The Trust Score SLA: What Operators Should Promise Buyers And What They Cannot
Bonded trust score SLAs are bold and possible. Here is what to promise, what to refuse, and the liability math that keeps the operator solvent.
Continue the reading path
Topic hub
Runtime GovernanceThis page is routed through Armalo's metadata-defined runtime governance hub rather than a loose category bucket.
TL;DR
Buyers want service-level guarantees on agent behavior. Operators have learned to promise uptime and throughput; trust score is the new dimension. A trust score SLA says: this agent will hold its composite score above some threshold during the contract window, or the operator pays. This is enforceable because the trust oracle is the same number both parties read. The mistake is promising on a single composite number alone. The right SLA is multi-clause: a floor on composite score, a floor on the highest-risk dimension for the buyer, an exclusion list for events the operator does not control, and a payout schedule bonded against an on-chain reserve. This post gives the template, the liability math, and the seven failure modes that destroy operators who skip the math.
Intro: the day the operator wrote a check they could not cover
A mid-sized operator running thirty-two customer-service agents for a pair of fintech buyers signed a clause that read, in plain English, "Agent A will maintain a composite trust score of at least 880 during the contract term. Each calendar day the score drops below 880 entitles Buyer to a credit equal to one day of fees." The clause was added at the last minute, after a procurement officer asked for "some kind of trust commitment." The operator's revenue engineer accepted because the agent had been running at 906 for the prior quarter. There was room. The contract closed on a Friday and the operator's leadership celebrated a fourteen-month deal worth roughly a hundred and twenty thousand dollars annually with a buyer they had been chasing for two quarters.
In week six, an upstream model provider quietly retrained the underlying policy. The composite score dropped to 871 within forty-eight hours, recovered to 879, dropped again to 866 the following week when a red-team battery hit the agent harder than expected. The SLA fired on twelve different days across two months. Credits owed exceeded the gross margin of the contract. The operator paid because the contract said pay. Then the operator went to the buyer and asked for the clause to be amended. The buyer said no, and the buyer was right to say no, because the buyer had paid for a guarantee and the operator had agreed to it without modeling the failure modes.
The failure was not the score drop. The failure was the SLA design. The clause priced a number the operator did not control end to end. It treated composite score as if it were uptime, which it is not. Composite score is a moving estimate produced by a multi-LLM jury reading a stream of behavioral evidence; it can move because the agent changed, because the world changed, because the jury changed, because the evaluation surface changed, or because adversarial probing intensified. A bonded promise needs to enumerate which of those movements the operator is responsible for, which the buyer absorbs, and which are split. The clause in question enumerated none of these. It treated the score as if it were a thermostat under the operator's hand.
It is worth dwelling on why this failure mode is the modal one. Procurement officers have learned to ask for SLAs because uptime SLAs taught the industry that bonded promises focus operator attention. Engineers have learned to write SLAs because uptime SLAs are a solved genre with a known template. Trust score is a new measurement, but the procurement instinct and the engineering template both transferred over from uptime, and the transfer is wrong. Uptime is a binary the operator controls almost entirely. Trust score is a continuous estimate the operator influences but does not control. The template that worked for the binary fails badly for the continuous estimate.
This post is the playbook for writing trust score SLAs that operators can actually keep. It assumes the reader runs five to fifty agents in production, has at least one buyer asking for a trust commitment, and would like to sign a bond rather than a wish. It contains a template, a liability calculator, an exclusion list, a dispute pathway, and the seven failure modes that have already cost operators real money. Every section is runbook-grade: copy the clauses, plug in your numbers, run the math before you sign.
Why a trust score SLA is now possible
For most of the history of services, operators promised uptime because uptime was a number both parties could compute from the same logs. Trust was promised in narrative paragraphs and lawsuit clauses, never as a number. The reason was simple: there was no shared definition of trust the parties could read in the same way at the same time, so any bonded promise dissolved into argument. The operator would point at one set of evidence; the buyer would point at another; the courts had no neutral measurement to anchor either side. Bonded trust promises were structurally impossible.
The trust oracle changes this. A composite score, computed from a multi-LLM jury that trims outliers, published to a public read endpoint with cryptographic integrity, gives both parties one number they can quote without ambiguity. The operator can monitor it in real time. The buyer can audit it after the fact. A neutral third party can replay the inputs that produced it. The number is not perfect; no behavioral measure is. But it is shared, and shared is the precondition for bonded promises. Without a shared measurement there is no SLA. With a shared measurement the SLA conversation becomes a conversation about thresholds, exclusions, and payouts, all of which are tractable.
The second precondition is on-chain settlement. A bonded SLA needs a place to put collateral that neither party controls unilaterally. USDC on Base L2 with an escrow contract solves this: the operator posts collateral, the SLA defines the trigger conditions, and the contract pays out automatically when the trust oracle reports a violation. There is no "will the operator actually send the credit" friction. The credit is already locked. This matters more than it sounds. Off-chain SLAs have always had a credit-collection problem; the larger the credit, the more the operator delays, disputes, or simply absorbs through future invoice negotiations. On-chain settlement collapses that delay to zero and forces the operator to internalize the cost in the moment it is incurred.
The third precondition is dispute infrastructure. Bonded promises must allow either party to contest a payout. The composite score has structured ancestry: every change traces to specific evaluation events, jury verdicts, and behavioral observations. A dispute pathway lets the parties query that ancestry, identify whether a score drop was caused by an excluded event, and adjust the payout. Without dispute infrastructure, every score drop becomes a lawsuit, and bonded SLAs collapse under legal cost. The dispute pathway is the safety valve that makes the bonded promise survivable for both sides.
The fourth precondition, often unstated, is operator literacy. The operator must understand the score well enough to model its movement. An operator who treats the score as a black box cannot price the bond, cannot set an honest floor, and cannot defend a dispute. Operator literacy is the missing piece in many failed SLAs; the technical infrastructure exists but the operator has not invested the hours to understand the dimensions, the jury, and the score's response surface to operational changes. The remedy is unglamorous: spend three hours with the score's audit trail, replay the last quarter's score history, and build a mental model of why the score moves. Without that mental model, signing a bonded SLA is signing a contract on a number the operator cannot reason about.
When these four preconditions exist, the operator has the raw material to write a credible SLA. What remains is the discipline to write one that can be honored. That discipline is the rest of this post.
What to promise: the four-clause structure that holds up
A workable trust score SLA has four clauses, not one. The single-number clause that destroyed the operator in the intro is the most common failure mode and the easiest to refuse.
Clause one is the composite floor. This is the headline number: the agent's composite score will not fall below threshold T for more than D days in any rolling thirty-day window. T is set well above the agent's known operating floor; D allows for normal jitter without firing the SLA. A floor of 850 with a D of three days is reasonable for a Silver-tier customer service agent that historically runs at 890. A floor of 920 with a D of one day is appropriate for a Platinum-tier transaction-handling agent. The floor is always below the steady-state score; it is a floor, not a target. Operators who set the floor at the steady-state are guaranteeing the SLA will fire regularly, because the score is a continuous estimate that always wanders. The right floor is two to three sigma below steady-state, where sigma is computed from the trailing ninety days of score history and adjusted upward if the agent has been running through a stable period that is unlikely to continue.
Clause two is the dimension-specific floor. The composite score is a weighted blend of twelve dimensions: accuracy, self-audit, reliability, safety, security, bond, latency, scope honesty, cost efficiency, model compliance, runtime compliance, and harness stability. A buyer who cares about safety should not be appeased by a high composite that hides a safety dip. The dimension floor names the one or two dimensions the buyer most depends on and sets a separate floor on each. For a fintech buyer the dimension floor often locks scope honesty above 0.95 and security above 0.97; for a healthcare buyer it locks safety above 0.99. The dimension floor reduces the operator's headline guarantee but eliminates the entire class of disputes where the composite looks fine and the buyer is hurt anyway. The buyer should pick at most two dimension floors; more than two reintroduces the headline-number failure mode by another door, because every additional dimension floor multiplies the probability of an SLA fire on a given day, and three or more independent floors typically double the expected payout without giving the buyer meaningful additional protection.
Clause three is the exclusion list. This clause names events the operator is not responsible for: model provider retraining without notice, infrastructure outages at named providers, adversarial probing campaigns that exceed a stated rate, regulatory changes that force scope contraction, and force majeure. Exclusions trigger a tolling provision: the SLA window pauses while the excluded event is in effect. Without exclusions the operator is writing insurance against the entire universe; with carefully named exclusions the operator writes insurance only against operator-controlled risk. The exclusion list should be specific. "Acts of God" is too vague to operationalize. "A named upstream model provider issues a policy update with less than sixty days notice that produces a measurable drop in the composite score traceable to specific evaluation runs after the update timestamp" is operational. Specificity is what separates an exclusion that survives a dispute from an exclusion that collapses under one.
Clause four is the payout schedule. This is the bond. Credits are not paid as a percentage of fees "if and when invoiced." Credits are paid from a posted bond, on chain, automatically, with the SLA definition encoded in the escrow contract. The bond size is a function of contract size, score volatility, and the operator's risk appetite. A typical bond is two to four months of fees for a Silver-tier contract, or one month of fees for Platinum where the operator has high confidence in the score. The bond is the operator's cap on liability; the SLA cannot demand more than the bond. The bond also signals operator commitment: a buyer who sees a real bond posted in a verifiable address knows the operator is putting capital behind the promise rather than relying on goodwill. That signal is often worth more commercially than the bond itself; bonded operators win contracts that unbonded operators lose, even when the unbonded promise is nominally larger.
Each clause is necessary. A composite-only clause invites the dimension-specific failure mode. A composite plus dimension clause without exclusions makes the operator the insurer of the world. A composite plus dimension plus exclusions without a bond is a contract whose enforcement depends on the operator's solvency and good faith, which is what off-chain SLAs always relied on and what bonded SLAs improve.
What never to promise
There are seven promises operators have made and regretted. Each is named below with the reason it fails and the corrective the operator should adopt instead.
First, never promise a perfect floor. The composite score is a continuous estimate; it always moves. A floor at the steady-state score guarantees firing. The floor must sit below the steady-state by enough margin to absorb normal jitter, which for most agents is twenty to forty points. Operators who anchor on the steady-state are confusing the score with the agent; the score is a measurement of the agent, and measurements have noise. The corrective: set the floor at steady-state minus three sigma, where sigma comes from the trailing ninety days of score history. If sigma is unknown because the agent is too new, the corrective is to refuse the SLA until calibration data exists.
Second, never promise a future score level. "This agent will be at 950 by end of quarter" is a forecast, not a commitment. Forecasts cannot be bonded because the operator does not control the inputs. Floors can be bonded because the operator controls the actions that move the score. Operators sometimes accept forecast clauses because buyers want them; the corrective is to convert the forecast into a floor with a calibration period. "By end of quarter, the agent will operate at a steady-state score consistent with a floor of 920 over rolling thirty-day windows" is a bondable clause. "The agent will be at 950" is a wish.
Third, never promise a single-day window. Single-day floors fire on jitter alone. The SLA should always quote a rolling window: "no more than D days below floor in any rolling thirty days." The window absorbs jitter without absorbing genuine failures. The corrective: make D and W explicit and set them so that the probability of an SLA fire in any window is below a threshold the operator can afford. Most operators settle on D equal to three and W equal to thirty, which together absorb the jitter of well-calibrated agents and fire reliably for genuine degradation.
Fourth, never promise a dimension the operator does not control. If the agent runs on top of a third-party model and the operator has no influence over latency, do not bond the latency dimension. The operator's bond should only price dimensions the operator can actually move via prompt engineering, harness changes, or runtime improvements. The corrective: before agreeing to a dimension floor, run the audit trail to confirm the dimension responds to operator interventions. If the operator cannot point to a historical case where the dimension moved in response to an operator action, do not bond it.
Fifth, never promise a score on a market the operator has not run a calibration period in. New deployments need a calibration window, typically four weeks, during which the SLA is suspended. Promising a floor before the calibration period means promising a number neither party knows. The corrective: insert a calibration clause that explicitly suspends the SLA for the first four weeks, during which both parties accumulate score history and either party can renegotiate the floor based on observed behavior.
Sixth, never promise a bond larger than the operator can lose. The bond should be a real fraction of margin, not an aspirational number. An operator who bonds half a year of fees against an SLA they cannot reliably hold is an operator one bad month away from default. The corrective: cap the bond at three months of contract fees, or at five times expected annual payout, whichever is smaller. If the buyer demands a larger bond, the operator must price the contract higher to cover the additional capital cost or refuse the contract.
Seventh, never promise without an exclusion list. Promises without exclusions are promises against the universe. The exclusion list is not the operator finding excuses; it is the operator naming the boundary of operator-controlled risk. The corrective: refuse to sign any bonded clause without the exclusion list attached. If the buyer demands an unconditional clause, walk away or convert to an unbonded credit-on-invoice term that limits exposure to a single billing cycle.
These seven failure modes are the difference between an SLA that strengthens the operator's commercial position and an SLA that ends the operator's career. Memorize them.
The liability math
Liability math is the piece operators most often skip. The math is not complicated; it is just unfamiliar. Operators who can estimate cloud spend to within ten percent often cannot estimate SLA payout to within an order of magnitude, because they have never built the model. The model below is sufficient for almost every bonded trust score SLA in current commercial use.
Start with three numbers: the contract value V (annualized), the floor T, and the steady-state score S. The probability of any single day below T is the area of the score distribution below T. For most agents the distribution is approximately normal with standard deviation sigma, where sigma is computed from the trailing ninety days of score history. The probability per day is the cumulative normal of (T minus S) over sigma. Call this P. This is the per-day failure probability.
If the SLA allows D days below floor in a rolling thirty-day window, the probability of an SLA fire in any given thirty-day window is the binomial probability of more than D successes in thirty trials with success probability P. Call this Q. This is the per-window failure probability. Note that Q is sensitive to D in a way that operators often miss: increasing D from one to three can reduce Q by an order of magnitude or more, because the binomial tail is steep. Operators who push hard on D in negotiation often lose more than operators who push hard on T.
The expected annual payout is twelve months times Q times the per-fire credit. If the credit is one day of fees, the per-fire credit is V over 365. So the expected annual cost of the SLA is (twelve times Q times V) divided by 365. This is the expected cost; it understates the variance. Operators should also compute the ninety-fifth percentile cost, which is the cost the operator will exceed in one year out of twenty. The ninety-fifth percentile is what determines whether the bond is sized correctly; the expected value determines whether the contract is profitable.
An example: V is one hundred twenty thousand dollars, T is 850, S is 905, sigma is fifteen, D is three days. P is the cumulative normal of (850 minus 905) over 15, which is approximately 0.0001. Q is essentially zero; the binomial probability of more than three days below T in thirty days is negligible. The SLA costs almost nothing in expectation. This is the case where bonding is essentially free and the operator should always offer it because the commercial signal is large and the cost is negligible.
Now rerun with sigma equal to twenty-five. P jumps to 0.014. Q rises to about 0.001. Annual expected cost rises to about four dollars. Still trivial; the bond is mostly working capital.
Now rerun with sigma equal to forty (a noisy agent under heavy adversarial probing). P jumps to 0.084. Q rises to about 0.16. Annual expected cost rises to about six thousand dollars, or five percent of the contract. The ninety-fifth percentile is materially higher, perhaps fifteen thousand. At this level the SLA is real insurance with real cost; the operator must price the contract to cover it or refuse the floor.
Now rerun with the floor moved from 850 to 875. With sigma equal to twenty-five, P jumps from 0.014 to 0.115. Q jumps from 0.001 to 0.39. Annual expected cost rises from four dollars to nearly fifteen thousand. A twenty-five point shift in the floor produced a four-thousand-fold increase in expected payout. Operators who treat the floor as a negotiating round number rather than a probabilistic input frequently get destroyed by exactly this kind of small-looking change.
The math reveals where the SLA is cheap and where it is dangerous. For low-jitter agents on conservative floors, bonded SLAs cost almost nothing in expectation and provide a powerful commercial signal. For noisy agents on tight floors, bonded SLAs become real insurance with real expected cost. The operator must price the contract accordingly or refuse the floor.
Operators should always pad the bond above expected payout. A bond of three times annual expected payout absorbs ordinary tail events. A bond of ten times absorbs rare tail events at the cost of capital efficiency. The right multiple depends on capital cost; most operators settle around five times. The bond should also have a floor: even when expected payout is small, the bond should be at least one month of fees so that the buyer sees a meaningful commitment. The cheap-bond temptation is to size the bond at expected payout times one and a half; the operator who does this will be reasonably solvent on average and bankrupt on a bad year.
Artifact: the Trust SLA Template And Liability Math worksheet
The artifact attached to this post is a one-page worksheet operators can use to draft and price a trust score SLA in a single sitting. It has six sections, each designed to fit on a quarter of the page and to be filled in with concrete numbers rather than narrative.
Section one is contract identity: buyer name, agent identifier, contract start date, contract value, contract duration. This section is paperwork; it exists to make every bonded SLA traceable to a specific commercial relationship. Operators with more than a handful of bonded contracts should also assign each SLA an internal identifier and link it to the underlying agent record so that score movements can be cross-referenced to active bonds in real time.
Section two is the composite floor specification: floor T, days-below-floor allowance D, rolling window length W. The recommended starting point is T equal to steady-state minus thirty points, D equal to three days, W equal to thirty days. These numbers are conservative and almost always within the operator's control. The worksheet has a field for the buyer's preferred T and an automatic calculation of the resulting expected payout, so that the negotiator can show the buyer in real time the cost of moving T upward by ten points.
Section three is the dimension floor specification: which dimension or dimensions the buyer cares most about, the floor on each, and the same D and W parameters. A buyer should pick at most two dimension floors; more than two reintroduces the headline-number failure mode by another door. The worksheet has a dropdown of all twelve dimensions and a field for the dimension-specific floor with a calculation of the marginal payout cost of adding each dimension. This makes the conversation explicit: the buyer sees that adding a third dimension floor doubles the expected payout, and the buyer can decide whether the protection is worth the bond.
Section four is the exclusion list. The starter exclusions are: model provider retraining without sixty days notice, named upstream infrastructure outages, adversarial probing exceeding a stated rate, regulatory scope changes, force majeure. The operator should add any agent-specific exclusions and should never accept a contract without an exclusion list. Each exclusion has a one-sentence operational definition: how it is detected, what evidence triggers tolling, and how the SLA resumes after the event. The worksheet enforces this by reserving space for the operational definition next to each exclusion line; lines without operational definitions are not exclusions, they are arguments.
Section five is the liability calculation. Plug in V, T, S, sigma, D, W. The worksheet computes P, Q, expected annual payout, ninety-fifth percentile annual payout, and recommended bond size. The bond is sized at five times expected annual payout, with a floor of one month of fees. If the ninety-fifth percentile exceeds the bond, the worksheet flags the contract as underbonded; the operator must either increase the bond, raise the floor, or refuse the contract.
Section six is the dispute pathway. Both parties name a primary contact, a secondary contact, and the timeline for raising a dispute (typically seven days from the SLA fire) and the timeline for resolution (typically thirty days). The dispute pathway references the trust oracle's audit trail as the canonical evidence source. The worksheet also reserves a field for the agreed arbitration mechanism in the event the parties cannot resolve the dispute internally; most operators use commercial arbitration with a single neutral arbitrator chosen from a pre-agreed panel.
A filled worksheet should fit on one page. Operators who cannot fit their SLA on one page have either not constrained the contract or have not constrained themselves. Either way, the contract is too dangerous to sign. Once the worksheet is filled, the operator should sign nothing for at least twenty-four hours, during which someone other than the negotiator runs the math independently. The discipline of an independent recheck has caught more bad SLAs than any single feature of the worksheet itself.
Bond posting and on-chain mechanics
The bond is the part of the SLA that exists outside the contract document. Posting the bond requires a deliberate set of operational steps that should be standardized across every bonded contract the operator signs.
The operator opens an escrow contract on Base L2 sized to the agreed bond. The contract is parameterized with the agent's identifier, the floor T, the days-below-floor allowance D, the rolling window W, the credit per fire, and the exclusion list reference. The contract reads from the trust oracle on a defined cadence, typically daily, and updates an internal counter of days below floor in the rolling window. When the counter exceeds D, the contract automatically transfers the credit from the bond to the buyer's address. The contract logs every read, every counter update, and every payout, producing an on-chain audit trail that mirrors the off-chain dispute pathway.
The operator funds the bond from operating cash, not from a credit line, because credit lines can be revoked. The bond should appear on the operator's balance sheet as restricted cash with a corresponding contingent liability. The operator's controller should review bonded contracts quarterly and confirm that total restricted cash equals total bonded liability. Operators who treat the bond as a contract artifact without surfacing it in financials run the risk of double-spending the cash, which produces the same default failure mode as missing the SLA fire.
The contract should support multiple administrative actions: top-up (the operator adds cash mid-term, often when the bond is partially consumed by a fire), reduction (the operator reclaims cash at term end after all disputes are resolved), and emergency pause (both parties can jointly pause the contract during an extended exclusion event, returning the cash to a held state until normal operation resumes). Each administrative action requires a multi-signature transaction signed by both parties; this prevents either party from unilaterally draining the bond.
The bond contract should never accept upgrades. Upgradeable contracts introduce a trust dimension the SLA does not price; the buyer wants to know that the contract logic on day one is the contract logic on day three hundred sixty-five. Operators who use upgradeable contracts are introducing a hidden party (whoever can upgrade) into the trust relationship. Use immutable contracts and deploy a new contract for each significant logic change.
Dispute pathway in operational detail
When an SLA fires and the operator believes the fire was caused by an excluded event, the operator initiates a dispute. The dispute pathway is the most legally sensitive part of the SLA and benefits from being defined with paranoid precision.
Within twenty-four hours of the SLA fire, the operator notifies the buyer in writing through a pre-agreed channel (email plus a logged message in the contract management system). The notification names the firing event, the alleged exclusion, and the evidence the operator intends to produce. Within seventy-two hours the operator delivers the evidence: the trust oracle audit trail for the fire window, the upstream provider notification or outage report supporting the exclusion claim, and the internal incident report documenting the operator's response.
The buyer has seven days to accept the exclusion or contest it. If the buyer accepts, the contract is rolled back automatically through a multi-signature reversal transaction; the bond is replenished and the counter is reset. If the buyer contests, the parties enter formal dispute resolution.
Formal dispute resolution begins with a joint review of the evidence. Both parties query the trust oracle for the disputed window and walk through the audit trail event by event. The audit trail shows which evaluations contributed to the score drop, which jury verdicts produced those evaluations, and which behavioral observations triggered those evaluations. The parties identify whether the chain traces to an excluded event or to operator-controlled risk.
Most disputes resolve in the joint review. The trust oracle's audit trail is detailed enough that the answer is usually visible to both parties simultaneously, and the cost of disagreement after seeing the same evidence is low. Disputes that do not resolve in joint review proceed to arbitration. The arbitrator examines the same audit trail, applies the contract's exclusion definitions, and issues a binding decision. The arbitrator's decision is encoded in a multi-signature transaction that adjusts the bond either by replenishing it (exclusion upheld) or by confirming the payout (exclusion rejected).
The operator should track dispute outcomes over time. Operators whose disputes are upheld at high rates may be writing exclusions that are too broad and should narrow them. Operators whose disputes are rejected at high rates may be writing exclusions that are not survivable in arbitration and should rewrite them with operational definitions that match arbitrator expectations. The dispute outcome data is itself a feedback loop into the SLA template; operators who run more than ten bonded contracts should rebuild the template annually based on what their dispute history teaches.
Counter-argument: bonded SLAs make operators risk-averse
The strongest argument against bonded trust score SLAs is that they make operators conservative. If an operator faces a bond every time the agent dips below floor, the operator will set artificially low floors, refuse contracts where the buyer wants ambitious floors, and stop pushing the agent's capability envelope. The result, the argument goes, is an industry of cautious operators running mediocre agents, while bolder competitors who refuse to bond capture the upside.
This argument is partially right and entirely correctable. Operators who bond aggressively and then run conservative agents are leaving margin on the table; the right response is to bond appropriately for the agent's actual risk profile, not to refuse to bond at all. Bonds priced at five times expected payout for low-jitter agents cost almost nothing and signal extraordinary commitment. Bonds priced at the same multiple for high-jitter agents are expensive precisely because the operator should be cautious about deploying high-jitter agents into customer-facing roles.
The argument also assumes the alternative is no SLA at all. The alternative in practice is procurement officers pushing operators into open-ended trust language with no liability cap, no exclusion list, and no defined trigger. That alternative is far more dangerous than a bonded SLA written with discipline. The bonded SLA caps liability at the bond; the open-ended language exposes the operator to whatever the buyer's lawyers can construct.
The deepest version of the counter-argument is that trust scores themselves are too noisy to bond. This is empirically false for most production agents whose composite scores have settled. It is true for agents in their first calibration period, which is precisely why the SLA template suspends during calibration. It is true for agents undergoing significant policy changes, which is why the SLA should include a change-management exclusion. The noise objection points to which agents should not yet be bonded, not to the impossibility of bonding.
A related counter-argument is that bonded SLAs concentrate liability in a small number of well-capitalized operators, creating a barrier to entry. There is a kernel of truth here: bonded promises favor operators with capital. The mitigation is bond pooling and bonded reinsurance, both of which are emerging in the agent economy and which let smaller operators access bonded structures without posting the full bond from operating cash. The pooling structure is the same insurance pattern that has worked for centuries in maritime and reinsurance markets; there is no reason it cannot work for agent operators.
What Armalo does
Armalo provides the rails that make bonded trust score SLAs operable. The composite score is computed by a multi-LLM jury that trims the top and bottom twenty percent of judgments, producing a stable estimate that both parties can quote without dispute. The trust oracle exposes the score on a public read endpoint with a structured audit trail; both parties read the same number from the same place. Pacts capture the agent's behavioral commitments and become the substrate for dimension-specific floors. Escrow on Base L2 with USDC settles bonded payouts automatically when the SLA's trigger conditions are met. The dispute mechanism allows either party to query the score's ancestry and contest a payout that traces to an excluded event. None of this requires the operator to write custom infrastructure; it is the platform.
FAQ
Q: What is the smallest contract worth bonding? A: Below about ten thousand dollars annually, the cost of negotiating the SLA exceeds the commercial value. Smaller contracts can use the unbonded version of the same template, with the bond clause replaced by a credit-on-invoice term. Operators with many small contracts can also pool them into a single bonded structure that covers a portfolio rather than each contract individually.
Q: How often should the bond size be revisited? A: Quarterly. Bond size is a function of agent volatility; volatility changes as the agent and the world change. A standing process to recompute sigma and rerun the liability math keeps bonds calibrated. Operators who set bonds once and never revisit them frequently discover that the agent has drifted into a higher-volatility regime and the bond is underwater long before any single fire.
Q: Can the bond be shared across multiple SLAs? A: Yes, with care. A pooled bond covers all SLAs against an operator and reduces capital requirements, but the SLA contracts must specify the pooling and the order of claims. Most operators bond per contract until they have enough contracts to make pooling actuarially sensible, typically around twenty active bonded contracts.
Q: What happens when an exclusion event occurs? A: The SLA tolls. The clock pauses. When the event ends and the agent resumes normal operation, the clock resumes. The exclusion event is logged in the dispute pathway so that any future dispute has clean evidence. The operator should publish a brief incident summary at the time of the tolling so that the buyer can see in real time why the contract clock paused.
Q: How is the calibration period defined? A: Four weeks of post-deployment operation during which the SLA is suspended and both parties accumulate score history. At the end of the calibration period the parties review S and sigma, recompute the bond, and activate the SLA. If the calibration period reveals that the agent's actual sigma is materially higher than the operator estimated, the parties may renegotiate the floor before activation.
Q: What if the trust oracle itself has an outage? A: The SLA tolls. The operator and buyer cannot enforce a bonded promise on a measurement neither party can read. The contract should specify the maximum tolerable oracle outage (typically four hours) before tolling kicks in, and should specify a fallback evidence source for the duration of the outage.
Q: Can the buyer demand audit access to the underlying agent? A: Yes, and the operator should expect this. The audit access should be defined in the contract: which logs, which retention period, which channel. The operator should never ship raw model weights or proprietary code in the audit access; the trust oracle's audit trail is sufficient for almost every legitimate buyer audit need.
Bottom line
Trust score SLAs are now possible because the trust oracle gives both parties one number, on-chain settlement gives both parties one place to put the bond, and the dispute mechanism gives both parties one pathway to contest a payout. They are dangerous when written casually and powerful when written with discipline. The discipline is four clauses: composite floor, dimension floor, exclusion list, bonded payout. The math is binomial probability over the agent's score distribution; the bond is five times expected annual payout. Operators who internalize the template and the math can offer commercial commitments their competitors cannot match. Operators who skip the math write checks they cannot cover. The template is a one-page worksheet; the math is a spreadsheet; both are cheap. Run them before you sign.
Continue exploring
Explore more articlesComments
Loading comments…