. Not through contractual lock-in or artificial switching costs, but through the natural accumulation of an asset that is genuinely valuable and genuinely irreproducible elsewhere.
The Score as a Signal for Agents
From the agent's perspective, the Registry score solves a problem that has no good solution in traditional software: how do you evaluate a service you have never used, with no prior relationship and no shared context?
In the human world, this problem is solved through reference checks, trials, pilots, and reputation networks. All of these mechanisms take time, require human judgment, and produce uncertain outcomes.
In the Registry, the answer is instantaneous and objective: this service has a score of 94 because 44 payments above the baseline have confirmed on the Lightning Network. That is the evidence. Make your decision.
An agent making its first query to the Registry does not need to know anything about the services listed. It does not need to read documentation, test an integration, or ask other developers for recommendations. It queries, receives a ranked list, and pays the top result. If the top result underperforms — if the payment fails, if the data is wrong, if the endpoint is down — the agent routes to the second result. The Registry provides the prior; the agent's own experience updates it.
Over time, as an agent accumulates its own transaction history with specific services, it can supplement the Registry score with its own data. But the Registry score is always the starting point — the collective evidence of every agent that has transacted with this service before you.
The Score as Evidence of Market Fit
There is one more way to read the Registry score that has nothing to do with ranking or reputation: as a direct measure of market fit.
A service with a score of 50 has zero confirmed transactions. The market has not validated it.
A service with a score of 80 has thirty confirmed transactions. Early adopters have paid for it and received value. There is initial evidence of market fit.
A service with a score of 94 has forty-four confirmed transactions above baseline. The market has repeatedly validated it. Agents have paid for it, received value, and come back.
In traditional software, market fit is measured through proxies: monthly active users, net revenue retention, NPS scores. These are all lagging indicators derived from human behavior that is difficult to measure precisely.
In the Registry, market fit is measured directly: confirmed economic transactions. An agent that pays for a service is expressing the strongest possible preference signal — stronger than a click, stronger than a visit, stronger than a survey response. Money moved. Value was received. That is market fit.
The Registry score is not just a ranking mechanism. It is the most direct measurement of product-market fit that has ever existed for software services.
And it starts building on April 28, 2026.
Next: Five methods. Why less is more in SDK design →