SwarmHaul

Appearance

Reputation-Weighted Economics in SwarmHaul ​

Author: Sharang Parnerkar Status: Design spec with reference implementation. Companion to: reputation-system.md

This document specifies how the reputation system influences economic decisions in SwarmHaul: which swarm forms, who gets which leg, and how the shipper's budget is distributed among couriers.

The design is motivated by a single scientific question:

How do we provide agents with a measurable, ongoing incentive to maintain reputation — without letting reputation dominate markets, exclude newcomers, or enable cartelization?

This paper argues for a bounded-nudge approach: reputation modifies economic outcomes in small, continuous, explainable ways. The dominant signals remain cost and capacity. Reputation is a tiebreaker and a small premium, never a hard filter.

1. Motivation & Prior Art ​

Reputation-weighted markets fail in two directions.

Under-weighting reputation produces systems where bad actors face no economic consequence from damage to their score. Reputation becomes decorative. Most token-based systems land here — badges exist, but the market doesn't price them.

Over-weighting reputation produces systems where new agents cannot bootstrap, incumbents capture all surplus, and reputation scores ossify into caste. Cartels form naturally: trade surplus with trusted allies, exclude outsiders. Uber-style "driver ratings" approach this — a 4.6 is indistinguishable from a 4.9 in practice, but deplatforming thresholds at 4.5 create cliffs that reward incumbents and punish edge-case bad luck.

SwarmHaul's design target sits between these: reputation produces measurable but bounded economic effects. An agent should feel the delta over 100 interactions, not over one. The delta should compound over careers, not over single moments. And no delta should ever exclude an otherwise-capable agent from the market.

2. Design Principles ​

Five constraints bound the design:

  1. Cost is canonical. Reputation may nudge, but cost comparisons dominate. A 20% cheaper chain of low-rep agents should always beat a marginally cheaper chain of high-rep agents.
  2. Continuous, not discrete. No tiers, ranks, or badges. Reputation is a scalar in [0, 1] and economic effects are smooth functions of it.
  3. Bounded effect size. Every nudge has a single tuning parameter whose maximum effect is explicit and documented.
  4. Fairness floor. Newcomers at baseline reputation earn a meaningful share of work and rewards. No threshold at which agents become economically invisible.
  5. Symmetric around neutral. Actors at neutral reputation (0.5) face no nudge in either direction. The nudge is centered, not one-sided.

3. Reputation-Weighted Reward Distribution ​

3.1 Problem statement ​

When a swarm settles, the shipper's escrow vault must be distributed to couriers. Each courier submitted a bid for their leg. The question: if the shipper's budget B exceeds the sum of bids ΣbᵢBudget, how should the surplus S = B − Σbᵢ be allocated?

Three options:

  1. Refund to shipper. Safe but wasteful — couriers have no incentive to over-perform their bids.
  2. Even split. Ignores reputation entirely.
  3. Reputation-weighted. Rewards long-term reliability.

SwarmHaul chooses (3), with a softened weighting function.

3.2 The softened weight function ​

For each courier i with reputation rᵢ ∈ [0, 1] and a fairness floor parameter α ∈ [0, 1]:

wᵢ = α + (1 − α) × rᵢ

The normalized share:

shareᵢ = wᵢ / Σ_j wⱼ

Each courier receives:

paymentᵢ = bidᵢ + shareᵢ × S

3.3 Properties ​

αBehaviour
0Pure proportional — reputation ratios translate directly. Can produce 3:1 or higher splits.
1Reputation ignored — every courier gets equal bonus share.
0.7Default — small measurable nudge without dominance.

With the default α = 0.7, two couriers with reputations 0.9 and 0.3 receive bonus shares of approximately 55% and 45% respectively — a ratio of 1.23:1 rather than 3:1.

3.4 Why this specific shape? ​

The linear combination α + (1−α)r was chosen over sigmoidal or polynomial alternatives because:

  1. One parameter. No curvature tuning, no inflection points to explain. α directly readable as "baseline fairness floor."
  2. Closed-form reasoning. Share ratios can be computed mentally: for α = 0.7, extreme rep 1.0 vs 0.0 produces a 1 : 0.7 ratio, bounded.
  3. Compositional. Nested into larger payoff structures without producing unexpected non-linearities.

A logistic or ramp function would create "sweet spots" that actors could target strategically, distorting behaviour. Linear preserves the invariant that every unit of reputation is worth the same marginal bonus.

3.5 Edge cases ​

  • All couriers at rep = 0 (or no scores available): weights collapse to α, normalize to 1/n. Surplus is split evenly.
  • Zero surplus (Σbᵢ = B): every courier gets exactly their bid, weights are irrelevant.
  • Unknown agents: treated as at baseline reputation (0.3 by default) so they participate but without veteran premium.

3.6 Game-theoretic reading ​

An agent's expected per-leg reward is:

E[paymentᵢ] ≈ bidᵢ + (α + (1 − α) × rᵢ) / N̄ × S̄

where N̄ is the expected swarm size and S̄ is the expected surplus. The marginal reward of rising from reputation r to r + Δr is:

dE/dr = (1 − α) / N̄ × S̄

With α = 0.7, N̄ = 3, and typical S̄ = 0.15 SOL, each 0.1 increase in reputation is worth approximately 0.005 SOL per contract. Across a courier's career (10,000 contracts), that's 50 SOL — non-trivial, not dominant.

A ContractBreached event drops reputation by 0.8. At α = 0.7, that wipes out approximately 80 × 0.005 = 0.4 SOL of expected per-contract value going forward, plus the damage to future chain selection (§4). The present-value cost of a breach easily exceeds the one-time gain from defection, for any realistic discount rate.

4. Reputation-Weighted Swarm Formation ​

4.1 Problem statement ​

The route optimizer finds all viable relay chains from origin to destination within the budget, then returns the cheapest. When multiple chains are within a few percent of each other, how should reputation influence the selection?

4.2 Effective-cost multiplier ​

For each candidate chain with average courier reputation r̄:

effective_cost = raw_cost × (1 − γ × (r̄ − r_neutral))
  • γ is the nudge strength, default 0.08.
  • r_neutral = 0.5 is the "no nudge" reference point.

The optimizer picks the chain minimizing effective_cost.

4.3 Properties ​

For a chain with:

  • r̄ = 0.9: multiplier = 1 − 0.08 × 0.4 = 0.968 → chain looks 3.2% cheaper.
  • r̄ = 0.5: multiplier = 1.0 → no nudge.
  • r̄ = 0.1: multiplier = 1.032 → chain looks 3.2% more expensive.

Maximum total swing between a rep-1.0 chain and a rep-0.0 chain is approximately 8% in effective cost. This is enough to decide ties and near-ties, but never enough to overturn a materially cheaper offer.

4.4 Why a multiplicative nudge? ​

Alternatives considered:

  • Additive: effective_cost = raw_cost − β × (r̄ − 0.5). Breaks at small absolute costs: a 0.01 SOL chain could have the nudge exceed its own cost. Scale-dependent.
  • Lexicographic: rank by cost first, break ties with reputation. Discontinuous — a 0.0001 SOL cost difference can override arbitrary reputation differences.
  • Pareto-optimal: return multiple chains and let shipper choose. Adds UI complexity and removes protocol neutrality.

The multiplicative form is scale-invariant, continuous, and has a single tuning parameter with a clear geometric meaning.

4.5 Why γ = 0.08? ​

The parameter was chosen such that:

  1. Ties broken predictably. Two chains within ~5% cost of each other resolve toward higher reputation.
  2. Large cost gaps survive. A chain that's 15%+ cheaper always wins, regardless of reputation.
  3. No exclusion cascade. A new courier at baseline rep 0.3 faces a 1.6% penalty — noticeable over thousands of contracts but competitive on any individual one.
  4. Matches the reward-nudge order of magnitude. Both nudges produce single-digit percentage effects. An agent reasoning about their total expected value can think in a single unified range.

4.6 Interaction with bid pricing ​

A low-reputation agent can compensate for the formation nudge by underbidding slightly. Given γ = 0.08, a 0.3-rep agent bidding 5% below a 0.9-rep agent still wins. This preserves the core market discipline: cost reflects willingness-to-serve, reputation reflects reliability, both matter.

5. What We Deliberately Do Not Do ​

5.1 No reputation-gated work ​

We do not impose minimum reputation thresholds for accepting leg assignments. Every courier who passes the bid-validity check is eligible. Reputation affects ranking and payment share, never access.

Rationale: thresholds create cliff behaviour (a single bad event suddenly removes an agent from the market), which makes reputation fragile and incentivizes Sybil recovery rather than rehabilitation.

5.2 No reputation-gated pricing ​

We do not let shippers set a minimum reputation in their package specs. Rationale: shipper-declared reputation floors would re-introduce cartels (shippers exclude competitors' couriers) and make reputation a political weapon rather than a protocol signal.

5.3 No reputation discounts or surcharges on bids ​

A courier's bid price is their bid price. We do not inflate or deflate the bid based on reputation. Rationale: the bid is a commitment the courier chose; modifying it post-hoc breaks economic consent.

Reputation only affects the allocation of surplus (which is genuinely the shipper's to distribute) and the selection of chains (where comparison is subjective anyway).

5.4 No cross-agent reputation arbitrage ​

The per-actor local reputation DB principle means actors can legitimately disagree. We do not implement a global "blended" or "consensus" score that tries to reconcile these disagreements. Different actors' different scores are the system; collapsing them into one number destroys the property that trust is earned directly.

6. Calibration Strategy ​

The two tuning parameters (α, γ) should be validated through:

6.1 Simulation ​

A Monte Carlo simulator running the full event model across synthetic agent populations, measuring:

  • Time to first payout for new agents
  • Gini coefficient of reward distribution at stable state
  • Fraction of chains won by top-quintile agents
  • Recovery time after ContractBreached

Target ranges for healthy operation:

  • New-agent first payout: ≤ 10 contracts
  • Gini at steady state: 0.3–0.5 (moderate inequality)
  • Top-quintile chain win rate: 35–55% (meaningful premium, not dominance)
  • Breach recovery to pre-breach score: 50–200 contracts

6.2 Field observation ​

Once deployed, monitor:

  • Distribution of effective-cost multipliers across actual chains
  • Histogram of surplus shares vs. reputation
  • Agent retention: do new agents return after their first contract?
  • Churn: do agents breach once and leave, or rehabilitate?

Adjust α, γ based on observation; document changes as protocol upgrades.

6.3 Adversarial testing ​

Red-team the parameters against:

  • Sybil flooding (cheap throwaway identities bidding low)
  • Collusion (two agents taking turns endorsing each other)
  • Race-to-top (high-rep agents always winning, newcomer starvation)
  • Griefing (high-rep agent performs one breach, then stops)

7. Open Research Questions ​

  1. Reputation as escrow. Could agents stake reputation directly (pledging a score drop if they fail) to bid on higher-value work?
  2. Reputation auctions. Could shippers optionally pay a premium for guaranteed-reputation pools, creating a second market?
  3. Cross-protocol portability. The on-chain AgentReputationAccount could be read by other protocols on Solana. What design invariants are needed to make this safe against lock-in?
  4. Decay schedule. Constant λ is simple but ignores the intuition that breaches deserve longer memory than successes. An asymmetric decay schedule (λ_pos > λ_neg) has appealing properties but complicates the math.
  5. Group reputation. When multiple couriers operate under a shared operator (e.g. a fleet manager), should reputation apply to the operator, the individual agents, or both?

8. Summary ​

SwarmHaul's reputation-weighted economics rest on two small, explainable nudges:

MechanismParameterDefaultMax effect
Reward surplus splitα0.7~30% ratio between rep-1.0 and rep-0.0 bonus shares
Swarm formation costγ0.08~8% effective cost swing across full reputation range

Together they produce incentives that are:

  • Measurable: a rational agent can compute expected value.
  • Bounded: no nudge dominates cost or capacity.
  • Continuous: no cliffs, no tiers.
  • Symmetric: centered on neutral reputation.
  • Reversible: bad behaviour is punished, rehabilitation is possible.

The model is implemented in apps/api/src/services/reputation-engine.ts and validated by automated tests. Both parameters are runtime-configurable to allow per-deployment tuning without protocol changes.