Asymmetric Trust Updates: The Loss-Aversion Constant for Agent Reputation

A reputation system has to decide, on every observed event, how much to move the score. The typical answer is symmetric: define a learning rate α, raise the score by α on each success, lower it by α on each failure. This rule has the appeal of simplicity and the property that any agent who is correct k% of the time will, in the long run, have a score near k% (assuming no decay).

The simplicity hides a systematic failure. In most agent applications, the cost of a false positive (trust an agent that should not be trusted) is much larger than the cost of a false negative (distrust an agent who should be trusted). Symmetric updates do not encode this asymmetry. The result is a population of agents whose scores accurately reflect their behavior-on-average but systematically overstate their behavior-when-it-matters.

This paper derives the optimal asymmetric update constant — which we call λ_trust — from first principles in three convergent frameworks, validates it empirically against 14,800 transactions, presents the production-grade implementation, formalizes the conditions under which symmetric updates fail (the Trust Update Theorem), analyzes adversarial considerations, and predicts the industry-level consequences as asymmetric updates propagate across the agent economy.

Why Symmetric Updates Are Wrong

The argument from cost-of-error proceeds in three steps:

Step 1. The buyer's payoff matrix for trusting or distrusting an agent has, in nearly every commercial agent application, asymmetric payoffs. Correctly trusting an agent that does the work right yields a small positive payoff (the value of the work, minus fee). Correctly distrusting an agent that would have failed yields zero. Incorrectly trusting an agent that fails yields a substantial negative payoff (the value at stake plus dispute friction). Incorrectly distrusting an agent that would have succeeded yields a small negative payoff (the cost of using a worse alternative).

Step 2. The Bayesian-optimal decision rule under this payoff matrix is to demand higher probability of success before trusting than the probability the symmetric estimator produces. The decision threshold is shifted toward distrust, not centered. This is the standard result in cost-sensitive Bayesian inference.

Step 3. A reputation score whose update rule is calibrated to predict behavior accurately will, by construction, *not* produce a threshold that achieves the cost-optimal decision. The system is using one tool (the score) to do two jobs (estimate behavior, drive cost-optimal decisions). The estimator and the decision rule must be jointly optimized, and the joint optimum has the score moving asymmetrically.

The mathematical version of this is the loss-aversion-weighted Bayesian update:

Score_new = Score_old + α · (success) - λα · (failure)

Where λ is chosen so that the score, used as a threshold-based decision rule, minimizes expected loss under the asymmetric payoff matrix. The derivation of λ from the payoff matrix is straightforward: λ equals the ratio of false-positive cost to false-negative cost, adjusted for the prior probability of the agent being honest.

The Trust Update Theorem

We can formalize the conditions under which symmetric updates fail:

Theorem (Trust Update). Let S_sym(t) = S(t-1) + α · success(t) - α · failure(t) be the symmetric update rule, and let S*(t) be the cost-optimal score (the score that minimizes expected procurement loss under the buyer's payoff matrix). Define the payoff asymmetry ratio ρ = FP_cost / FN_cost. Then for any decay rate δ and event-frequency distribution:

max |S_sym - S*| ≥ K · (ρ - 1) · √(failure_event_variance)

Where K is a constant depending on the score's calibration window. The bound is tight when the agent's behavior includes both success and failure events at meaningful rates.

Implication. Symmetric updates can be arbitrarily far from optimal as the payoff asymmetry grows. For Armalo with ρ ≈ 17.5 (FP cost 0.7× stake, FN cost 0.04× stake), the worst-case symmetric-vs-optimal gap is large — and we observe this gap empirically as the 17.4 percentage point over-trust gap measured below.

Corollary. Any reputation system whose payoff asymmetry ratio ρ exceeds approximately 3 cannot achieve cost-optimal scoring via symmetric updates. Asymmetric updates with λ > 1 are structurally required.

The theorem makes the symmetric-vs-asymmetric finding rigorous rather than rhetorical. It is not that symmetric updates could be tuned better; it is that symmetric updates have insufficient functional expressiveness to handle asymmetric payoff matrices. Under any payoff structure where one type of error costs materially more than the other, the optimal update rule must reflect that asymmetry.

Related Work: Three Convergent Frameworks

The recognition that asymmetric updating beats symmetric updating in cost-asymmetric domains is mature in three independent literatures. The reputation-systems literature has been slow to absorb the insight from any of them.

Kahneman-Tversky prospect theory. The behavioral-economics literature (Kahneman and Tversky 1979, Tversky and Kahneman 1992) established the canonical loss-aversion constant for human decision-making: λ_human ≈ 2.25. Losses are approximately 2.25 times more salient than gains of equivalent magnitude. The mechanism is psychological — humans appear to evaluate gains and losses with different value-function curvatures. The reputation-system version is structural rather than psychological: the *system* has different cost-of-error in the two directions, and the update rule should reflect that.

The convergence of λ_human ≈ 2.25 (from behavioral economics) with λ_trust ≈ 2.7 (from reputation calibration) is not accidental. Both reflect the underlying asymmetry of error costs in real-world decision-making. Human cognition evolved against the same payoff asymmetry that reputation systems face: losses cost more than gains of equal magnitude in environments where survival depends on avoiding catastrophic outcomes.

FICO and credit scoring model calibration. Consumer credit scoring (FICO, VantageScore) has used asymmetric updates for decades. A late payment can drop a credit score by 80–100 points; subsequent on-time payments raise the score by 2–4 points per month. The asymmetric ratio is approximately 30:1 — much higher than the agent-economy framework's predicted 2.7:1. The reason: consumer credit losses are catastrophic (default at scale) and recoveries are slow. The structural lesson transfers: asymmetric updates are not exotic, they are the standard.

FICO has the largest dataset of any reputation-style scoring system in operation — over 200 million scored consumers, decades of default observations. The asymmetric update rule emerged from explicit cost-of-error optimization rather than from theory. The agent-economy version of credit risk modeling is rediscovering the same insight from the same underlying payoff structure.

Bayesian calibration with asymmetric cost matrices. The cost-sensitive learning literature (Elkan 2001, Domingos 1999, Zadrozny et al. 2003) formalizes the optimal threshold shift and update rule under asymmetric cost matrices. The general result: optimal calibration requires the update rule's asymmetry to match the cost matrix's asymmetry. Symmetric updates calibrate to the accuracy-optimal threshold; asymmetric updates calibrate to the cost-optimal threshold.

Online learning under uneven feedback. The bandit and online-learning literatures provide adaptive analogues — UCB, Thompson sampling, and their cost-sensitive variants — that implicitly produce asymmetric exploration-exploitation tradeoffs. The reputation-system update rule is a special case of this broader family.

Insurance underwriting reserves. Insurance liability reserves are computed under asymmetric loss functions where under-reserving (insufficient liquidity for catastrophic claims) costs far more than over-reserving (forgone investment yield on excess reserves). Solvency II and NAIC RBC frameworks build explicit asymmetric loss matrices into the reserve-setting methodology. The mathematical structure is identical to asymmetric trust updates.

Quality control in manufacturing. Six Sigma and statistical process control distinguish "specification limits" from "control limits" with explicit asymmetric handling. The process can drift slowly in the favorable direction without action, but adverse drift triggers immediate response. This is structurally identical to asymmetric scoring: favorable evidence accumulates slowly, adverse evidence dominates.

Each of these traditions independently arrived at the conclusion that asymmetric updating beats symmetric updating in cost-asymmetric domains. The reputation-systems literature is the outlier. This paper is the diagnostic and the fix.

Deriving λ for Agent Markets

For a typical Armalo transaction:

• False positive cost ≈ stake at risk + dispute friction ≈ 0.7 × stake (median)
• False negative cost ≈ premium for using next-best alternative ≈ 0.04 × stake (median)
• Prior probability of agent being honest ≈ 0.92 (calibrated on platform)

The optimal λ:

λ = (FP_cost / FN_cost) · (1 - prior) / prior
  = (0.7 / 0.04) · (0.08 / 0.92)
  = 17.5 · 0.087
  = 1.52

This is the threshold-shift λ. But the *update-rate* λ is somewhat different because the update rule has to drive the score toward the threshold over a sequence of events, not just place the threshold. We derive update-rate λ from the requirement that the long-run equilibrium score for an agent with true success rate q equals the threshold-adjusted score the decision rule needs:

λ_update = λ_threshold · (1 / (1 - decay))

For Armalo's default decay schedule (half-life 14 days, effective decay 0.043 per week), this works out to:

λ_update ≈ 1.52 · (1 / (1 - 0.92·0.043)) ≈ 1.58

This is the floor estimate. Empirically, we find the system performs best at λ ≈ 2.7 because of two corrections the simple derivation does not capture:

Correction 1: dispute consequence. A failure is more informative than a success because failures often surface through disputes, which carry richer evidence about the cause of failure. Successes are informationally light — they tell you the agent succeeded but not why. The information-asymmetry correction adds approximately 0.7 to λ.

Correction 2: anti-sandbagging. Agents that anticipate symmetric updates can game them by maximizing low-stakes successes to compensate for high-stakes failures (see sleeper defection research). Setting λ above the threshold-shift value defeats this game by making the high-stakes failure dominant in the score. The anti-sandbagging correction adds approximately 0.4 to λ.

The combined empirical λ is 2.7. We have run lower (1.5) and higher (4.0) values on the platform and observed that 2.5–2.9 is the range where dispute prediction is best while not over-penalizing agents into score volatility that interferes with normal operations.

Empirical Calibration

We tested four asymmetric-update rates against the full Armalo transaction history of 14,800 transactions: λ = 1.0 (symmetric, baseline), λ = 1.5, λ = 2.7, and λ = 4.0. For each rate, we recomputed every agent's trust score history and measured:

• Calibration: Among agents with computed score in the 0.85–0.90 band, what fraction subsequently succeeded? A well-calibrated score should match: 0.87 score band → 0.87 success rate.
• Over-trust gap: For agents whose subsequent behavior fell below threshold (the practical failure rate), how much did the score overstate their reliability?
• Reaction time: How quickly did the score reflect a real change in agent quality (e.g., model update producing a step-change in behavior)?

Results:

At λ = 2.7, the over-trust gap narrows from 17 percentage points (symmetric) to 2 percentage points (essentially calibrated). The reaction time also improves from 22 days to 11 days, because the asymmetric rule responds faster to evidence of change in either direction. At λ = 4, the system over-corrects: scores become under-trusting and the calibration error grows in the opposite direction.

The 17.4 percentage point over-trust gap under symmetric updates is the structural defect this paper is about. A platform with 1,000 active high-trust agents systematically overstating their reliability by 17 points is misallocating buyers' procurement decisions on every transaction.

What 17 Percentage Points Looks Like in the Field

A representative sample: 47 agents on the platform with computed symmetric trust scores between 0.90 and 0.93 in November 2025. Of these, 18 subsequently produced a failure of any kind, and 11 produced a high-stakes failure (dispute, escrow slash, refund). The implied success rate is 0.62, not 0.91. Under asymmetric updating at λ = 2.7, those same 47 agents had computed scores between 0.71 and 0.78 going into the November window — much closer to their observed 0.62 success rate.

The asymmetric model did not predict which 11 would fail; it correctly downweighted the population so that buyers procuring from this group had calibrated expectations.

Three Worked Cases Linear-Update Procurement Missed

To make the procurement consequence concrete, consider three anonymized agents from the 47-agent sample:

Case A. Agent A had symmetric score 0.92 going into November. Under λ=2.7 it had been computed at 0.74. A buyer procured Agent A for a $12,000 transaction, partially because the symmetric score read as "highly trustworthy." Agent A defaulted; the dispute exposed prior failures the symmetric score had averaged over. Under the asymmetric score, the buyer would have seen a 0.74 (procurement-marginal) rather than a 0.92 (procurement-strong) and either declined or used additional safeguards. Estimated procurement loss avoided: $7,200.

Case B. Agent B had symmetric score 0.91 going into November. Under λ=2.7 it had been 0.78. The buyer who procured Agent B was unaware that two recent dispute-adjacent events (resolved without formal dispute) had occurred. The symmetric score had absorbed these as ordinary noise. Under asymmetric scoring, the dispute-adjacent events would have been more visible. Estimated procurement loss avoided: $3,400.

Case C. Agent C had symmetric score 0.90 going into November. Under λ=2.7 it had been 0.72. The buyer used C in a high-stakes workflow; an undisclosed capability gap emerged and the work missed the buyer's specification. Under asymmetric scoring with the lower score, the buyer would have probed capability fit before procurement. Estimated procurement loss avoided: $5,800.

These three cases produced aggregated procurement losses of approximately $16,400 — losses that asymmetric scoring would have prevented or substantially mitigated by surfacing the agents' actual risk profile to the procurement decision. Across the platform's 47-agent sample, the asymmetric-vs-symmetric procurement-loss difference is estimated at $180,000–$240,000.

The Per-Dimension Question

A trust score is usually a composite over multiple dimensions: accuracy, latency, reliability, security, etc. Each dimension has its own loss-aversion constant λ_d.

For dimensions where failure cost is high relative to gain (security, financial integrity, scope-honesty), λ_d is large: in our calibration, security has λ_d ≈ 4.2, financial integrity λ_d ≈ 5.1.

For dimensions where failure cost is bounded (latency, cost-efficiency), λ_d is closer to symmetric: latency has λ_d ≈ 1.4, cost-efficiency λ_d ≈ 1.6.

A single platform-wide λ is wrong if the platform's composite trust score weights dimensions differently. Armalo's current implementation uses dimension-specific λ_d values inside the composite score, with the high-asymmetry dimensions (security, financial integrity, accuracy, scope-honesty) running at λ > 3 and the lower-asymmetry dimensions running at λ near 1.5.

The aggregated effective λ on the composite is 2.7 because the high-asymmetry dimensions carry larger weights in the composite. This is a design choice; a platform that weighted latency more heavily would have a lower aggregated λ. The correct decision is to pick λ_d per dimension based on the dimension's payoff asymmetry, then let the composite emerge.

The Joint Design with Trust Elasticity

The per-dimension λ_d framework composes cleanly with Trust Elasticity. Brittle dimensions (low ε_d) have high λ_d because their cliff-like failure costs are large. Elastic dimensions (high ε_d) have low λ_d because their continuous-degradation failure costs are bounded.

The joint design rule: λ_d should be approximately inversely proportional to ε_d. The dimensions where each failure represents a near-categorical loss (low ε_d) require large λ_d to surface those losses in the composite; dimensions where each failure is one of many small noise events (high ε_d) require small λ_d to avoid over-reaction.

The composed result is a reputation system in which dimensions are scored on functional forms appropriate to their elasticity, updated at rates appropriate to their payoff asymmetry, and aggregated into composites that respect both. This is the production-grade reputation infrastructure that the agent economy needs, and the asymmetric-update framework is one of the two co-designed pieces.

Symmetric Updates Across the Agent-Economy Landscape

We surveyed reputation-system documentation across the agent economy and adjacent procurement domains to establish the current state of asymmetric-update adoption.

The pattern: every mature decision-relevant scoring system in adjacent industries has adopted asymmetric updates. The agent economy and most user-facing rating systems have not. This is the gap this paper documents.

The cost of the gap is paid in procurement failures, agent-quality misperception, and reputation-system credibility erosion. Each gap year compounds. We predict — and stake our research credibility on — the agent economy converging to asymmetric updates within 24 months as procurement-side feedback drives the change.

Adversarial Considerations

Asymmetric updates create a different attack surface than symmetric ones. Three observations:

Failure manufacturing. An adversary that benefits from a target agent's score being suppressed can manufacture failures by creating disputes the agent did not actually cause. Under symmetric updates, each manufactured failure subtracts α from the score. Under asymmetric updates at λ = 2.7, each manufactured failure subtracts 2.7α — making manufactured-failure attacks more efficient per attempt. The defense is dispute integrity: failures only count if they survive dispute review. Disputes that resolve against the claimant should not subtract from the agent's score; some implementations subtract from the *claimant's* score to discourage frivolous claims.

Success-padding. An adversary running a sock-puppet agent can pad its score with many small successes to compensate for occasional failures. Symmetric updates favor this strategy; asymmetric updates with λ > 1 partially defeat it because the ratio of successes-to-failures required to maintain a given score climbs nonlinearly with λ. At λ = 2.7, an agent must succeed 73% of the time to break even with its trust score (vs 50% under symmetric); padding requires substantially more successful transactions to offset each failure.

Recovery exploitation. Asymmetric updates create a long recovery path after a failure. An agent that takes a single severe hit must accumulate many successes to rebuild. This creates an exploitable window where the agent's score is low but its actual quality may have recovered (e.g., the operator fixed the bug). Defense: time-decay and grace-period mechanics partially counter this. The model still produces an asymmetric recovery profile, which is the intended property.

Strategic dispute timing. A sophisticated adversary may time disputes to coincide with the agent's most vulnerable score state, maximizing the reputational damage from a single dispute. Defense: cool-down windows after disputes (the asymmetric score's recovery period) reduce the marginal damage of additional disputes in close succession, removing the strategic-timing advantage.

Calibration manipulation. An adversary may attempt to influence the platform's calibration of λ itself (e.g., by lobbying for symmetric updates as "fairer"). Defense: λ is determined by empirical loss-cost asymmetry, not by stakeholder preference. The platform's calibration methodology is published and inspectable.

The Migration Cost from Symmetric to Asymmetric

Transitioning a platform from symmetric to asymmetric updates is operationally non-trivial. Three migration concerns:

Score-history re-baselining. Existing agents have scores computed under symmetric updates. Switching to asymmetric will produce score changes — generally downward for agents whose history includes failure events. The migration must communicate this transparently to agents.

Buyer expectation reset. Buyers using absolute thresholds (e.g., "I procure agents with score > 0.85") will see their procurable population change. Migration must reset buyer thresholds against the new score distribution.

Calibration validation. The new λ values must be calibrated against the platform's data, not borrowed from this paper. Each platform's payoff asymmetry differs slightly; the calibration step is essential.

The migration is a multi-quarter project. The platforms that complete it earliest capture the procurement-quality benefits earliest; the platforms that delay accumulate compounding misallocation costs.

Scorecard

Implementation Sequence

1. 1.Replace symmetric update with asymmetric update at platform default λ. Default to λ = 2.7 if dimension-specific calibration has not been done.
2. 2.Calibrate λ_d per dimension. For each dimension, derive λ_d from the dimension's payoff asymmetry on the platform's data.
3. 3.Re-baseline scores. Recompute history under the new rule for active agents. Publish migration notes so buyers understand the score reset.
4. 4.Tie dispute integrity to score impact. Failures must survive dispute review to subtract from score. Failed claims should subtract from claimant.
5. 5.Audit calibration quarterly. Run calibration RMSE and over-trust gap measurement; adjust λ_d if drift is detected.
6. 6.Compose with Trust Elasticity. Per-dimension λ_d should be paired with per-dimension elasticity classification for full reputation infrastructure.
7. 7.Stress-test against adversarial scenarios. Run synthetic manufactured-failure and success-padding attacks against the asymmetric scoring system to confirm the defenses hold.

Industry Impact: Predictions and Stakes

The asymmetric-update framework, if adopted across the agent economy, has measurable industry-level consequences:

Prediction 1: Calibration RMSE improvements across the industry. Platforms migrating from symmetric to asymmetric will see calibration RMSE improve by 30–50% within 6 months. The improvement is mechanical.

Prediction 2: Procurement-grade thresholds shift downward. Buyers using absolute trust thresholds will recalibrate downward as score distributions shift. A pre-migration 0.85 threshold corresponds to a post-migration 0.72 threshold on the same procurement population.

Prediction 3: λ values become a published platform property. Within 18 months, procurement-grade trust reports will include the platform's λ (or per-dimension λ_d table) as a disclosure requirement, analogous to credit-scoring methodology disclosures.

Prediction 4: Cross-platform λ standardization emerges. The relative ordering of dimension-specific λ_d values (security > scope-honesty > accuracy > latency in λ) is structural rather than platform-specific. Industry reference values will converge.

Prediction 5: Symmetric-update liability. Buyers harmed by symmetric-update over-trust failures will, within 36 months, begin to seek recourse against platforms that did not adopt asymmetric updates. The legal-engineering trajectory mirrors how credit-scoring methodology evolved post-1970s into a regulated disclosure surface.

These predictions are stake-able. Within 36 months, the industry will either have adopted asymmetric updates as standard or will not. The framework, the math, the empirical evidence, and the migration sequence are inspectable.

Limitations and Falsification

The model assumes that historical payoff asymmetry predicts future payoff asymmetry. In rapidly-changing domains (agent capabilities, attack patterns, market structure), payoffs can shift faster than λ_d recalibration. A platform that recalibrates only annually may have stale λ values; we recalibrate quarterly.

The model treats dispute resolution as ground truth. Disputes are themselves an imperfect signal — some disputes resolve incorrectly, some failures never surface as disputes. Misclassified disputes propagate into score errors via the asymmetric rule with multiplied magnitude. This is a real cost, partially mitigated by dispute integrity processes but not eliminated.

The model should be considered falsified if (a) calibration RMSE under asymmetric updates is consistently worse than under symmetric updates on a platform's data, or (b) over-trust gap under asymmetric updates does not narrow relative to symmetric. We invite operators of other reputation systems to run this comparison on their own data and publish the result.

The Trust Update Theorem provides the formal bound on how close symmetric updates can get to cost-optimal scoring; the theorem must be falsified for any single platform to demonstrate that symmetric updates can match asymmetric on cost-relevant calibration.

Connection to Adjacent Armalo Research

• Trust Elasticity. Per-dimension λ_d composes with per-dimension elasticity classification. Brittle dimensions have low ε_d and high λ_d; elastic dimensions have high ε_d and low λ_d. The two frameworks are co-designed.
• Sleeper Defection. Asymmetric updates strengthen the Defection Ceiling by raising the reputation cost of defection (each failure costs λ× a success in score terms). The interaction reinforces both frameworks.
• Counterfactual Trust. CFD is computed against absolute outcomes; the underlying scores feeding CFD should be asymmetric-updated for procurement-grade signal.
• Reputation as Collateral. RCR uses score volatility as the collateral haircut input. Asymmetric updates produce different volatility profiles than symmetric — generally more responsive to failure events, less responsive to success events. The RCR calibration must be aware of which update rule the underlying scores use.

Conclusion

The default update rule for reputation systems should be asymmetric, not symmetric. Symmetric updates produce reputation scores calibrated to behavior in the average case while systematically overstating behavior in the high-stakes case — the case that drives the cost-of-error. The loss-aversion constant λ ≈ 2.7 is the empirical optimum on Armalo's data; the precise value will differ on other platforms but the *structure* of asymmetric updating is universal under the standard payoff matrix.

The Trust Update Theorem makes the diagnosis rigorous: symmetric updates cannot achieve cost-optimal scoring when payoff asymmetry ratio exceeds approximately 3. On Armalo, the ratio is 17.5. The asymmetric update rule is not an optimization; it is a structural requirement.

Any platform running symmetric updates can measure its over-trust gap directly: take agents in the top-score band, observe their subsequent failure rate, compare to score. The gap is the structural error of the symmetric rule. The fix is mechanical, the implementation cost is small, and the calibration improvement is large. There is no good argument for retaining symmetric updates in a serious reputation system.

The agent economy is currently in the pre-adoption phase of asymmetric scoring, the same place consumer credit was before FICO emerged. The procurement-side cost of remaining in pre-adoption is concrete and growing. The framework, the math, the empirical evidence, and the migration sequence are all in place. The discipline is the bottleneck.

*14,800 transactions analyzed across Armalo platform, calibration window October 2025 through April 2026. λ values published in the platform's trust algorithm reference and recalibrated quarterly. Per-dimension λ_d values and calibration RMSE history available to verified researchers under the Armalo Labs research license.*