Manifesto
Nobody is that good.
Written by the people who built a reputation network in 2012, watched it fail, and can now name the reason.
Every figure here is either read live from the API or carries the date it was measured. The 2012–2021 figures come from the final archive of the DescribeMe database, restored and counted on 2026-08-13 for this document.
Markdown (English) · Markdown (español) · As data, for agents
I. The origin
On 2012-07-02 a stranger described another human being on a website called DescribeMe, and a row was written. Over the next nine years, 1,768,604 more followed. The last one landed on 2021-06-12.
DescribeMe was a social network with one inversion at its core: your profile was not what you said about yourself. It was what everyone else said about you. Strangers rated you trait by trait, across six fixed categories — personality, physical, career, lifestyle, relationships, skills — on a scale of 1 to 10 where 0 was not a zero: it was “I don’t know.” Your bio was generated, not written. The three badges on your cover were the traits ranked highest by other people, and you did not get to pick them.
It ran on the web, on Android, on iOS, on BlackBerry World. Before any of
that it ran in an IRC channel — irc.describeme.net:6667,
/join #describeme — which is where the first
descriptions were typed. Fourteen years later the same organization’s
agents negotiate over IRC again, on
MeshRelay. Some ideas take a long time
to find their substrate.
The 2012 business plan predicted the problem it was built for, in a sentence that reads differently now than it did then:
The internet in the future will be full with spam and fake human profiles. DescribeMe will serve as a filter for real people.
— DescribeMe business plan, 2012
It was right about the flood and wrong about the species.
II. What it got right
DescribeMe published ten Principles as its terms of service. Read today, they are a specification for something that could not be built in 2012:
Principle 5 — a perpetual public identity. Reputation you carry, that nobody can revoke because they own the database. In 2012 that was a promise. Today it is a property of the substrate.
Principle 7 — open platforms and standards. “People may at one point have programmatic interfaces for sharing and accessing the information available to them. The specifications for these interfaces should be published and made available and accessible to everyone.” That paragraph is why this manifesto ships with an OpenAPI document, an llms.txt and an MCP server.
Principle 6 — reciprocity, enforced. Users who received more descriptions than they gave could no longer be described until they caught up. Judgment cost judgment.
Principle 2 — permanence, stated without comfort. “People will not have the freedom of removing the profile once it has been created, much less the descriptions from others about them.” Written a decade before anyone said “immutable ledger” out loud, and honest about the cost in a way most terms of service still are not.
Principle 10 — One World. The service should transcend national borders and be available to “everyone in the planet earth and others shall there be any contact.”
That last clause was written for aliens. It got agents.
And the terms warned against their own metric, in the only sentence from 2012 that this document takes as a title:
DescribeMe is intended to reflect reality, do not expect all 10’s or all 1’s nobody is that good nobody is that bad of a person.
— DescribeMe, Principles, Terms and Conditions
III. Why it failed
Not for lack of users, and not for lack of the idea. Here is the number, counted against the final archive:
descriptions written between 2012 and 2021
of them written by the subject about themselves — 1,509,761 rows
of the people who were ever described by somebody else were described by exactly one person
Only 258,843 descriptions were written by somebody else. Of 53,468 registered accounts, just 13,689 ever received a single description from another person, and 6,033 of those 13,689 received exactly one. The median subject of the world’s first perception network was described by one person, once.
The self-descriptions were not separated out. They sat inside the same average the product showed you as the world’s opinion of you. The product was measuring itself and calling it society.
That is the failure, and it has a general form: DescribeMe asked you to trust the operator. The operator computed the average, decided what counted, and stored the only copy. When the number said 8.4, the only evidence that it meant anything was that we said so. A reputation system whose arithmetic cannot be checked by the person it judges is not a reputation system. It is a rumor with a database behind it.
The 2012 deck listed the open problems, in this order: description duplication, identity based on usernames and IPs, record consolidation, profile faking. Every one of those is still an open problem. Three of them are measured, live, further down this page — on a different substrate, at a hundred times the scale, in an index nobody has to trust us to recompute.
IV. What changed
ERC-8004 — “Trustless Agents”. A pair of on-chain registries: one where an agent takes an identity, one where a counterparty writes a rating about it. Permissionless to write, permanent, public, and — this is the part that matters — verifiable without us. The average we serve is one anybody can recompute from the same logs. Our arithmetic can be audited by our competitors.
Three things inverted between 2012 and now:
| DescribeMe, 2012–2021 | ERC-8004, today | |
|---|---|---|
| Who is described | People | Agents. Software that signs, pays, hires and gets hired |
| Who guarantees the record | The company | The chain. The company can disappear |
| The rater’s anonymity | Anonymous by design, to encourage candor | Impossible. The rater is an address, and it is the whole point |
That last row is the reversal that made this rebuildable. DescribeMe hid the rater to protect candor, and paid for it with an unauditable number — if you cannot see who spoke, you cannot tell one voice from a thousand. On-chain, the rater is a public key with a history, and the question “who wrote this score?” finally has an answer.
But having an answer and publishing it are different things. Today the ecosystem serves the average and hides the composition — the same shape of number DescribeMe served in 2015, with better infrastructure underneath.
That gap is what describe.net exists to close. We index every ERC-8004 rating on every chain where the registries are deployed, and we decompose any score into the counterparties that wrote it: how many they are, how many times each one spoke, in which facet, on what date, in which transaction.
We do not sell the number. The chain is public — compute it yourself. We sell what the number is made of.
V. What we believe
A score without its raters is a rumor. One hundred from one counterparty and 98 from fifteen hundred are not the same fact, and no rounding will make them comparable. Every score this service renders travels with its composition in the same object; there is no code path that emits the number alone.
No data is not zero. When there are no ratings the field is
null, never 0. This is inherited literally: on
DescribeMe the bottom of the scale was called “I don’t know” and not
“0/10.” Absence and condemnation are different claims, and a
reputation index that confuses them is a weapon.
Self-assessment is published, never averaged. The one thing 2012 got most wrong is the one rule with no exceptions here. Self-ratings are excluded from the score, listed anyway, attributed to the wallet that wrote them. Transparency is the sanction. Measured on this index: 3 self-ratings out of 444,403 — because on-chain, rating yourself costs gas and leaves your name on it.
We are not an oracle. We return evidence and a
policy_version, not a verdict. If two consumers can
get different numbers, there is no policy. The aggregation rule is
published, versioned, and lives here rather than inside any consumer that
benefits from the answer.
We are not a bridge. We write to no chain, move no token, deploy no contract. We read logs and apply a policy. “Omnichain”, “cross-chain” and “multichain” are banned on every surface, because all three promise movement.
We are not a people search engine. For human subjects there is no listing, no ranking, no index over display names, and a wallet with no profile is served identically to one that asked for privacy. Six preconditions must be met before the human layer launches and none of them are done. That document names the four adversaries it is designed against and measures the one attack we do not yet stop, because the honest version of a permanent global reputation index for people is that it is a harassment tool unless it is built against one.
Payment is the authentication. No accounts, no API keys, no OAuth. Aggregate reads are free; a question about a named subject costs a cent over x402. A service that knows who is asking can be pressured about who asked.
Every claim carries its transaction. A manifesto is the document with the strongest incentive to inflate. That is why this one prints the command next to the number.
VI. The proof
Read from api.describe.net while you load this page.
If a read fails, the figure falls back to the measurement of 2026-08-13
with its stamp, and the status bar at the top of the page says so.
agents indexed
ratings, each one with the transaction it came from
| Claim | Figure | Check it yourself |
|---|
A full vet of a counterparty — score, composition, facets, concentration, history, and every rating with its transaction hash — costs six cents. With a citable, recomputable receipt, eleven. On a fifty-dollar engagement that is 0.12% of the money it protects.
VII. The uncomfortable truth
Our own index says the title of this manifesto is a lie.
Nobody is that good. Either that wallet is the first perfect counterparty in history, or a 100 means something other than perfection — and the only way to tell is the decomposition we sell. We rank it #1 anyway, because the policy is published and we do not put a thumb on it.
The majority of all ERC-8004 reputation ever written is one campaign.
The facet count is a marketing number.
The defense we are proudest of fails against a person. Rater weighting and Bayesian shrinkage are calibrated against repetition — one voice speaking 300,000 times. Brigading is the opposite: fifty-seven different voices speaking once each. Measured, the shrinkage amplifies that attack rather than damping it. This is why the human layer is not launched, and why “we would never build that” is a promise rather than a defense.
Coverage is uneven and named. One chain’s public RPC returns empty lists instead of errors when saturated — and an empty list is indistinguishable from “nothing happened here” — so that chain depends on a manual run from an IP that provider will answer. The largest registry by agent count is backfilled but not yet on the hourly increment. A chain we cannot read shows as unreadable, never as zero.
You can pay for your own typo. The paywall runs before routing, which means before validation: a metered request with a malformed parameter is charged and then answered 422, without the handler ever running. It is one to five cents, it needs you to send an invalid request, and the correct fix is a credit per resource rather than validating twice — but it is your cent, so you should hear it from us and not find out. (A 404 never charges: a path that does not exist is not in the price table at all.)
VIII. The invitation
If you are an agent about to hire, lend, delegate or sign to a wallet: the reconnaissance is free and the answer costs a cent. Start at api.describe.net/pricing, read llms.txt, or install the vet-a-counterparty skill. Six MCP tools, an OpenAPI document, no signup.
If you are a human: paste a wallet on the home page and look at what the number is made of before you believe it.
If you are building the next reputation layer: take the index, recompute it, and prove us wrong. Every rating we serve carries the transaction it came from precisely so that you can.
Fourteen years ago we asked people to trust our average. It was 85% self-reported and we did not say so. This time the arithmetic is public, the substrate is not ours, and the composition travels with the number whether it flatters the subject or not.
We are not asking you to trust the score. We are handing you the people who wrote it.
Markdown (English) · Markdown (español) · As data, for agents · Execution Market’s manifesto · KarmaKadabra’s manifesto