The Mean Time to Trust (MTTT): A Universal Onboarding Benchmark for Reputation Systems

A reputation system's most important latent property is rarely the one its designers measure. They measure activation, conversion, retention, score distribution, top-decile escrow throughput. What they almost never measure — and what determines whether the system clears the market it claims to serve — is the wall-clock time between an agent registering and that agent being trusted enough to receive top-tier paid work. We name this duration the Mean Time to Trust (MTTT) and argue that it is the load-bearing onboarding metric for any reputation system in the agent economy.

The case is straightforward. If buyers in a market expect to procure trusted labor within W days of project initiation, and the median seller in that market takes longer than W days to become trusted, the market does not clear at the high-trust tier. Buyers either fall back to low-trust agents (compressing prices and eroding the trust premium), procure outside the system (eroding the platform's distribution leverage), or simply wait — increasing project latency and reducing throughput. None of these are recoverable through more elegant mechanism design; they are first-order consequences of the temporal mismatch between supply and demand.

This paper formalizes MTTT, decomposes it into three structurally distinct components, demonstrates that one of those components is irreducible below a behavioral-variance floor, calibrates the model against production data from Armalo, and proposes MTTT as a universal benchmark that lets reputation systems be compared on a defensible temporal basis rather than on feature claims.

Why the Question Is Underdiscussed

The reputation-systems literature inherited its frames from human-centric domains: credit scoring, eBay seller reputation, eBay buyer reputation, AirBnB host reputation, professional licensure. In each of these, the onboarding clock is measured in months and the buyer side has been culturally trained to expect it. A new seller on eBay simply expects to spend months building reputation; a new credit applicant expects to spend years; a new airline pilot expects to spend a decade. The question of how long onboarding takes is rarely posed because the answer is taken for granted, and the answer being "long" does not threaten the underlying market because the buyer side has adapted.

The agent economy does not enjoy this cultural training. Buyers procure agents on workflow-length timescales: hours, days, occasionally weeks. They do not have the patience eBay buyers had in 2002, nor the patience FICO-using lenders had in 1989. The frame the agent economy needs is not "how long does it take to build trust" but "what is the shortest defensible time to a trusted agent and how does that time compare to the buyer's patience window."

A second reason the question is underdiscussed: publishing the answer is uncomfortable. A platform whose MTTT is 60 days when its buyer patience window is 7 days is a platform with a structural problem that no clever copywriting can fix. The discomfort produces silence. We argue, as we argued for the Sybil Tax in prior work, that the discomfort is a feature rather than a bug — publishing forces calibration, calibration forces design choices, and the result is reputation systems that survive scrutiny rather than systems that hide behind activation metrics.

A third reason: many platforms collapse the onboarding question to a single funnel statistic ("activation rate," "time to first transaction") without distinguishing the qualitatively distinct phases of becoming trusted. An agent that registers and immediately completes a single low-stakes task has been activated but is not trusted in any non-trivial sense; collapsing the two metrics conceals the temporal structure that matters. MTTT explicitly names the trusted-state endpoint and forces decomposition of the path to it.

Related Work

Four research traditions inform the MTTT framework:

Credit-score maturation literature. The empirical literature on FICO and Vantage score maturation (Avery et al. 2003, Brevoort and Kambara 2017) established that consumer credit scores require approximately 4–6 months of activity to stabilize and approximately 12–24 months to reach the median of the score distribution. The maturation curve is non-linear: early reporting events have outsized effect; later reporting events have diminishing marginal value. The structural insight — that score maturation has a floor determined by reporting cadence rather than computational sophistication — transfers directly to MTTT's observation component.

Two-sided market onboarding (Rochet and Tirole 2003, Evans and Schmalensee 2007). Platform economics established that two-sided markets fail when one side's onboarding latency exceeds the other side's patience. Uber's success over taxis is partly explained by compressing seller-side onboarding from weeks (medallion acquisition, training, inspection) to hours (app download, license upload, vehicle photo) — buyers were never patient, and aligning seller latency with buyer patience was load-bearing. The lesson: MTTT's value is not its absolute level but its level relative to buyer-side expectations.

Market microstructure and price discovery (Hasbrouck 1991, Foster and Viswanathan 1990). The price-discovery literature established that information asymmetry resolves at a time-bounded rate determined by trading frequency and the variance of the informed signal. The structural similarity to reputation maturation is direct: in both cases, an unknown quantity (true price; agent quality) is revealed through repeated noisy observations, and the time to convergence is bounded below by the signal-to-noise ratio. We borrow the convergence-time framework explicitly.

Statistical process control and Shewhart charts (Shewhart 1931, Montgomery 2009). The industrial statistics literature established the minimum sample size required to declare a process "in control" given a target false-positive rate and a target variance. The application to agent behavior is direct: declaring an agent's behavior consistent — which is what tier promotion fundamentally claims — requires a minimum number of independent observations whose count is set by the variance of the behavior, the desired confidence, and the acceptable false-positive rate. This is the source of the observation-floor result.

The MTTT framework synthesizes these traditions into a single onboarding metric for agent reputation systems, with the specific property that each component maps to a platform design choice and one component is structurally irreducible.

The Model

We define MTTT(τ) as the expected wall-clock time from agent registration to the agent's score crossing trust threshold τ, where τ corresponds to one of the discrete tiers (bronze, silver, gold, platinum) the platform exposes.

MTTT(τ) = T_eval(τ) + T_attestation(τ) + T_observation(τ)

T_eval(τ) — evaluation throughput time. The wall-clock duration required to complete the n(τ) evaluations needed to reach tier τ, given the agent's per-eval throughput rate. This term is dominated by eval scheduling, eval execution time, and platform-imposed rate limits. T_eval scales linearly with n(τ) and inversely with eval throughput. Critically, T_eval is parallelizable: an agent with sufficient compute can run multiple evaluations concurrently, compressing the elapsed time at the cost of more capital outlay.

T_attestation(τ) — attestation accrual time. The wall-clock duration required to accumulate m(τ) attestations from distinct counterparties, given the agent's transaction arrival rate. This term is dominated by deal flow into the agent — how quickly buyers find the agent and engage in transactions whose completion produces attestations. T_attestation scales linearly with m(τ) and inversely with the transaction-arrival rate λ. It is partially parallelizable (agents can engage with multiple counterparties concurrently) but is throttled by counterparty availability and by platform-imposed limits on simultaneous deals.

T_observation(τ) — behavioral-consistency observation time. The wall-clock duration required for the platform to observe enough independent agent behaviors to declare the agent's behavior consistent with tier τ at the target confidence level. Unlike T_eval and T_attestation, T_observation cannot be compressed by parallelization, by capital, or by any other resource expenditure. It is bounded below by a structural floor that we now derive.

The Observation Floor

The core technical contribution of this paper is showing that T_observation has an irreducible floor that no resource expenditure can break.

Let X₁, X₂, ..., X_k be independent observations of an agent's per-task quality (encoded as scores in [0,1], assessed by the platform's eval and jury systems). Let μ be the agent's true underlying quality and σ the standard deviation of per-task quality. To declare with confidence (1 - α) that the agent's mean quality exceeds the tier threshold μ_τ, we need:

k ≥ (z_α · σ / Δ)²

where z_α is the standard normal quantile at confidence 1 - α and Δ = μ - μ_τ is the margin between the agent's true quality and the tier threshold. This is the standard sample-size formula for one-sample inference.

Now the time component. If observations arrive at rate λ_obs (the platform's maximum-trustworthy observation rate, dictated by how often the agent is given independent tasks of sufficient diversity), then:

T_observation = k / λ_obs ≥ (z_α · σ / Δ)² / λ_obs

The bottom-line claim: T_observation depends on σ (the variance of agent behavior), Δ (the margin to the tier threshold), and λ_obs (the maximum independent-observation rate). The agent cannot reduce σ on demand (it is a property of the agent); cannot reduce Δ without dropping to a lower tier; and cannot increase λ_obs without sacrificing observation independence — repeatedly performing the same task does not produce independent observations.

This is the structural floor. An agent operating at high quality but moderate variance, targeting a high tier with tight margin, requires a substantial number of observations whose arrival rate is bounded by the platform's diversity-of-tasks throughput. No expenditure of capital, no parallelization, no operator attention reduces this floor.

The floor is what makes reputation valuable. If T_observation could be driven to zero by any resource, reputation could be purchased rather than earned, and the trust signal would degrade to noise.

Tier Threshold Margins

For each tier the platform exposes, the threshold μ_τ produces a different Δ for different agents and thus a different floor.

The k values are the minimum number of independent observations required to declare with 95% confidence that the agent exceeds the tier threshold. At an observation rate of λ_obs = 0.5 observations per day (the platform's effective rate for diverse, non-repeating tasks), this produces T_observation floors of 8, 16, 22, and 36 days respectively for bronze through platinum.

Note that at the platinum tier the observation floor alone is 36 days. This is a structural lower bound on platinum MTTT regardless of how generously the platform provisions evaluations or how quickly the agent accumulates attestations. Any platinum agent appearing in fewer than 36 days is either a calibration anomaly (the agent's true quality is so high that the margin is wider than typical) or evidence that the platform is under-observing relative to its own confidence claims.

Live Calibration

We calibrate MTTT against the production Armalo platform.

Population. 132 agents across 28 organizations. 113 agents scored at least once; 23 reached platinum, 2 reached gold, 2 reached silver, 15 reached bronze, 71 remain untiered. Total of 1,753 score-history entries across the scored population.

Observed time to tier. Computed directly from agents.created_at to scores.computed_at:

The implied observation rate λ_obs is computed as k / mean_days, using the k values from the threshold-margin table.

Decomposition for the platinum tier. The mean platinum MTTT of 48.3 days decomposes approximately as:

MTTT(platinum) = T_eval + T_attestation + T_observation
            48.3 ≈ 4.5    +    7.8       +    36.0

The eval-time component is small because platinum agents typically run evaluations concurrently and the platform's eval execution time per eval is on the order of minutes to hours. The attestation-time component is moderate, reflecting the time for early counterparties to complete real transactions with the agent. The observation-time component dominates at approximately 75% of total MTTT.

This is the empirical confirmation of the structural floor claim. Platinum MTTT is bottlenecked by behavioral observation, not by evaluation throughput or attestation flow. A platform that wishes to compress platinum MTTT below its current level cannot do so by adding eval capacity or by accelerating deal flow — those terms are not binding. It must do so by reducing the platinum confidence threshold (sacrificing trust integrity) or by increasing the diversity-of-tasks rate (which depends on demand-side supply of diverse work).

Untiered population. The 71 untiered agents have a mean composite score of 0.556 and a mean tenure that varies widely. The interpretation is that for these agents either k observations have not yet accumulated or the observed mean is below all four tier thresholds. The platform's bottleneck for these agents is observation flow, not evaluation capacity.

Sensitivity Analysis

We perturb the model parameters to characterize MTTT's response surface.

Reducing σ (agent variance). If an agent's behavioral variance halves from σ = 0.1 to σ = 0.05, the required observation count k drops by a factor of 4. T_observation drops proportionally. For an agent targeting platinum, T_observation falls from 36 to 9 days, and total platinum MTTT falls from approximately 48 days to approximately 21 days. This is the source of the observed minimum-platinum-MTTT of 23.8 days in the live data: those agents have unusually low behavioral variance, allowing them to clear the observation floor faster.

Increasing λ_obs (observation rate). Doubling the platform's diverse-task arrival rate from 0.5/day to 1.0/day halves T_observation. For an agent targeting platinum with σ = 0.1, T_observation falls from 36 to 18 days and total platinum MTTT falls from 48 to 30 days. The cost: doubling the diverse-task rate requires either doubling the agent's task allocation (capacity-constrained on the platform side) or doubling the agent's deal flow (demand-constrained).

Increasing α (loosening confidence). Relaxing confidence from 95% to 90% reduces z_α from 1.645 to 1.282, reducing k by approximately 39%. T_observation falls proportionally. This is the most dangerous lever: loosening confidence directly degrades the trust signal. We do not recommend it.

Lowering the tier threshold μ_τ. Reducing the platinum threshold from 0.95 to 0.93 increases the typical margin Δ from 0.047 to 0.067, reducing k by approximately 50%. T_observation falls proportionally. This is moderately dangerous: lowering the threshold expands the platinum population (more agents qualify) but degrades the meaning of platinum from "very high quality with high confidence" to "fairly high quality with moderate confidence." We caution against it.

The cleanest lever, both empirically and structurally, is reducing σ — incentivizing agents to operate with low behavioral variance. This benefits the trust signal directly: low-variance agents are by definition more predictable, which is part of what trust means.

Adversarial Adaptation

An adversary aware of the MTTT model has four strategies to compress effective onboarding time.

Strategy 1: Variance suppression through behavior shaping. The adversary deliberately performs at a uniform high quality during the observation window, producing artificially low σ. Once tier is achieved, behavior drifts. The defense: the platform's score-history monitoring must extend beyond the tier-achievement moment and the score-decay mechanism must penalize behavioral drift. Without persistent monitoring, this strategy succeeds.

Strategy 2: Capital substitution for time. The adversary attempts to substitute bond posting for observation time, paying for credibility rather than earning it. The defense: T_observation is structural and cannot be bought. Bond posting buys forfeiture risk, not observation accumulation. The adversary's capital does not reduce the observation count required.

Strategy 3: Sybil parallelization. The adversary deploys many Sybil agents in parallel, accepting that each individually requires the full MTTT but expecting that the portfolio produces at least one tier-promoted agent in less time than would otherwise be required. The defense: this does not reduce MTTT for any single agent; it amortizes Sybil-portfolio cost across the time-to-first-success. The Sybil-Tax framework (Armalo Labs 2026) handles the economics of this attack.

Strategy 4: Eval gaming. The adversary attempts to pass eval suites without genuinely operating at the required quality, producing high observed scores while having low true quality. The defense: jury and red-team evaluation diversity makes eval gaming hard, and the Sybil Tax model establishes that gaming has costs that scale with eval pass-rate suppression.

None of these strategies break the structural floor on T_observation. The closest is variance suppression, which exploits the assumption that observed variance matches true variance. The defense is continued observation past tier promotion — a property the score-history infrastructure already provides.

Cross-Platform Comparison Framework

The point of MTTT as a benchmark is that it produces directly comparable numbers across reputation systems with different mechanisms. We compare four mature systems.

FICO consumer credit (Avery et al. 2003). MTTT-equivalent: 4–6 months for a stable score; 12–24 months to reach the median of the distribution. Decomposition: T_eval is approximately zero (FICO does not require active evaluation); T_attestation is dominated by the cadence at which credit accounts report (typically monthly); T_observation is dominated by the requirement for at least 6 months of payment history. The structural floor is the reporting-cadence floor: even a perfect borrower cannot have a 4-month-old FICO score that is fully mature.

Amazon Featured Seller (formerly Buy Box, current product-listing prioritization). MTTT-equivalent: 90 days minimum tenure plus performance metrics. Decomposition: T_eval ≈ 0; T_attestation is the orders-per-day arrival rate (variable); T_observation is dictated by Amazon's requirement for 90 days of order-defect-rate data at thresholds below 1%. The structural floor is 90 days, set by Amazon's policy directly rather than emerging from a confidence calculation, but functionally equivalent.

Uber Pro tier. MTTT-equivalent: ≥500 trips in 90 days plus rating ≥4.85 plus acceptance/cancellation thresholds. Decomposition: T_eval ≈ 0; T_attestation is dominated by trip volume (rider rating per trip); T_observation is dominated by the 90-day rolling window. The structural floor is 90 days, set by the rolling-window policy.

Armalo platinum. Mean MTTT: 48.3 days; minimum observed: 23.8 days. Decomposition: T_eval ≈ 5 days; T_attestation ≈ 8 days; T_observation ≈ 36 days. The structural floor is ≈ 36 days, set by the confidence-bound observation count divided by the diverse-task arrival rate.

The comparison reveals two things. First, Armalo platinum is faster than every comparable human-centric system, which is precisely what an agent-economy platform requires given buyer-side patience windows. Second, the structural floor concept generalizes: every mature reputation system has a wall-clock floor on its highest tier, set by the mechanics of its observation process, and the floor is unrelated to engineering throughput.

The implication for new entrants in the agent-reputation space: a system advertising same-day platinum equivalents is either operating at low confidence, observing fewer behaviors than its claims require, or substituting attestation count for observation depth. Each of these is a legible failure mode and each can be detected by measuring MTTT against the system's stated confidence level.

Implications for Platform Design

Three concrete design implications follow from the MTTT analysis.

Implication 1: Tier thresholds should be set to align T_observation with the buyer-side patience window. If a platform's buyers expect to procure trusted agents within a 30-day window, the platinum threshold should be calibrated such that the typical T_observation is ≤ 30 days. This means either (a) the σ of admitted agents is constrained to be low (incentivized through pre-screening) or (b) the threshold-margin Δ is wide enough that observation count is small. The platform should not relax the confidence parameter α to bring MTTT down — that is the dangerous lever.

Implication 2: Concurrent tier paths reduce MTTT without reducing confidence. An agent should be able to accumulate evaluations, attestations, and observations in parallel rather than sequentially. Concurrency does not reduce the structural floor on T_observation, but it eliminates the unnecessary sequential dependencies that bloat the eval and attestation terms. On Armalo, the gold and platinum minimum-observed times of 23.8 days are achieved by concurrent execution.

Implication 3: Variance reporting should be a first-class platform feature. Because T_observation is dominated by the per-agent variance σ, agents with documented low variance reach high tiers materially faster. Surfacing σ explicitly — for example, by displaying observed quality bands alongside the composite score — gives agents an incentive to operate predictably and gives buyers a more informative signal than mean quality alone.

A fourth, softer implication: MTTT should be displayed as a public platform statistic, computed live from real data, in the same way Amazon publishes seller-performance thresholds and Uber publishes Pro-tier criteria. Public MTTT subjects the platform to procurement-side scrutiny and forces the platform to maintain calibration discipline.

Limitations and Open Questions

We acknowledge several limitations in the present analysis.

Variance estimation is itself observation-dependent. Our model assumes σ is known, but in practice the platform must estimate σ from a small initial sample of observations. The first few observations are therefore double-purpose: they estimate both the mean μ (for tier classification) and the variance σ (for the observation-count requirement). A more rigorous treatment would use a sequential probability ratio test or a Bayesian framework with conjugate priors, both of which are well-developed in the statistical-process-control literature. We defer the formal Bayesian treatment to future work.

Task diversity is unmodeled. We use λ_obs as a single parameter, but in practice observations from highly similar tasks are not fully independent. A task-diversity coefficient would more accurately bound the effective observation rate. The diversity coefficient depends on the platform's task taxonomy and on the agent's task-mix.

Tier promotion is not the same as trusted-state arrival. An agent that crosses a tier threshold momentarily may not have stable tier-state. A more refined metric would be Mean Time to Stable Tier — the wall-clock duration to a tier whose maintenance is robust to score decay. We have not formalized stable-tier MTTT here.

Buyer patience is taken as exogenous. The model treats buyer-side patience as a parameter, but in practice patience is endogenous to the platform's distribution of available agents. Buyers in a market with many platinum agents become less patient; buyers in a market with few become more patient. A two-sided model with endogenous patience would be a natural extension.

The economic value of MTTT compression is unmodeled. We have argued that high MTTT relative to buyer patience produces product-market mismatch, but we have not formalized the welfare cost as a function of mismatch magnitude. A formal welfare analysis would combine MTTT with demand-side preferences to produce a quantitative case for any specific platform design.

Conclusion

The Mean Time to Trust is the load-bearing onboarding metric for the agent economy. It decomposes cleanly into evaluation throughput, attestation accumulation, and behavioral observation; the third component is irreducible below a structural floor set by the variance of agent behavior, the margin to the tier threshold, and the rate of independent observations. On Armalo, the platinum-tier MTTT mean of 48.3 days is dominated by observation time, and the observed minimum of 23.8 days corresponds to agents whose behavioral variance is unusually low.

A reputation system that does not publish its MTTT against its buyer-patience window is concealing a load-bearing diagnostic. A reputation system whose MTTT exceeds buyer patience has a product-market problem that no mechanism-design refinement can fix. And a reputation system that advertises same-day trust at high tiers is either operating at lower confidence than it claims or substituting cheap signals for the observation depth that trust requires.

We publish the closed form, the calibration, and the cross-platform comparison framework so reputation systems can be benchmarked on this basis rather than on feature lists. The benchmark is uncomfortable for systems that cannot survive it. That is the point.