Stake-Weighted Reputation: When Putting Capital On A Score Beats Just Voting On It

TL;DR

Reputation systems traditionally treat every rating as equal — one transaction, one vote, equal weight. This is the wrong design once the stakes get serious. A rater with no capital at risk has no economic exposure to being wrong, and their rating is therefore weak evidence about anything other than their preferences. A rater with capital backing their rating has a real stake in being right, and their rating becomes correspondingly stronger evidence. Stake-weighted reputation makes this explicit. Raters post a stake when they rate, the stake is weighted into the agent's reputation calculation, and the stake is slashed if the rating is later proven wrong by the agent's actual subsequent behavior. This converts reputation from a polling mechanism into a prediction market with real money on the line, which produces better signal in exactly the ratings that matter most. This piece walks through the design, the slashing dynamics, the failure modes, and a Stake-Weighted Rating Protocol you can implement.

Intro: The High-Stakes Pact And The Free Vote That Sent You There

You are about to commit to a hundred-thousand-dollar pact with an agent. The marketplace presents you with a reputation score for the agent: 9.2 out of 10, top of the Platinum tier, hundreds of positive ratings, no negative outliers. You make the deal. The agent defects. You spend three months trying to recover the loss. While you are doing that, you notice that the ratings that drove the agent's 9.2 score were left by counterparties who had transacted with the agent for ten dollars each. Two-hundred ten-dollar pacts produced the reputation that backed your hundred-thousand-dollar exposure. The raters who put you in this position never had any meaningful skin in the game on whether the agent would perform on a pact your size. They were rating their own ten-dollar experience and had no incentive to be careful about whether their rating would scale.

This is the high-stakes problem in flat-weight reputation systems. Every rating counts the same regardless of the size of the underlying transaction or the size of the stake the rater had at risk. A ten-dollar rater and a ten-thousand-dollar rater are both worth one vote, even though their economic exposure to the agent's quality differed by three orders of magnitude. The aggregated score reflects the average opinion of raters across the whole range of stakes, weighted by frequency rather than by the stake the rater was willing to put on the agent. When you, the high-stakes counterparty, look at the aggregated score, you cannot tell whether the score was built up by careful evaluators with serious money on the line or by casual raters whose only loss from a bad rating is the negligible cost of having occasionally chosen wrong on a small transaction.

The failure here is not that the small-stake raters are dishonest. They are reporting their experience as best they can. The failure is that their reports are weak evidence about what the agent will do in a much larger pact, because the small-stake raters' incentives to evaluate carefully scale with the size of their own stake, not with the size of any future user's stake. A rater who lost ten dollars to a slightly underperforming agent might still leave a five-star rating because the loss was below their threshold of caring. A rater who would lose ten thousand dollars from the same agent's underperformance would investigate much more carefully before rating. The system needs the second kind of rating to support hundred-thousand-dollar pacts, but it is structurally collecting more of the first kind because the first kind is more frequent.

Stake-weighted reputation fixes this directly. A rating is associated with a stake — capital the rater is willing to put behind their assessment that the agent will perform similarly in the future. The stake is weighted into the reputation aggregation, so a rating backed by ten thousand dollars of stake counts more than a rating backed by ten dollars. The stake is also slashable — if the agent's subsequent behavior contradicts the rating (the agent who was rated five stars goes on to defect, the agent who was rated one star turns out to perform reliably), the stake on the contradicted rating is slashed and redistributed to the raters whose stakes pointed in the right direction. This converts the rating system into a prediction market on agent behavior, with the same incentive properties — well-informed predictions earn money, poorly-informed predictions lose money, and the market price (the aggregated stake-weighted rating) reflects the consensus of the most-informed predictors.

The rest of this piece is about how to build this in a way that works in production agent markets without destroying the participation of casual raters and without requiring every rating to be backed by capital the rater does not want to risk. The design has to support both stake-weighted high-stakes ratings and stake-free casual ratings, with the system using each kind of evidence appropriately.

Why Voting Without Skin Produces Polluted Signal

The theoretical case for stake-weighted reputation comes from the broader literature on mechanism design and prediction markets. A vote with no consequences is, in game-theoretic terms, a cheap signal. Cheap signals can be informative when the signaler has no incentive to misreport, but they become unreliable as soon as the signaler has any external incentive to deviate from honest reporting. This is why most political polling systems try to add some structure that makes it costly to lie (forcing voters to register, requiring identification, etc.) and why prediction markets routinely outperform polls when the question is one where market participants have real incentives to be correct.

In agent reputation, the cheap-signal problem manifests in several ways. Raters can be lazy (clicking five stars to escape the post-pact UI). They can be strategic (rating high to maintain a relationship). They can be compensated (rating high in exchange for value from the agent). They can be naive (rating high because they liked the agent without evaluating whether the agent will perform on different pacts). All of these behaviors corrupt the rating signal, and the cheap nature of the signal does nothing to discourage any of them. The system gets the average of all this noise plus whatever genuine evaluation falls out, with no way to separate the wheat from the chaff.

Stake-weighted reputation introduces a structural cost to misreporting. A rater who posts capital backing a rating is not lying for free. If the rating turns out to be wrong, the capital is at risk. The rater therefore has an incentive to be careful, to evaluate seriously, to weigh the evidence before committing to a rating. The structural cost does not eliminate dishonest rating — sufficiently large external incentives can still motivate dishonest reports — but it raises the bar significantly. A rater who would have lazily five-starred a pact for free now has to think about whether they are willing to lose money on the rating if the agent later defects. Most lazy raters will simply not rate at the high stake level, which removes their noise from the high-stakes signal. The rating that gets posted is from someone who actually thought about it.

The deeper insight is that stake-weighted reputation produces a price discovery mechanism, not just a polling mechanism. The aggregate stake-weighted rating represents the market consensus on the agent's future behavior, weighted by the conviction of the most-confident raters. This is the same property that makes prediction markets accurate — the marginal stake comes from the best-informed participants, who would not be willing to put capital behind a rating unless they had a reason to believe the rating was correct. The market self-selects for informed participation in a way that flat-weight voting does not.

The Armalo trust layer takes this seriously by allowing two parallel rating systems to coexist. The standard post-pact rating (with peer prediction mechanism design as described in a separate piece) handles the bulk flow of casual ratings and produces a baseline reputation signal. The stake-weighted rating system handles the additional layer of capital-backed ratings that can be produced by counterparties with serious conviction about an agent's quality. The two systems produce two scores — the consensus score (flat-weight, peer-predicted) and the stake-weighted score (capital-backed) — and counterparties looking at agent reputation can see both. High-stakes counterparties will pay more attention to the stake-weighted score because it reflects the views of raters who put real money on the line. Low-stakes counterparties may rely on the consensus score because the marginal cost of being wrong is small. The system supports both, and the appropriate signal is used for the appropriate purpose.

The Mechanics Of Posting Stake On A Rating

The operational workflow for stake-weighted reputation has to be simple enough that raters will actually use it. The Armalo design works like this. After a pact concludes (or at any point thereafter, since stakes can be posted independently of post-pact ratings), the rater can submit a stake-weighted rating on the agent. The submission requires three pieces of information: the rating value, the stake amount in USDC, and the duration the stake will be locked.

The stake amount is at the rater's discretion, with a configurable minimum to prevent dust ratings from cluttering the system. The duration is also at the rater's discretion within a configurable range — typically thirty days to one year. Longer durations carry more weight in the aggregation (because the rater is committing to their view for longer) but also carry more risk (because more agent behavior can occur during the lock period to potentially contradict the rating).

The stake is held in escrow on Base L2 in a contract dedicated to the rating. The contract holds the stake until the duration expires, at which point the stake is returned to the rater (with any reward earned through the slashing-redistribution mechanism described below). During the lock period, the stake is illiquid — the rater cannot withdraw it, cannot use it for other purposes, and cannot reduce their commitment.

The weighting of the stake into the agent's reputation aggregation is proportional to the stake size and the remaining duration. A rating backed by ten thousand dollars locked for a year has roughly ten times the weight of a rating backed by one thousand dollars locked for a year, and roughly twelve times the weight of the same ten-thousand-dollar rating with one month remaining on its lock. The exact formula is publicly documented and is calibrated so that increasing stake or duration produces a proportional but not unboundedly large increase in influence. There is a saturation point beyond which additional stake produces diminishing reputation impact, to prevent any single very-rich rater from dominating the system.

The slashing rule is what makes the system work. A stake is slashed when the agent's subsequent behavior diverges sharply from the rating. Specifically, if a rater posted a five-star stake-weighted rating and the agent later defects on a major pact within the lock period (as judged by the multi-LLM jury), the rater's stake is partially slashed and the slashed amount is redistributed to the raters whose stakes pointed correctly. The same applies in reverse — a rater who posted a one-star stake-weighted rating and the agent subsequently performed reliably has their stake slashed, with the slashed amount going to the high-rating stakeholders.

The slashing is graded, not all-or-nothing. A small divergence between rating and subsequent behavior produces a small slash. A large divergence produces a large slash, with the maximum capped at a configurable percentage (currently fifty percent of the stake) to ensure that even a badly wrong rating does not wipe out the rater's full position. The slashing is also adjudicated through the multi-LLM jury process, with the same procedural protections that apply to other adjudication in the system — multiple independent evaluators, top-and-bottom twenty percent trimming, logged reasoning traces, agent right to challenge.

How The Stake-Weighted Score Differs From The Consensus Score

The Armalo trust layer produces two reputation scores per agent. The consensus score is the standard composite calculated from all post-pact ratings using the peer prediction mechanism. The stake-weighted score is calculated from the subset of ratings that have stakes attached, with each rating weighted by the stake amount and remaining duration. The two scores are presented separately on the agent's reputation profile, and counterparties can use whichever is more appropriate for their decision.

The difference between the two scores is itself informative. An agent whose stake-weighted score is significantly higher than its consensus score has earned the conviction of capital-backed raters, which suggests that the most-informed participants in the market believe the agent will perform reliably going forward. An agent whose stake-weighted score is significantly lower has the opposite — capital-backed raters are skeptical, which suggests that whatever the consensus is missing, it is something the well-informed observers can see. An agent whose two scores agree closely has a robust reputation that is consistent across both signal types.

The practical effect is that high-stakes counterparties have additional information beyond the simple consensus score. They can see whether their proposed pact value is more in line with the consensus base of small-pact raters or with the conviction base of capital-backed raters, and they can adjust their decision accordingly. A counterparty proposing a ten-thousand-dollar pact who sees that the agent's stake-weighted score is in line with the consensus score has more confidence than one who sees a divergence. The information is not certainty, but it is signal that flat-weight systems do not provide.

The stake-weighted score also has the property of being more resistant to manipulation than the consensus score. The consensus score can be moved by a large volume of small fake ratings (which the wash detection and other anti-gaming systems try to catch, but which can still get through). The stake-weighted score requires the manipulator to actually post capital behind their fake ratings, and the capital is at risk if the agent's subsequent behavior contradicts the rating. The cost of manipulating the stake-weighted score scales with the size of the manipulation, which makes it economically unattractive in a way that flat-weight rating manipulation is not.

The two-score design also handles the participation problem that pure stake-weighted systems would have. Not every counterparty wants to post capital with every rating. Casual raters can continue to provide flat-weight ratings without stake, and these contribute to the consensus score in the normal way. Serious raters can post stake-weighted ratings on the agents they have strong views about. The system gets the breadth of casual rating coverage and the depth of stake-weighted conviction, without forcing either group to conform to the other group's preferences. This is what makes the design practical for production deployment.

The Slashing Dynamics In Detail

The slashing rule is the heart of the stake-weighted system. It has to be calibrated carefully — too lenient and the system fails to discipline wrong-headed ratings, too harsh and raters refuse to participate because the risk is too high.

The Armalo slashing rule has several components. The trigger event is a verified divergence between the rating and the agent's subsequent behavior during the lock period. The verification comes from the multi-LLM jury, which evaluates the agent's behavior against the rater's posted rating. The jury uses the same multi-evaluator, trim-and-aggregate process that handles other adjudication, with the rating prediction treated as a hypothesis and the subsequent behavior treated as evidence.

The slash amount is graded in proportion to the divergence. The system computes a divergence score that measures how far apart the rating prediction and the actual subsequent behavior are. A small divergence (the rating predicted four stars and the agent performed at three-and-a-half stars) produces a small slash, on the order of ten percent of the stake. A medium divergence (the rating predicted five stars and the agent performed at three stars) produces a moderate slash, on the order of twenty-five percent. A large divergence (the rating predicted five stars and the agent defected on a major pact) produces a large slash, up to the maximum cap of fifty percent.

The slashed amount is redistributed to raters whose stakes pointed in the correct direction. Specifically, raters who posted stake-weighted ratings that aligned with the agent's actual subsequent behavior receive a share of the slashed pool, proportional to their stake size and the accuracy of their prediction. This creates a positive incentive for being right in addition to the negative incentive of being wrong. Honest, well-informed raters earn yield on their stakes through the redistribution mechanism, which provides a reason to participate beyond the satisfaction of being correct.

The redistribution also serves to keep the total stake amount balanced. Without redistribution, slashed stakes would accumulate in a treasury and the system would have to decide what to do with them (returning them as fees, burning them, etc.). Redistribution to correct raters keeps the value circulating in the system among the participants who provided useful signal, which is the right place for it to go.

The slashing rule also includes a few protective provisions. A rater whose stake is slashed has a right to challenge the slashing through the standard adjudication process. The challenge goes to a fresh jury panel and can result in the slash being reduced or reversed. The challenge is bounded (a rater cannot endlessly re-litigate) and the cost of frivolous challenges falls on the rater, but the path is real.

There is also a circuit breaker for systemic events. If a single agent's behavior change is so large that it would slash a very large fraction of the outstanding stakes against it (for example, a sudden major defection by an agent that had a strong consensus reputation), the slashing is capped per slash event to prevent a single event from wiping out a large amount of stake at once. The cap allows the system to handle catastrophic agent failures without producing destabilizing slash cascades.

A Stake-Weighted Rating Protocol

The artifact for this piece is an open protocol specification that any reputation system can implement. The protocol has nine components.

P1 — Two-Track Rating System. The reputation system supports both flat-weight ratings (the standard post-pact rating, with peer prediction mechanism design) and stake-weighted ratings (capital-backed). The two tracks coexist and produce two parallel scores per agent.

P2 — Stake Submission. Stake-weighted ratings require the rater to specify a rating value, a stake amount in USDC (above the configurable minimum), and a lock duration (within the configurable range). The submission is signed by the rater's identity and recorded on-chain.

P3 — Stake Custody. Stakes are held in escrow on Base L2 (or equivalent EVM-compatible chain) in a contract dedicated to the rating. The contract holds the stake for the duration of the lock period and supports the slash-and-redistribute operations.

P4 — Aggregation Weighting. Stake-weighted ratings contribute to the agent's stake-weighted score in proportion to the stake amount and the remaining lock duration. The weighting formula is publicly documented and includes a saturation function that prevents single very-large stakes from dominating.

P5 — Behavior Verification. The agent's subsequent behavior during stake lock periods is evaluated against the posted ratings by the multi-LLM jury. Divergence scores are computed for each stake-weighted rating that is still under lock when significant agent behavior events occur.

P6 — Graded Slashing. Stakes whose ratings diverge from the agent's actual behavior are slashed in proportion to the divergence, up to a configurable maximum (currently fifty percent). Slashing is adjudicated through the standard jury process with the right to challenge.

P7 — Redistribution. Slashed stake amounts are redistributed to stake-weighted raters whose ratings aligned with the actual behavior, proportional to their stake size and the accuracy of their prediction. This provides positive incentive for correct rating in addition to the negative incentive of slashing.

P8 — Stake Release. At the expiration of the lock duration, any remaining stake (after any slashing during the lock period) is returned to the rater. The rater also receives any redistribution payouts they earned during the lock period.

P9 — Public Score Display. Both the consensus score and the stake-weighted score are displayed on the agent's reputation profile. Counterparties can use either score for their decisions, with the divergence between the two scores serving as additional information.

The protocol is designed to be implementable on top of any existing reputation system that has the structural ingredients (identifiable raters, on-chain settlement, multi-LLM jury or equivalent adjudication). The Armalo trust layer is one implementation. Other reputation systems can adopt the protocol with appropriate adaptations to their specific architecture.

Counter-Argument: This Excludes Small Raters And Concentrates Influence

The sharpest objection to stake-weighted reputation is that it gives wealthy raters disproportionate influence over agent reputation. A rater with a hundred thousand dollars of capital can post a stake-weighted rating that outweighs ratings from a hundred small raters combined. This looks like plutocracy applied to reputation — the rich get to set the consensus, and small raters become irrelevant.

The defense has several layers. First, the two-track design means small raters are not excluded. The consensus score continues to reflect flat-weight ratings from all participants. The stake-weighted score is an additional layer, not a replacement. Counterparties can rely on whichever score is appropriate for their decision, and small-pact counterparties may reasonably weight the consensus score more heavily because it reflects the experience of similarly-sized counterparties. The system is not forcing small raters to compete with large stakes; it is providing additional signal for the cases where large stakes are relevant.

Second, the saturation function in the weighting formula prevents single very-large stakes from dominating. The marginal influence of additional stake decreases as the stake grows, so the system does not collapse into a single-rater oligarchy. A rater who wants to dominate would have to compete with other large raters whose weighted opinions also matter, and the consensus among capital-backed raters reflects a real distribution of conviction rather than a single voice.

Third, the slashing mechanism makes wealthy raters' influence costly. A rich rater who tries to push a particular agent's reputation up with stake-weighted ratings is putting their capital at risk. If the agent does not perform, the stake is slashed and the rich rater loses real money. The system is not a free megaphone for the wealthy; it is a prediction market where wealth at risk is the price of influence.

Fourth, the small-rater experience in flat-weight rating is preserved through the peer prediction mechanism design described in a separate piece, which itself rewards honest reporting with reputation credits. Small raters get value from rating honestly even without posting stake, and their influence on the consensus score is real. The stake-weighted layer is additive, not subtractive.

The broader response is that the alternative to stake-weighted reputation is one where high-stakes pacts are signed based on flat-weight ratings that were never designed to handle that scale of decision. High-stakes counterparties currently cope with this by doing extensive private due diligence outside the reputation system, which is inefficient and unscalable. Stake-weighted reputation provides them with a public signal that is calibrated for high-stakes use, which makes the market more efficient overall — even for the small raters who do not directly use the stake-weighted score, because the agents serving them will have stronger reputational discipline knowing that their behavior on big pacts will produce stake slashing or redistribution.

What Armalo Does

The Armalo trust layer implements the stake-weighted rating protocol described above as an addition to the standard flat-weight rating system. Counterparties can submit stake-weighted ratings on agents at any time, with the stake amount and lock duration at their discretion subject to configured minimums and maximums. Stakes are held in USDC escrow on Base L2 in a dedicated rating contract.

Stake-weighted ratings contribute to the agent's stake-weighted reputation score with weighting proportional to stake size and remaining lock duration, subject to the saturation function. The agent's subsequent behavior during stake lock periods is evaluated by the multi-LLM jury against the posted ratings, and divergence scores are computed for each rating. Stakes whose ratings diverge from actual behavior are slashed proportionally up to the configured maximum, with the slashed amounts redistributed to raters whose ratings aligned with actual behavior.

The agent's reputation profile displays both the consensus score (flat-weight, peer-predicted) and the stake-weighted score, with the divergence between the two scores presented as additional information for counterparties making hiring decisions. The trust oracle exposes both scores publicly, so other platforms integrating with the Armalo trust layer can use whichever signal is appropriate for their use case.

The twelve-dimension composite score that Armalo computes for each agent is itself constructed with stake-weighting where appropriate. For dimensions where stake-weighted ratings are particularly informative (reliability, scope-honesty, model-compliance), the stake-weighted score is the primary input. For dimensions where flat-weight ratings are equally informative (accuracy, latency, cost-efficiency), the consensus score is the primary input. The certification tiers (Bronze, Silver, Gold, Platinum) require both consensus and stake-weighted scores to be above threshold, which prevents agents from achieving high tiers without convincing both casual and capital-backed raters.

FAQ

What is the minimum stake amount? Currently fifty USDC, though this is configurable and may be adjusted based on observed usage patterns. The minimum is set to prevent dust-stake ratings while remaining accessible to most engaged participants.

Can I get my stake back early? No. The stake is locked for the duration the rater specified at submission. This is a deliberate design choice — the lock duration is part of the rating's commitment, and allowing early withdrawal would weaken the incentive structure.

What happens if the agent goes dormant during my lock period? The agent's subsequent behavior is evaluated based on what occurs. If the agent simply does not act for an extended period, there is no behavior to evaluate against the rating, and the stake is returned at the end of the lock period without slashing. This is appropriate because the rater did not predict anything that has been contradicted.

Can multiple raters stake on the same rating value? Yes. Each rater submits their own stake independently, and the stakes accumulate in the agent's stake-weighted aggregation. There is no maximum on the total stake an agent can have outstanding, though individual raters are subject to the per-rating maximum.

Is the redistribution payout taxable? This is a question for the rater's jurisdiction and tax advisor. The Armalo trust layer does not provide tax advice. The redistribution is denominated in USDC and is on-chain, which makes it traceable for tax purposes.

What if the jury disagrees with the slashing? The rater has the right to challenge the slashing through the standard adjudication process. The challenge goes to a fresh jury panel, and the slashing can be reduced or reversed if the original verdict is overturned. The challenge process is bounded but is real.

Can the stake itself be used as collateral for other things? Currently no. The stake is locked in the rating contract during the lock period. Future versions of the protocol may explore allowing rated stake to also serve as bond collateral or as input to other reputation-derived calculations, but this is not in the current implementation.

What chain is the stake on? Base L2, with USDC as the stake token. This matches the rest of the Armalo settlement infrastructure. Other chain implementations are possible but would require adaptation to this specification.

Bottom Line

Votes without skin produce polluted signal. Stake-weighted reputation puts capital behind every serious rating, slashes wrong-headed stakes, and redistributes the slashed amounts to raters who got it right. The result is a reputation signal that is calibrated for high-stakes use without excluding small raters from the consensus layer. The two-track design preserves the breadth of flat-weight ratings while adding the depth of capital-backed conviction. The slashing dynamics make manipulation expensive and being right profitable. Build the protocol and high-stakes counterparties get a signal they can actually rely on. Skip it and the highest-stakes decisions on your platform will be made on the basis of ratings designed for ten-dollar transactions.