Every reputation system fails. Most fail the same ways. This is a field guide to the failures.
Ten specimens below, each a serious attempt, each with people who cared. Some got funding, some got users, some got press. All of them tried to make reputation portable, corroborable, or automated. None of them survived the way their designers hoped. What killed each is instructive: the failures cluster, and the clusters name the laws that any successor has to obey.
Read the specimens first. The laws follow, each paired with a survivor. That rare system, sometimes centuries old, that obeys the law and lives.
Ten specimens
Open Badges (2011–ongoing, hollow)
An open standard for digital credentials, developed at Mozilla in 2011 with MacArthur Foundation backing, later stewarded by IMS Global. A “badge” is a portable digital credential: JSON metadata about who earned what from whom, cryptographically signed by the issuer. By 2013, over 1,450 organizations were issuing badges; millions were minted. By the mid-2010s the standard was technically excellent and functionally unused. The problem was that nobody checked. Employers did not query badge repositories before hiring. Universities did not verify inbound badges before granting credit. A credential nobody looks at is a credential nobody has. Worse, adoption skewed toward populations without conventional credentials, and the badges came to signal the absence of the credential the reader was actually looking for. The standard survives today as plumbing inside enterprise learning systems, absorbed by the incumbent it was built to disrupt. Never a rival to it.
Died from: Law 1 (checking never emerges on its own) and Law 4 (adverse selection stigmatizes alternative credentials).
W3C Verifiable Credentials (2019–ongoing, hollow)
A W3C recommendation for cryptographically signed claims about a subject, expressed in a portable data format. Verifiable Credentials solved the technical problems Open Badges did not: better cryptography, better privacy, better composability. Adopted in specific verticals (educational transcripts, government identity documents) where a specific institution demands the credential. Not adopted as general reputation infrastructure, and for the same reason as Open Badges: outside those verticals, nobody checks. A well-designed credential is not a solution to the checking problem; it is a prerequisite for one.
Died from: Law 1 (checking never emerges on its own).
LinkedIn Endorsements (2012–ongoing)
One-click endorsements of colleagues’ skills, launched September 2012. Users click a button, endorsement appears on the recipient’s profile. Frictionless and viral: two hundred million endorsements in the first month, a billion within six. And within about a year, recruiters had learned to ignore them entirely, recognizing them as reciprocity spam and algorithm-prompted strangers endorsing skills they’d never witnessed. The written recommendation, costly and specific, retains modest value to this day; the one-click endorsement retains none. The word for a sign-off that costs nothing is noise.
Died from: Law 2 (attestation value is proportional to attestation cost).
POAP (2019–2026)
“Proof of Attendance Protocol”: a system for issuing on-chain badges to event attendees, born at the ETHDenver hackathon in 2019. Show up, scan a QR code, receive a POAP that travels in your wallet. Portable, verifiable, permanent, with millions minted. The design solved attendance-recording cleanly. It also demonstrated that attendance is not attestation. A POAP proves you were physically present; it does not prove that you contributed, or that anyone would vouch for you. Collectors accumulated POAPs by the hundreds. Gatekeepers do not read them, because “was in the room” is a very small unit of information about a person, and one available essentially for free. Cryptographic verifiability adds nothing when the underlying claim is costless. The wind-down announced in August 2026 confirmed a second failure mode: the badges live on-chain, but the meaning-and-discovery layer dies with the platform.
Died from: Law 2 (attestation value is proportional to attestation cost; showing up costs almost nothing).
Klout (2008–2018)
A single 0–100 score of a person’s online influence, computed from social media activity — for hundreds of millions of people, most of whom never asked. For a moment it had real traction: conference organizers checked Klout scores when booking speakers, and the number showed up in hiring conversations. Then people noticed the score was gameable, opaque, unrequested, and imposed from outside any community the person belonged to. Public revolt. Lithium Technologies acquired the company for roughly $200 million in 2014 and shut it down in May 2018. The deep failure was structural: Klout tried to make reputation portable by extracting a score from every context and averaging it. Reputation resists that extraction. A number abstracted from the settings where it was earned is not portable reputation; it is a fiction dressed up as one.
Died from: Law 3 (reputation resists leaving its context; portability requires federation, not extraction).
Braintrust (2018–ongoing, hollow)
A talent network for freelance knowledge workers, founded in 2018, out of stealth in 2020, structured as a decentralized cooperative and monetized with a native token launched in September 2021. Talent kept full contract value; clients and members earned tokens for hiring and for referring peers. The mechanism that hollowed the community was subtle: paying members for the act of referring converted referrals from social signal into economic activity. Referrers optimized for volume rather than fit. “I refer you because I know your work” became indistinguishable from “I refer you because I earn tokens if you sign.” The token peaked within weeks of launch and has since lost effectively all of its value. Meanwhile the community discovered it had been running on economic activity that no longer paid enough to sustain the behavior.
Died from: Law 5 (paying for prosocial acts tends to destroy them; the referral was the prosocial act).
Stack Overflow reputation (2008–ongoing, decaying)
Question-and-answer platform for programmers. Users earned reputation by answering questions upvoted by peers, and for a decade that reputation was a genuinely useful signal. Rigorous public participation, visible to peers, audited by voting. Note why it worked: reputation gated real internal resources such as editing rights, moderation tools, closing votes. But it lived entirely inside Stack Overflow. The platform tried to externalize it and failed: Stack Overflow Jobs was discontinued in March 2022, because employers never checked SO reputation at scale despite fifteen years of perfect data. Then came the harder ending: as language models trained on the site’s own answers reduced question volume and veteran contributors drifted away, the reputations stayed on profile pages and stopped meaning anything outside them. Users who spent a decade accumulating rank discovered the rank was a lease from a platform that no longer prioritized it. The coda is sharper still: the platform went on to sign data-licensing deals with major AI labs, selling access to the very answers its contributors had written for free with no record of answerers receiving any cut. The standing was never convertible, even at the moment the work itself was finally monetized. By someone else.
Died from: Law 6 (platform-custodied standing dies with the platform), with a supporting diagnosis of Law 1 (external checking never formed, even with perfect data).
SourceCred (2018–~2022)
An open-source contribution measurement system that used graph algorithms on GitHub, Discord, and Discourse to compute per-person “cred” scores, then distributed treasury allocations proportionally. Well-designed, mathematically rigorous, thoughtfully governed by the projects that adopted it. Two failure modes appeared quickly. First, the moment cred paid, contributors optimized for what the algorithm counted rather than what the community needed: comment frequency skyrocketed, and it was clearly farming. Second, communities fractured over the scoring function; every parameter change felt like reallocating money because it was reallocating money. Several major DAOs adopted, then abandoned. The project wound down around 2022.
Died from: Law 7 (measurement politicizes; peers collude around whichever parameters favor them).
Coordinape (2021–ongoing)
Peer-attribution mechanism for crypto-based decentralized autonomous organizations (DAOs). Each member gets tokens to allocate to other members each epoch; allocations become the basis for treasury distribution. Elegant, small-group, low-friction. Fails when the group is either too small (collusion is trivial; three people can agree to allocate to each other) or too large (voters don’t know each other well enough for the allocation to reflect real observation). The sweet spot is roughly ten to thirty people who all work together intensely which is a limited market, and one that many DAOs outgrow within a year.
Died from: Law 7 (peers collude; the allocation game reveals coalitions faster than it reveals contribution).
On Deck (2019–2022 in its original form)
Cohort-based programs marketed as elite communities of ambitious founders and operators. Selectivity was the value proposition: get accepted and you’re in a room with peers who matter. Growth pressure led to more programs, larger cohorts, weaker selection: over ten thousand founders and professionals at peak. The acceptance rate climbed to accommodate the business model’s need for revenue, and “On Deck alum” carried less signal each year precisely because there were so many of them. In 2022 the company went through two rounds of layoffs months apart, shuttered and refunded programs, and retrenched to a fraction of its ambition. On Deck arrived at the selectivity trade-off the way every credential does, and made the choice most business models require: accept more, dilute the signal, extract revenue from the volume.
Died from: Law 8 (credential value is proportional to selectivity; growth pressure destroys the credential it was distributing).
The Eight Laws
Each of the ten specimens is diagnosable by the same eight laws. The laws are not novel; the specimens surface them repeatedly. Each comes with its survivor: proof the law can be obeyed.
Law 1: Checking never emerges on its own
You cannot build a credential system by publishing well-formed credentials and hoping the demand side arrives. Open Badges and Verifiable Credentials are the paradigm cases: technically sound infrastructures that solved every problem except the one they most needed to solve, which is that nobody looks. Every project in this graveyard solved supply. Not one solved demand. Checking emerges only when the population that signs is also the population that reads: the same person attests on Saturday and sits on the grant panel on Tuesday. There is no other documented mechanism. Any system that treats issuance and verification as separate populations will discover that the verification population never forms.
The survivor: the volunteer coordinator who signs records and later reads them from the other side of the table. She checks records because she makes records; she knows what they cost. Demand isn’t a market you build. It’s a habit that travels inside individuals who inhabit both sides of the record.
Law 2: Attestation value is proportional to attestation cost
The signature that survives is the one where the signer has skin in the game. LinkedIn endorsements and POAPs sit at the same costless end of the spectrum, one on a social graph and one on a blockchain: the one-click endorsement is nearly free, “I showed up” is nearly free, and both produce records that carry the corresponding weight. The substrate changed; the physics didn’t. The cost need not be money; it can be reputation, time, or the observable act of putting one’s own standing next to someone else’s. But it cannot be zero. Costless testimony is a different thing that borrows the vocabulary of real testimony.
The survivors: the term sheet. An attestation collateralized with actual money. The reference letter: costly, specific, still read. The guild mark. Collateralized with collective reputation and enforced by the members who share it. Note the kind of cost that works: staked reputation on a specific witnessed claim. Not solely effort. Effort just kills completion. The signer must have something to lose if the claim is false.
Law 3: Reputation resists leaving its context
Klout tried to build portability by extracting a score from every context and averaging it. The extraction destroyed the very thing it was trying to make portable. Reputation earned in one community means what it means because the people reading it (the gatekeeper, the grant panelist, the receiving local) belong to the community that watched the work happen. Strip the context away and the signal collapses into a number everyone can game and nobody can trust. Every universal score built so far has been rejected by the people it purported to represent.
The survivor: the IBEW traveler system. A journeyman electrician can walk into a distant local, sign the books, and work. Not because his reputation was converted into a universal score, but because the receiving local trusts the issuing local’s training, through a shared institution in which both locals have standing staked. That’s the ingredient Klout never had: a vouching party with skin in the game. Nationalization solved the same problem to a more limited degree: state licensing makes competence portable within a jurisdiction by enforcing minimum standards. But a license ports only the floor, and only for work the state chooses to license. The traveler card ports the whole reputation. Portability was never a cryptography problem. It’s a federation-of-trust problem, and only guild-shaped institutions have ever fully solved it.
Law 4: Adverse selection stigmatizes alternative credentials
Open Badges was rejected in most markets not because the technology failed but because the population that adopted it, on average, was the population that could not or did not go through the standard credential. A credential adopted disproportionately by people avoiding the standard credential gets read as evidence of failure to meet the standard. This is the trap every alternative credential falls into.
The survivor (really an inversion): Facebook launched Harvard-only. A credential’s status is set by its least needy holders. Seed a new form of recognition with people who conspicuously don’t need it, people with impeccable establishment credentials who choose the new one anyway, and the stigma reverses; the new credential borrows their standing before it can lend any of its own.
Law 5: Paying for prosocial acts tends to destroy them
This is the Titmuss lesson (1970), refined by fifty years of motivation-crowding research that followed: paying for a prosocial act tends to reduce it, because the act stops signaling generosity and starts signaling economic activity. Titmuss’s specific claims about the 1960s blood supply have been substantially refined (paid plasma donation now supplies most of the world’s therapies because plasma can be industrially processed), but the underlying finding has replicated across domains, and the mechanism turns on what the payment is contingent on. Braintrust rediscovered it for referrals. The moment reward attaches to the act of endorsing, referring, or vouching, members convert into optimizers of the reward function; the behavior that used to signal care now signals farming, and the signal is dead within a quarter.
The fine print of the crowding-out literature is a design specification, and it turns on what the payment is contingent on. Payment corrupts a prosocial act when it is offered as a bounty for the act, to the person performing it. Titmuss’s donor, Braintrust’s referrer.
The survivor: the honor. The earned title (master, fellow, laureate) the standing that gates access and voice but cannot be sold. And the law’s “tends to” is precise, because the destruction is reversible: a Swedish field experiment (Mellström and Johannesson, 2008) that reran Titmuss found cash payment cut blood donation among women nearly in half, but that the effect disappeared entirely when donors could redirect the payment to charity. Same money, prosocial frame, signal restored. The requirement was never that money be invisible, it’s that the system be inspectable: signing pays nothing, standing can’t be transferred, and anyone can verify both.
Law 6: Platform-custodied standing dies with the platform
Stack Overflow proves this in slow motion. Reputation held only inside a platform is on lease from the platform’s continued attention; when the platform’s priorities changed and its traffic collapsed, twenty years of earned standing stayed on profile pages and stopped meaning anything outside them. The successor pattern separates the notarization of a record from the custody of its display: the record’s authenticity should not depend on any single platform’s continued existence, only on the durability of the notarization substrate.
The survivor: the county recorder of deeds. Land titles outlive the clerks, the software, and sometimes the governments that recorded them, because the record’s permanence was engineered as a separate concern from any institution’s fortunes. Community members’ records should be able to outlive any company, including the one that hosts them.
Law 7: Measurement politicizes; peers collude
Systems that compute a score from an algorithm are politically fragile: every parameter is a fight, every recalculation is a reallocation, and every value chosen is contestable. SourceCred experienced this repeatedly. Systems that let peers allocate directly face the opposite failure: coalitions form and allocate to each other regardless of contribution, as Coordinape found in every group large enough for coalitions. The stable point sits between these failure modes: human attestation of specific witnessed work, attributed to specific witnesses. That structure is durable because the graph of who signed what is not a fight; it is a fact, subject to interpretation but not to negotiation. The pressure toward favoritism never disappears on that ridge, which is why the signing graph must be visible and governed from the start — not bolted on after the first scandal.
The survivor: Modern court systems have needed to establish contribution and fault for centuries, and they have never once done it by computing a score or by polling the room. They call a witness, a named person who saw a specific thing, and make the testimony costly: under oath, on the record, punishable if false, cross-examined in front of everyone. Not a measurement, not a preference; a fact with a name attached, governed visibly from the first word.
Law 8: Credential value is proportional to selectivity; growth wants volume
The credential earns its worth by excluding. Growth strategies that dilute the excluding set destroy the credential they were trying to distribute — On Deck’s story in one sentence.
The survivors: Y Combinator, which stays ruthlessly selective despite every incentive to expand. Harvard, which hasn’t meaningfully grown its class in decades despite a fifty-billion-dollar endowment, because the rejection rate is the product. Michelin, whose anonymous inspectors stake the Guide’s name on every star. Michelin also passes Law 3: a star means the same thing in Tokyo as in Lyon because a single institution with staked standing binds every context it appears in. The law’s reach extends to anyone who builds infrastructure for recognition: the infrastructure should never try to be the credential. The selectivity burden belongs to each community that issues one, and infrastructure that monetizes member volume will produce credentials whose value is inversely proportional to its own success.
Photo by Michelle Li on Unsplash
What this rules in
Read backward, the eight laws are ruthless. Laws 1 and 3 alone kill the grand version of every credentialing dream, because external verification demand has never spontaneously formed and reputation has never survived leaving home. But the laws don’t prohibit everything.
Overlaid, the eight laws describe the shape of a successor. Checking flows through the same population that signs. Each signature risks the signer’s own standing, and never earns them anything. Reputation remains rooted in where it was earned; portability works through federated recognition between communities, not by extraction into a universal score. Records outlast the platforms that host them. Seeding uses people who don’t need the credential, so its early adoption resists stigma. The infrastructure records specific witnessed acts rather than computing scores from parameters, and never tries to become the credential itself. The credential belongs to each community, on the community’s own terms.
None of the ten specimens got all eight right. Several got most of them wrong. The successor need not be more clever than any of them. Only more disciplined about which failures the graveyard has already documented, and structurally incapable of repeating the ones for which the record is now clear.
Every one of these projects believed it was the exception. The laws don’t care what you believe. They only care what the design costs the signer, who owns the record, and whether anyone was ever going to check.
Ken Leiserson builds tools for trust between people who work together. He can be reached at extol.work or on X @kenserson.