ReputationChain: Robust Trust Updating for Blockchain-Enabled Supply Chains

ReputationChain: Robust Trust Updating for Blockchain-Enabled Supply Chains Adnan Iftekhar

a,e

, Chengliang Zhengb , Xiaohui Cuia,* , and Mir Hassanc,d

a

arXiv:2608.03554v1 [cs.CR] 4 Aug 2026

b

Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, Wuhan 430072, China School of Artificial Intelligence (School of Blockchain Industry), Chengdu University of Information Technology, Chengdu 610225, China c Human Environment Technology Systems Centre, Mykolas Romeris University, Vilnius LT-08320, Lithuania d Azerbaijan State Oil and Industry University, 34 Azadliq Avenue, Baku, Azerbaijan e The University of Faisalabad, Faisalabad, Pakistan Emails: adnan@whu.edu.cn; zhengcl@cuit.edu.cn; xcui@whu.edu.cn; mir.hassan@mruni.eu * Corresponding author: Xiaohui Cui (xcui@whu.edu.cn).

Abstract—A blockchain can preserve the record of a supplychain interaction, but an intact ledger says nothing about whether the participant behind that record deserves trust in the next risk-sensitive transaction. Existing reputation designs concentrate on product evidence, global feedback aggregation, or review authenticity, and three problems that dominate permissioned supply chains receive far less attention: inflation through repeated bilateral interactions, the advantage gained from identity multiplicity, and decay rules that quietly punish honest participants whose history is simply sparse. We present R EPUTATION C HAIN, a participant trust framework that treats the blockchain as an evidence and provenance layer rather than as the source of trust itself. Governed interaction outcomes become bounded evidence; repeated interactions between the same pair are discounted, low counterparty diversity is penalized, governance-supplied identity confidence weights positive evidence, and scores decay toward a neutral prior at a rate set by verified interaction volume. Identity, contract, outcome, and update provenance stay on chain, while the nonlinear computation runs off chain and is checked on chain for admissibility. In controlled simulations over 30 seeded runs with matched interaction traces, the full model cuts mean collusive gain to 0.1443 (against 0.3688 for naive mean evidence and 0.3585 for static decay). With ten identities under one controller, the reputation inflation ratio drops to 0.8723 while three comparison baselines stay above 1.08. On identical newcomer traces, volume-aware decay raises mean newcomer reputation from 0.6626 to 0.7589 and lowers the false low-trust rate from 0.3633 to 0.1683; seed-paired analyses confirm that these differences hold run by run. What the evidence supports is a bounded reduction in reputation distortion, not attacker detection, and it is simulation evidence: deployment measurement and calibration against operational data remain necessary before production use. Index Terms—blockchain, supply chain, reputation, trust management, collusion, Sybil attack, newcomer fairness

I. I NTRODUCTION Blockchain platforms now serve as shared recordkeeping infrastructure across many supply-chain consortia, and the case for them usually rests on traceability and auditability [1], [2]. Those properties settle disputes over whether an event was recorded and whether the record was later altered. They leave a different question open. Should this supplier, this carrier, or this retailer be trusted in the next transaction that creates real financial or operational exposure? A ledger can hold a perfectly preserved record of a delayed settlement or a failed delivery

and still offer no defensible way of turning that evidence into a participant-level trust signal. The question is hardest in inclusive supply chains, where small firms, cross-border partners, temporary collaborators, and third-party providers arrive with little shared history [3], [4]. A trust mechanism for this setting has to steer between two failure modes at once. If repeated successful interactions inside a narrow group are allowed to compound without limit, reputation becomes something a closed clique can manufacture. If thin history is read as evidence of unreliability, the mechanism ends up protecting incumbents rather than the supply chain, and it locks newcomers out of exactly the interactions that would let them build a record. Prior work supplies important pieces of the answer. TrustChain fuses multiple evidence sources for food supply chains [5], and DeTRM models commodity and participant trust together in blockchain-supported supply chains [6]. MeritRank bounds the benefit of Sybil identities in graph-based reputation [7], while BlockRep ties review authenticity to cryptographic purchasing evidence [8]. Each of these solves an adjacent problem. None of them jointly targets participant-level trust updating under governed interactions, where repeated bilateral reinforcement, governance-side identity assurance, and sparse history all feed into later risk-sensitive decisions. We designed R EPUTATION C HAIN with that gap in mind. The framework uses the blockchain as an evidence and statemanagement layer, not as the source of trust. Registered participants interact under governing contracts; verified completion, delay, failure, and dispute events become bounded evidence; and the reputation layer then applies repeatedpair discounting, counterparty-diversity adjustment, identityconfidence weighting, and volume-aware decay toward a neutral prior. This paper makes four contributions. 1) We define an auditable participant trust architecture that separates blockchain-recorded evidence from the numerical reputation computation. 2) We develop a bounded, event-indexed update model that limits the influence of repeated bilateral interactions and of identity multiplicity while preserving adverse evidence in full.

3) We introduce volume-aware decay toward a neutral prior, [8]. We do not attempt to replace either. The gap we address which lowers the rate at which honest newcomers are is narrower and operational: bounded participant trust updates falsely classified as low trust. inside a governed supply-chain consortium. Table I summarizes 4) We evaluate the mechanisms against baselines, through where each system sits relative to ours. targeted ablations, and under parameter sensitivity One naming clarification is needed. RepChain [15] uses sweeps, with 30 seeded runs and matched traces per validator reputation for shard formation, leader selection, and condition, and we report seed-paired effects with confi- incentives, and therefore operates inside the consensus layer. dence intervals. R EPUTATION C HAIN operates above it: the scores describe business participants and never touch validator selection or The claims are deliberately narrow. R EPUTATION C HAIN consensus voting. does not detect every colluder, and it does not identify Sybil identities; what it does, under the tested assumptions, is III. S YSTEM AND T HREAT M ODEL reduce how much reputation an attacker can extract. The distinction is worth stating up front, because our strongest A. Participants and Governed Interactions identity-multiplicity experiment still leaves a high false trust Let elevation rate, and we report that number rather than bury it. E = {E1 , E2 , . . . , Ez } (1) II. R ELATED W ORK AND R ESEARCH G AP denote the registered supply-chain participants. Each participant A. Reputation in Distributed Systems is bound to a blockchain-recognized identity with a public EigenTrust, PeerTrust, and the Beta Reputation System key, an organization, and an admission state. Trust-sensitive set the main directions for distributed reputation aggregation interactions form a directed graph IG = (E, I), where an edge [9], [10], [11]: transitive global trust, feedback weighted by I records a governed interaction from E to E under a a,b a b transaction context, and probabilistic estimates derived from contract or policy object. binary outcomes. All three grew out of open peer communities, Every interaction leaves a record: the request identifier, the and their assumptions show it. In the setting we target, identities participating identities, a contract reference, an outcome class, are registered, interactions are governed by contracts, and a timestamp, and an evidence digest. Outcome classes cover a reputation score feeds directly into a participant’s future successful completion, delay, failure, default, and dispute. For business exposure. every admitted entity, the framework maintains a reputation The attack surface of reputation systems is well catalogued: score Rs i ∈ [0, 1]. collusion, reciprocal inflation, badmouthing, whitewashing, Figure 1 shows how the pieces fit together. The blockchain and identity manipulation [12], with the Sybil attack as the plane stores admission records, interaction state, settlement or canonical demonstration that cheap identity creation can distort dispute evidence, and the update history. The reputation service distributed decisions [13]. Permissioned membership raises reads committed evidence, computes the numerical update, the cost of entry but does not close the identity problem. A and submits a bounded result together with its evidence and participant can still control several valid identities through parameter references. A decision layer can then consume the subsidiaries or affiliated accounts, and governance processes score for credit, delivery, compliance, or any other risk-sensitive themselves can hand out identity assurance of uneven quality. decision. B. Blockchain-Supported Supply-Chain Trust B. Adversary and Fairness Conditions TrustChain combines sensor, buyer, and regulator ratings The evaluation covers three bounded conditions. under time decay [5]; DeTRM links commodity evidence, Under repeated bilateral collusion, a small set of entities agreement fulfilment, authority endorsement, and trust propagation [6]; and a broader line of work integrates blockchain keeps completing successful-looking interactions with the with supply-chain trust management and smart manufacturing same counterparties, trying to pass correlated evidence off [14]. Together, these systems establish that a blockchain can as independent. Under identity multiplicity, one controller operates k regispreserve evidence and automate trust-related state transitions. Product evidence and participant evidence are not the same tered identities. The system receives an identity-confidence thing, though. A sensor stream can support claims about the value ϕi ∈ [0, 1] for each identity as governance input, condition of goods without saying anything about how reliably and R EPUTATION C HAIN makes no attempt to infer ϕi from a counterparty behaves across partners, and a participant behavior. In an operational consortium that value would come can perform flawlessly for one closely connected partner from verifiable assurance evidence: legal-entity registration, while remaining untested everywhere else. R EPUTATION C HAIN the certificate assurance tier, independent organizational verifitherefore attends to the shape of the interaction evidence, not cation, or hardware attestation. Assurance that is set too high weakens the attenuation, which is exactly what the ϕ sweep in just its existence. MeritRank offers a stronger formal treatment of Sybil- Section V-G measures. tolerant graph reputation [7], and BlockRep offers stronger Under sparse-history fairness, honest newcomers simply cryptographic protection for review authenticity and anonymity interact less often than established entities, and the risk is that

TABLE I P OSITIONING OF R EPUTATION C HAIN RELATIVE TO REPRESENTATIVE REPUTATION SYSTEMS . System

Primary setting

Main mechanism

Boundary relative to R EPUTATION C HAIN

TrustChain [5]

IoT food supply chain

DeTRM [6]

Blockchain supply chain

MeritRank [7]

DAO feedback graphs

Sensor, buyer, and regulator evidence with time decay Commodity trust, agreement fulfilment, authority endorsement, and trust propagation Sybil-tolerant graph aggregation and decay

BlockRep [8]

IIoT retail reviews

R EPUTATION C HAIN

Permissioned supply-chain participation

Stronger product and multisource evidence context; weaker focus on repeated-pair discounting and newcomer fairness. Stronger commodity and production-operation modelling; different from bounded participant updates under narrow interaction diversity. Stronger graph-level Sybil theory; not designed around permissioned contractual outcomes and supply-chain participation. Stronger review authenticity and anonymity; does not define dynamic participant trust under sparse history. Targets bounded participant trust updates under concentrated evidence, identity uncertainty, and sparse history.

Purchasing proofs and cryptographic review authentication Repeated-pair discount, diversity, identity confidence, and volume-aware decay

ReputationChain framework architecture Blockchain recording and audit plane

Registered supply-chain entities Governed interaction

Entity Buyer / payer

Entity Seller / payee

Consortium authority Registration and policy

Layer 1: Entity admission and identity binding Admission records

Initial reputation prior

Public-key registration

Identity binding

Layer 2: Smart-contract-governed interaction Contract state

Contract request

Role and obligation definition

Trust-sensitive interaction

Outcome capture

Layer 3: Settlement, dispute, and outcome evidence

Settlement and dispute evidence

Successful completion

Delayed settlement

Dispute handling

Default or failure

Reputation-update history

Bounded bilateral influence

Identity-aware adjustment

Volume-aware decay and final update Trade-credit validation scenario Delayed payment risk Default loss Collateral exposure Virtual security deposit Cash security deposit

Layer 5: Risk-sensitive decision output Default-risk estimation

A successful high-risk event pushes the evidence further above the prior, and a failed high-risk event pushes it further below. One asymmetry runs through the whole design: when reinforcement looks suspicious, only positive evidence is weakened, and adverse evidence always lands at full strength. Without that asymmetry, colluders could turn the defense against itself and use the attenuation to soften the impact of their own failures. B. Repeated-Pair and Diversity Control (m)

Layer 4: Reputation computation and robustness controls Temporal decay

1

(m)

ωi,j =

(m)

1 + η(ci,j − 1)

Trade credit risk and collateral allocation

(m)

(m)

Fig. 1. R EPUTATION C HAIN architecture. Blockchain preserves identity, contract, outcome, and update provenance. Numerical reputation computation is separated from the smart-contract path. The trade-credit block is an example of downstream risk-sensitive use, not a separate trust mechanism.

IV. R EPUTATION C HAIN M ODEL A. Bounded Event Evidence For event m, let Yi ∈ {0, 1} be the verified outcome for entity Ei , and let rm ∈ [0, 1] be the interaction risk. With a

(3)

=

|Ni

|

(m) Ni

,

(4)

the ratio of distinct counterparties to total interactions; a low value signals concentrated evidence. In the collusion scenario, the raw attenuation is (m)

time decay alone pushes them into false low-trust classification without any adverse behavior on their part. Several conditions sit outside the evaluated scope: certificateauthority compromise, identity replacement, on-off attacks, strategic badmouthing, delayed reporting, and production network failures. We list them explicitly because, taken together, they define the current security boundary of the design.

η ≥ 0,

,

which gives the first interaction full weight and every later interaction with the same pair progressively less. Diversity enters through Di

(m)

(2)

Let ci,j count the observed interactions between Ei and counterparty Ej up to event m. The repeated-pair weight is

Positive and negative evidence

Evidence weighting

neutral prior R0 = 0.5, the bounded event evidence is    1 1 (m) (m) ei = clip[0,1] + rm Yi − . 2 2

qi

(m)

(m)

= min{1, ωi,jm Di

}.

(5)

Figure 2 contrasts this with the naive rule, which treats every repeated reciprocal event as if it were independent. C. Identity-Aware Adjustment For identity multiplicity, the attenuation becomes (m)

qi

(m)

= min{1, ϕi Di

}.

(6)

A lower ϕi attenuates positive evidence more strongly. Nothing here detects multiple identities; the mechanism is a controlled weighting applied after governance has already assigned an assurance level.

Proposition 2 (Sublinear repeated-pair influence). For N repeated interactions between the same pair and η > 0,

Collusive reputation inflation and bounded-influence control Naive update rule

Bounded-influence update rule

N X

Low-diversity signal Colluder 1

Colluder 3

Colluder 2

Colluder 1

Colluder 3

Focal entity Reputation target

Focal entity Reputation target

Inflated reputation score

Limited reputation increase

Low diversity, repeated bilateral reinforcement

Discounted repeated interactions Weighted by

Proof sketch. Apply the integral bound for the decreasing function 1/x to the terms after the first. Proposition 3 (Volume-aware attenuation). For κ ≥ 0, volumeaware decay moves a score toward R0 no faster than the corresponding static decay rule.

Fig. 2. Repeated reciprocal evidence can inflate reputation under a naive rule. R EPUTATION C HAIN bounds this reinforcement through repeated-pair weight (m) (m) ωi,j and diversity Di .

Both scenario-specific factors pass through a softening step, (m)

(m)

= 1 − ρsoft (1 − qi

),

0 ≤ ρsoft ≤ 1,

and the adjusted evidence is ( (m) (m) (m) R0 + si (ei − R0 ), ei > R0 , (m) ∆i = (m) (m) ei , ei ≤ R0 .

(11)

Thus raw repeated-pair weight grows at most logarithmically rather than linearly.

Repeated reciprocal interactions should not be treated as independent trust evidence

si

1 1 ≤ 1 + ln(1 + η(N − 1)). 1 + η(n − 1) η n=1

Colluder 2

(7)

Proof sketch. Since γ/(1 + κVi ) ≤ γ, the volume-aware exponential factor is at least the static factor for ∆t ≥ 0. These properties do not amount to attack detection. They establish boundedness and fix the direction of each mechanism; the experiments then measure how far those mechanisms actually move the outcomes. F. Blockchain Realization Boundary

The blockchain stores admission, interaction, outcome, and update records. Every submitted update carries an evidence digest, a parameter-set version, a sequence number, and the The sequential update is score itself, and the smart contracts check authorization,   (m+1) (m) (m) Rs i = clip[0,1] (1 − λ)Rs i + λ∆i , (9) evidence existence, contract state, update order, parameter version, and score bounds. The exponential calculation in (10) where 0 < λ ≤ 1 sets how strongly the score responds to and the event sequence in (9) run in an off-chain reputation current evidence. service. This separation buys two things and costs one. It keeps D. Volume-Aware Decay and Newcomer Fairness platform-specific floating-point behavior and repeated nonlinear A trust score should not sit frozen forever, but decay toward computation out of the smart contracts. The cost is that an zero would treat inactivity as if it were failure. R EPUTATION on-chain admissibility check cannot prove that a submitted C HAIN instead decays toward the neutral prior R0 : value is the unique correct result. Any consortium member   γ∆t can recompute an update from the committed evidence and Rs decay (t + ∆t) = R0 + (Rs i (t) − R0 ) exp − , i 1 + κVi (t) the referenced parameter version, so numerical correctness (10) is auditable rather than cryptographically enforced. Work on where Vi (t) is the verified interaction volume, γ > 0 is the trust as a service argues for the same split between evidence base decay rate, and κ ≥ 0 controls the volume sensitivity. management and reputation computation [16]. Setting κ = 0 recovers static decay. The interpretation needs some care. Higher volume slows V. E VALUATION the drift toward the neutral prior, which preserves established evidence; it hands nothing to newcomers for free. Newcomer A. Questions, Metrics, and Protocol fairness is therefore measured through the false low-trust rate, The evaluation asks three questions. not through any unconditional protection of scores. (8)

E. Formal Properties (m) (m) Proposition 1 (Boundedness). If Rs i ∈ [0, 1], ∆i ∈ [0, 1],

and 0 < λ ≤ 1, then the update in (9) remains in [0, 1]. Proof sketch. The un-clipped term is a convex combination of two bounded values. Clipping preserves the interval.

Q1: Does repeated-pair control reduce collusive reputation gain? • Q2: Does identity-aware weighting reduce the advantage of identity multiplicity? • Q3: Does volume-aware decay reduce false low-trust classification of honest newcomers?

•

Gcoll = Rs C (T ) − Rs C (0), Nmalicious high trust FTER = , Nmalicious Rs adversarial RIR = , Rs honest |{i ∈ Nh : Rs i (T ) < τ }| FLTR = . |Nh |

(12) (13) (14) (15)

0.40 Collusive reputation gain

Four metrics carry the results: collusive gain Gcoll , false trust elevation rate (FTER), reputation inflation ratio (RIR), and false low-trust rate (FLTR):

0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00

B1

B2

B3 Model

B4

B6

An RIR above one means the adversarial mean sits above the honest mean, and FLTR counts honest newcomers incorrectly Fig. 3. Collusive reputation gain. Error bars show standard deviation over 30 seeded runs. left below the threshold τ . The simulator is written in Python and driven by fixed configuration files. Every reported condition aggregates 30 than the raw standard deviations would suggest, which is the seeded runs, and within a scenario all methods consume expected consequence of common random numbers: the seedthe identical generated trace for each seed; this includes the to-seed workload variation cancels inside each difference. Four static and volume-aware newcomer comparison, which shares of the five effects are consistent across all 30 matched runs. traces in the same way. Sensitivity sweeps reuse the same The recovery-burden difference is the exception, and its interval seed identifiers across parameter values. Differences between includes zero. methods therefore cannot be attributed to workload noise. D. Repeated Bilateral Collusion Table II gives the populations, the outcome probabilities, and Table IV reports the collusion results. B6 produces the lowest the model settings needed to reconstruct each scenario. collusive gain at 0.1443, B4 reaches 0.1815, and the simple B. Baselines and Ablations baselines all stay above 0.35. Because B4 keeps diversity and softening active, the gap between B4 and B6 isolates what the Six labels are used across the scenarios. repeated-pair discount adds on its own. • B1 is naive mean evidence without sequential update, The residual FTER of 0.7333 deserves emphasis. Colluders temporal decay, diversity, or identity weighting. still complete successful-looking events, and many of them stay • B2 is a static-decay baseline. above the trust threshold, so nothing in this table amounts to • B3 is a TrustChain-inspired time-decayed multisource colluder classification. The claim the data supports is smaller: proxy [5]. repeated-pair and diversity control shrink the magnitude of the • B4 removes only the repetition-count discount. Diversity inflation. and softening remain active. • B5 removes identity-confidence adjustment. E. Identity Multiplicity • B6 is full R EPUTATION C HAIN . At k = 10, B6 reaches an RIR of 0.8723, while B1, B2, and B3 deserves a caveat. It is not a full TrustChain implemen- B3 remain above 1.08 and B5 lands at 0.9228. The B5 to B6 tation, because independent real sensor, buyer, and regulator difference isolates the identity-confidence weighting. Figure 4 streams were not available; the proxy generates correlated traces the trend over all tested values of k. synthetic components from the same underlying event. Its FTER stays near 0.98 even for B6, so the experiment results should not be read as a general defeat of TrustChain. establishes no Sybil detection at all. What it does establish is What it tests is whether time-decayed multisource aggregation reduced reputation amplification once low identity assurance alone holds up when the sources are correlated and manipulable. is already known, which makes the mechanism a policy C. Seed-Paired Analysis

instrument rather than a detector.

Because every method sees the same trace for a given seed, the mechanism comparisons are naturally paired. The scenario tables report mean ± standard deviation for the raw metrics; Table III reports the mean paired effect with a 95% Student-t confidence interval, together with a one-sided Wilcoxon signedrank test as a distribution-free check on the expected direction.1 Readers should note that the paired intervals are far tighter

F. Newcomer Fairness

1 Paired effects are computed from unrounded run-level values, so they may differ in the last digit from the difference of the rounded means shown in the scenario tables.

Table VI compares static and volume-aware decay on identical newcomer traces. Volume-aware decay lifts mean newcomer reputation from 0.6626 to 0.7589 and cuts FLTR from 0.3633 to 0.1683, while the conditional recovery burden barely moves (from 1.2772 to 1.2590, with a paired interval that includes zero). The benefit is preventive, in other words: fewer honest newcomers get pushed below the threshold in the first place, and recovery after a low-trust classification becomes no easier than before.

TABLE II C ONTROLLED SIMULATION SETTINGS AND WORKLOAD - GENERATION ASSUMPTIONS . I NTERACTION RISK rm IS SAMPLED UNIFORMLY FROM [0.60, 1.00] IN EVERY SCENARIO . Scenario

Population and workload

Outcome and interaction assumptions

Model settings

Global

30 paired seeded runs

Update mixing λ = 0.35

Collusion

100 entities; 10 colluders; 2,000 events over 60 days 80 honest entities; one controller; k ∈ {1, 3, 5, 10}; 1,500 events

Initial and neutral reputation R0 = 0.50; trust threshold τ = 0.65 Collusion-event probability 0.30; honest success 0.88; collusive success 0.98 Base attack probability 0.25, raised by 0.03 per additional identity and capped at 0.80; honest success 0.88; controlled-identity positive outcome 0.99; intra-group fraction 0.60 Daily Poisson interaction rates 4.0, 2.0, and 0.35; honest success 0.90 B3 draws correlated synthetic sensor, buyer, and regulator ratings from the same event, with noise 0.05

Identity multiplicity

Sparse history Comparison models

20 high-volume, 20 medium-volume, and 20 newcomer entities; 40 days Same trace per seed as the full model

TABLE III M EAN PAIRED EFFECTS OVER 30 MATCHED SEEDS . P OSITIVE VALUES FAVOR THE FULL OR VOLUME - AWARE MODEL . Comparison and metric B4 − B6 collusive gain B5 − B6 RIR at k = 10 Volume-aware − static newcomer rep. Static − volume-aware FLTR Static − volume-aware recovery

Mean effect [95% CI] 0.0372 [0.0361, 0.0383] 0.0506 [0.0496, 0.0515] 0.0963 [0.0929, 0.0997] 0.1950 [0.1613, 0.2287] 0.0183 [−0.1186, 0.1552]

Wilcoxon p < 0.001 p < 0.001 p < 0.001 p < 0.001 p = 0.168

η = 0.80; ρsoft = 0.60 ϕ = 0.35; ρsoft = 0.55

γ = 0.10; κ = 0.30; at most 30 recovery events B2 decay 0.003 per event; B3 weights (0.40, 0.40, 0.20) and decay 0.003 per event

TABLE V I DENTITY MULTIPLICITY AT k = 10, MEAN ± STANDARD DEVIATION . Model B1 B2 B3 B5 B6

RIR

FTER

Identity rep.

1.0838 ± 0.0124 1.0000 ± 0 0.8918 ± 0.0042 1.0908 ± 0.0277 0.9933 ± 0.0254 0.8935 ± 0.0171 1.1047 ± 0.0164 1.0000 ± 0 0.9077 ± 0.0049 0.9228 ± 0.0218 0.9833 ± 0.0531 0.7387 ± 0.0123 0.8723 ± 0.0207 0.9800 ± 0.0551 0.6983 ± 0.0111

1.15

Model B1 B2 B3 B4 B6

Gain

FTER

Colluder rep.

0.3688 ± 0.0055 1.0000 ± 0 0.8688 ± 0.0055 0.3585 ± 0.0277 0.9467 ± 0.0776 0.8585 ± 0.0277 0.3806 ± 0.0097 1.0000 ± 0 0.8806 ± 0.0097 0.1815 ± 0.0207 0.8533 ± 0.1167 0.6815 ± 0.0207 0.1443 ± 0.0197 0.7333 ± 0.1539 0.6443 ± 0.0197

1.10 Reputation inflation ratio

TABLE IV R EPEATED BILATERAL COLLUSION , MEAN ± STANDARD DEVIATION .

1.05 1.00 0.95

The one-factor-at-a-time sweeps vary η, κ, and ϕ under common seeds. Table VII lists selected endpoints beside the reference configuration. Three of the boundary values double as a consistency check on the implementation, because each one reduces the full model to a configuration we also ran as a separate code path: η = 0 is B4, ϕ = 1 is B5, and κ = 0 is static decay. The sweep endpoints reproduce those ablation results exactly. Pushing η past 0.80 buys only small further reductions in collusive gain and pays for them with lower honest reputation, a plain robustness-utility trade-off. Raising κ improves newcomer outcomes with diminishing returns, and raising ϕ weakens the attenuation and lets RIR climb, exactly as the model predicts. None of the sweeps points to a universal optimum, so calibration remains an operational task rather than a solved problem.

B5 ablation B6 full

0.90 0.85 0.80

G. Sensitivity and Trade-Offs

Honest parity B1 naive B3 proxy

1

3

5 Controlled identities, k

10

Fig. 4. RIR across controlled identity multiplicity. Values below one remain below the honest mean. B3 is a controlled TrustChain-inspired proxy, not a full reimplementation.

VI. D ISCUSSION AND L IMITATIONS A. What the Results Establish The strongest evidence comes from the matched baselines and ablations. B4 against B6 isolates the repeated-pair discount, B5 against B6 isolates the identity-confidence weighting, and static against volume-aware decay isolates the sparse-history mechanism. Common traces strip out workload noise, the paired intervals show that each effect holds run by run rather than only on average, and the sensitivity sweeps show that the directional findings do not hinge on one exact parameter value. Three bounded statements survive this scrutiny. Repeated bilateral evidence receives diminishing influence and yields lower collusive gain. Governance-supplied identity confidence reduces the advantage of multiplicity. Volume-aware decay

TABLE VI H ONEST NEWCOMER FAIRNESS , MEAN ± STANDARD DEVIATION . Decay

Mean rep.

FLTR

Recovery

Static 0.6626 ± 0.0168 0.3633 ± 0.0909 1.2772 ± 0.1529 Volume-aware 0.7589 ± 0.0221 0.1683 ± 0.0895 1.2590 ± 0.3933

0.9 0.8

TABLE VII S ENSITIVITY SUMMARY. T HE MIDDLE VALUE IS THE MAIN REFERENCE SETTING .

Mean newcomer reputation False low-trust rate

Sweep η: collusive gain η: honest rep. κ: FLTR κ: newcomer rep. ϕ: RIR ϕ: identity rep.

Low or boundary

Reference

High

0.1815 at 0.00 0.7817 0.3633 at 0.00 0.6626 0.8606 at 0.20 0.6889

0.1443 at 0.80 0.7623 0.1683 at 0.30 0.7589 0.8723 at 0.35 0.6983

0.1426 at 1.40 0.7566 0.1533 at 0.80 0.7852 0.9228 at 1.00 0.7387

0.7 Mean value

0.6

then test whether the mechanism generalizes past a single risk scenario. Each of these additions probes something the present study cannot: external validity, implementation cost, and attack coverage.

0.5 0.4 0.3 0.2 0.1 0.0

Static decay

Volume-aware

Fig. 5. Mean newcomer reputation and false low-trust rate. Error bars show standard deviation over 30 runs.

lowers false low-trust classification under the tested sparsehistory workload. B. What the Results Do Not Establish The evaluation is a controlled simulation, not a deployment benchmark: the interaction outcomes and adversarial behavior are synthetic, the parameters are transparent but never calibrated against operational supply-chain records, and the B3 comparison is a proxy. Real seasonal demand, dispute cycles, delayed reporting, and correlated organizational behavior could all shift the absolute reputation trajectories. The attack model is also incomplete. Whitewashing, certificate replacement, on-off attacks, badmouthing, strategic timing, and governance compromise all remain open, and the model assumes that adverse outcomes are captured correctly in the first place. A blockchain protects committed evidence from later alteration; it cannot certify that an external event was honestly observed before submission. Finally, the off-chain computation boundary leaves an integrity dependency. Consortium members can recompute every update, but the smart contract never proves the nonlinear calculation itself. Deterministic fixed-point arithmetic, replicated computation, threshold approval of updates, or succinct proofs could each strengthen this in future designs. C. Future Work The natural next step is a working permissioned-blockchain prototype with measured transaction latency, update cost, and storage overhead, driven by operational or realistically derived interaction traces and used to calibrate η, κ, ϕ, λ, and τ . Adaptive adversaries belong in that evaluation as well, including on-off behavior, whitewashing, badmouthing, identity replacement, and governance that assigns assurance incorrectly. A second application beyond trade credit would

VII. C ONCLUSION We presented R EPUTATION C HAIN, a blockchain-supported participant trust framework for governed supply-chain interactions. The model bounds repeated bilateral reinforcement, folds in governance-supplied identity confidence, and decays scores toward a neutral prior at a volume-aware rate. Across controlled simulations with 30 seeded runs and matched traces, it lowers collusive gain, identity-multiplicity inflation, and false low-trust classification of honest newcomers, and the seed-paired analysis shows those effects are consistent rather than averaged artifacts. The evidence supports a bounded reduction of reputation distortion rather than attack detection, and production use still calls for deployment measurement, calibration on real data, stronger numerical verification, and a broader adversarial evaluation. DATA AVAILABILITY Every value reported here is drawn from a frozen set of audited raw and summary tables, and the result figures are regenerated from those tables by a single plotting script. All five paired effects in Table III reproduce exactly from the retained per-seed records. The simulator, its configuration files, the raw and summary tables, and the plotting script are available at github. D ECLARATION ON THE U SE OF AI T OOLS The AI was used for language editing and for restructuring the prose of this manuscript, and for assembling the LATEX source. The research design, the simulation code, the experimental configurations, the data, and the figures are the authors’ own work and were neither generated nor selected by an AI system. The authors have verified every reported value against the underlying result files and accept full responsibility for the content. R EFERENCES [1] N. Kumar, K. Kumar, A. Aeron, and F. Verre, “Blockchain technology in supply chain management: Innovations, applications, and challenges,” Telematics and Informatics Reports, vol. 18, p. 100204, 2025. [2] Ö. Karaduman and G. Gülhas, “Blockchain-enabled supply chain management: A review of security, traceability, and data integrity amid the evolving systemic demand,” Applied Sciences, vol. 15, no. 9, p. 5168, 2025.

[3] P. L. S. Miguel, A. Lago da Silva, and M. E. Silva, “Unpacking supplier diversity in purchasing and supply management: What do we know and what should we know about?” Journal of Purchasing and Supply Management, vol. 32, no. 2, p. 101078, 2026. [4] M. Jeanneret Medina, C. Baudet, and J.-F. Lebraty, “Blockchain and agency theory in supply chain management: A question of trust,” International Journal of Information Management, vol. 75, p. 102747, 2024. [5] S. Malik, V. Dedeoglu, S. S. Kanhere, and R. Jurdak, “TrustChain: Trust management in blockchain and IoT supported supply chains,” in Proc. IEEE Int. Conf. Blockchain, 2019, pp. 184–193. [6] G. D. Putra, C. Kang, S. S. Kanhere, and J. W.-K. Hong, “DeTRM: Decentralised trust and reputation management for blockchain-based supply chains,” in Proc. IEEE Int. Conf. Blockchain Cryptocurrency (ICBC), 2022, pp. 1–5. [7] B. Nasrulin, G. Ishmaev, and J. Pouwelse, “MeritRank: Sybil tolerant reputation for merit-based tokenomics,” in Proc. 4th Conf. Blockchain Research and Applications for Innovative Networks and Services (BRAINS), 2022, pp. 95–102. [8] W. Zhao, X. Yang, S. Qi, J. Wei, X. Dong, X. Yang, and Y. Qi, “Secure blockchain-based reputation system for IIoT-enabled retail industry with resistance to Sybil attack,” Future Generation Computer Systems, vol. 166, p. 107705, 2025.

[9] S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina, “The EigenTrust algorithm for reputation management in P2P networks,” in Proc. 12th Int. Conf. World Wide Web (WWW), 2003, pp. 640–651. [10] L. Xiong and L. Liu, “PeerTrust: Supporting reputation-based trust for peer-to-peer electronic communities,” IEEE Trans. Knowl. Data Eng., vol. 16, no. 7, pp. 843–857, 2004. [11] A. Jøsang and R. Ismail, “The beta reputation system,” in Proc. 15th Bled Electronic Commerce Conf., 2002, pp. 2502–2511. [12] K. Hoffman, D. Zage, and C. Nita-Rotaru, “A survey of attack and defense techniques for reputation systems,” ACM Comput. Surv., vol. 42, no. 1, pp. 1–31, 2009. [13] J. R. Douceur, “The Sybil attack,” in Peer-to-Peer Systems, ser. Lecture Notes in Computer Science. Springer, 2002, vol. 2429, pp. 251–260. [14] Y. Wu and Y. Zhang, “An integrated framework for blockchain-enabled supply chain trust management towards smart manufacturing,” Advanced Engineering Informatics, vol. 51, p. 101522, 2022. [15] C. Huang, Z. Wang, H. Chen, Q. Hu, Q. Zhang, W. Wang, and X. Guan, “RepChain: A reputation-based secure, fast, and high incentive blockchain system via sharding,” IEEE Internet Things J., vol. 8, no. 6, pp. 4291– 4304, 2021. [16] S. Olariu, R. Mukkamala, and M. Aljohani, “Towards trust and reputation as a service in society 5.0,” Smart Cities, vol. 7, no. 5, pp. 2645–2669, 2024.

Related documents

• Rethinking Collaborative Trust for Verifiably Decentralized Blockchain Systems#321754
• Blockchain Empowered Trustworthy Agent Networks: Foundations, Taxonomy, and Future Directions#429270
• MHOT: Height-Optimized Authenticated Data Structure for Blockchain State Commitment#271813
• Towards a Zero-Trust Supply-Chain Assurance Rubric for ORAN RIC Applications#168261
• Physical synchronisation for anomaly detection in epidemic blockchain consensus#368895