Decentralised News · Evidence reviewed 12 November 2026 · Reputation portability and assessment scope
What Matters
Reputation can travel between systems only if the receiving party understands what was assessed, which version performed the work, who supplied the evidence and what remains unresolved. Copying a score does not transfer its meaning. DN now uses an evidence record instead of uncalibrated trust bands, keeping identity, past outcomes and permission for the next action separate.
DN Evidence Block
Documented: ERC-8004's draft feedback interfaces, A2A discovery and credential data models. Original analysis: evidence continuity, reviewer independence and context transfer. Not measured: named-agent rankings, Sybil probabilities, a universal cross-protocol reputation score or creditworthiness. The earlier weighted sliders, penalty points and “High-Assurance” band have been withdrawn because they were not calibrated against outcomes.
Why the Same Rating Can Mean Different Things
An agent's reported history can combine tasks with different consequences. Correctly summarizing a public page is not the same outcome as making an authorized payment, delivering usable code or resolving a disputed purchase. A single average hides the task, the value exposed and how success was determined.
Consider a constructed history of 100 successful tasks. If one operator created both the tasks and the reviews, it supports a different inference from 100 independently requested tasks whose outputs were checked against a declared acceptance rule. Neither record becomes independent merely because its events are on-chain. Keep the task identifier, relevant output, acceptance criterion, time and reviewer reference together.
If privacy requires redaction, record what is withheld and who can inspect it. Public availability is not the only route to credible evidence. An opaque aggregate with no defined review process is the harder problem.
What ERC-8004 Standardizes
ERC-8004 remained a Draft when reviewed. It proposes identity, reputation and validation registries. Its feedback interface includes a signed numeric value with decimal precision and optional tags and references. Its summary query requires specified reviewer addresses and warns about unfiltered feedback in the presence of Sybil and spam activity. More elaborate aggregation can occur outside the registry. [1]
A portable feedback value therefore needs its type, units, sign convention, reviewer selection and revocation treatment. “95” might mean a percentage, a marketplace score or another metric. Two registries storing the same number do not establish comparable performance. This article does not certify an interoperable reputation-oracle network or measured provider adoption.
Use an Evidence Vector, Not an Unvalidated Trust Total
| Dimension | Question | What cannot compensate for it |
|---|---|---|
| Task provenance | Can the work be connected to a request and defined outcome? | A large count with no acceptance rule. |
| Outcome validation | Who checked the result and what did the check cover? | A completion flag that only indicates the software stopped. |
| Reviewer independence | What supports separation of interests, and what remains unknown? | Multiple addresses treated as independent people. |
| Continuity | Which owner, keys and behavioral version produced the history? | A familiar name after an undocumented change. |
| Economic exposure | What was at stake and how was it valued? | Nominal volume without examining circular flows. |
| Recency and relevance | Does the evidence describe the current system and task? | Old results from a different permission scope. |
| Disputes and recovery | Which failures, corrections and pending cases are included? | Successful tasks after failures were excluded. |
These dimensions are not additive components of a probability. An unresolved delegation does not become acceptable because six other boxes have references. Identify the specific missing evidence and the decision it blocks.
Portability Is a Translation of Claims
Before moving a reputation record, retain the source schema and revision, registry or service, agent identifier, chain where applicable and metric meaning. Preserve cohort dates, eligibility rules, incentives, exclusions and the aggregation policy. Do not overwrite a signed source claim merely to fit another platform's field names.
Then record the receiving task and acceptance rule. A procurement platform may care about delivery conformity and disputes; a trading workflow may require execution evidence and different enforced authority. The source record can inform the review without determining permission. If the mapping loses material information, label it incomplete.
For example, a source marketplace may mark a task successful when a seller submits a file. A receiving business may require that file to pass validation and security review. A source success count is not yet a receiving success count. Missing validation is a data gap, not a numeric discount that DN can invent.
Identity Continuity Needs an Event History
A key rotation can preserve history when evidence connects the old and new control states. An ownership transfer, model replacement, new tool access or larger financial permissions can change its relevance. Keep original events and a dated change record; do not erase incidents or silently inherit approval.
A behavioral-version label is itself a declaration until supported by the relevant change-management or attestation process. Record assessment scope: one deployment's review does not automatically cover another region, wallet policy or dependent service. A material change can require reassessment even when the public agent name remains unchanged.
The DN Agent Passport organizes declarations and evidence references. A2A Agent Cards describe discovery, capabilities and authentication-related metadata; the specification also defines signed cards. Those features do not independently establish successful task history. [2]
Reviewer Clusters Are Leads, Not Ownership Verdicts
Shared funding, synchronized activity and reciprocal reviews can justify inspection. They do not alone prove common ownership, agent control, collusion or fraud. An exchange withdrawal system can fund unrelated users, and a legitimate campaign can produce correlated timing. State the observed relationship before asserting an explanation.
Distinguish a confirmed related party, a disclosed incentive, a suspected relationship and an unknown relationship. Document the basis for labels and provide a correction path. Keep independently validated evidence separate from reviews whose independence is unresolved; do not automatically call the latter malicious.
This limits the proposed swarm and DAO-astroturfing instruments too. Wallet activity and voting power are not a labelled population of humans or agents. A detector needs authorized data, independently established labels, false-positive and false-negative analysis, abstention and challenge procedures. This article does not publish a Sybil score, identify anonymous operators or attribute legal entities from transaction patterns.
Incentives, Failures and Large Exposures
Disclose paid reviews, referral incentives, related-party tasks and activity rewards. Treatment depends on the question and data; there is no universally justified fifteen-point penalty. A sensitivity analysis can separate those observations where meaningful, but its exclusions and denominators must remain visible.
Thousands of low-consequence successes should not conceal an unresolved incident affecting a different risk. Record severity, scope and remediation evidence. A refund can resolve a financial amount while leaving a confidentiality incident unresolved.
Higher exposure can justify stronger evidence and tighter permissions, but DN cannot derive a safe dollar limit from completed tasks. Credit, insurance and regulated decisions have additional contractual and legal requirements. A reputation record can be one input; it does not approve the activity or make software a legal borrower.
Credentials and Authority Remain Separate
W3C Verifiable Credentials 2.0 supplies a model for issuer claims and their exchange. Credential verification does not make every claim relevant to every task or replace the verifier's policy. Record issuer, subject binding, scope, time and status requirements. [3]
A claimed audit, licence or policy needs the actual covered entity, activity, version and period. Its title is not its coverage. Permissions to spend, disclose data or delegate work must remain independently enforced through limits, scoped access, revocation and escalation.
Use the DN Reputation Portability Record
The manual JSON record below preserves source meaning, receiving context and the seven evidence dimensions. A recorded reference means the reviewer entered it; the schema does not fetch it, check its signature or authenticate its author. Unknown and issue states remain explicit.
Ask a receiving reviewer to identify the decision the record supports and any additional evidence required. Retain earlier review versions when the evidence or system changes. The useful product is an auditable claim trail, not a grade whose formula conceals gaps.
Research Needed Before a Predictive Score
A predictive model needs a predeclared outcome, representative labelled observations, temporal separation between training and evaluation, and tests across relevant tasks and deployments. Report calibration, uncertainty, errors and abstentions. Testing on the same tasks used to choose weights does not establish future reliability.
Publish collection rules and raw dimensions so readers can assess scope. If a simple baseline performs as well on held-out outcomes, that result should inform the design. This review supplies no such validation and therefore no predictive trust bands.
Relevant DN Research
Open DN Pathfinder to compare infrastructure according to the task and constraints. Related pages may contain separately disclosed commercial links; they do not determine these evidence rules.
Personhood, Wallet History and Sybil Screening: Six Different Evidence Questions
Decision extension · 13 November 2026. An identity credential, an old wallet and an active address count can each be useful without establishing the same fact. Before using any of them to grant access, distribute rewards or describe adoption, specify the population, current controller and decision the evidence supports. This extension adds a screening and enrollment-cost protocol to the existing reputation record; it does not publish a measured fraud rate.
1. What does the credential actually assert?
The W3C Verifiable Credentials Data Model distinguishes issuers, holders, subjects and verifiers. It also separates verification from the truth of the underlying claim. A valid signature identifies an assertion under the scheme; the relying party still needs an acceptance policy for the issuer and the asserted property. Record validity periods and status information where the credential design supplies them. W3C credential standard.
Consequently, an attestation of account control is not automatically an attestation of legal identity, unique personhood, professional qualification or authority to spend. Record the issuer's evidence process and the obligation, if any, it accepts. Whether an issuer has legal liability is a jurisdiction- and contract-specific question, not something the signature algorithm decides.
2. Unique within which system?
World describes World ID as a way to prove unique humanity without disclosing a name or other identifying information to the relying service. That is the project's description of its intended functionality, not a DN measurement of false enrollment, recovery security or coverage. The relevant procurement question is what the chosen credential and application actually require. World's current FAQ.
Do not generalize one biometric enrollment architecture to every zero-knowledge identity system. A proof mechanism, the evidence collected during enrollment and the attributes disclosed during use are separate choices. Ask how duplicate enrollment, a lost device, an account-control change and credential revocation are handled. A proof associated with an enrolled person does not establish that every downstream action was manually performed by that person.
3. Does the history still belong to the relevant controller?
An address can retain its transaction history while the people or systems able to act through it change. For a smart account, inspect relevant owner, module and recovery changes. For a key-based account, public history may not reveal a private change in control. A longstanding identifier is therefore evidence of identifier age, not conclusive evidence of an unchanged counterparty.
Use a continuity file: previous and current controller assertions, the authority change, the date, the evidence supporting the association and the history that remains relevant to the present decision. A legitimate organizational handover can preserve useful records without making the new operator responsible for every historical promise. Do not facilitate the purchase of someone else's verified account or assume that such a purchase transfers its eligibility.
4. What error does an anti-abuse test make?
NIST SP 800-63A-4 describes identity resolution, evidence and attribute validation, verification, enrollment and fraud mitigation for natural persons. It also addresses privacy and usability. DN uses these distinctions to design a defensive review; the publication is not a universal exchange regulation or an agent-identification standard. NIST identity-proofing guidance.
A screening test needs independently established labels. Preserve genuine applicants incorrectly rejected, abusive applicants incorrectly admitted, unresolved cases and the appeals outcome. A cluster of similar transactions may justify investigation but does not conclusively identify AI control or multiple accounts belonging to one person. Nor does an airdrop's eligibility rule necessarily prohibit all automation.
Fictional test: a predeclared sample contains 80 eligible people and 20 ineligible cases. The system rejects ten eligible people and admits five ineligible cases. Its false-rejection rate is 10/80 = 12.5%; its false-acceptance rate is 5/20 = 25%. Of the 75 admitted cases, five are ineligible, a different ratio of 5/75 = 6.7%. Mixing those denominators can make the same result look much better or worse. These are invented counts, not observations about any provider.
Run testing only with permission, consented participants or authorized synthetic fixtures. Do not generate forged identity documents, acquire accounts or test a live service's defenses without authorization. Include accessibility, recovery and appeal paths in the acceptance criteria rather than optimizing one headline rejection percentage.
5. What does a valid enrollment cost?
Count the costs of the complete cohort: acquisition, incentives actually paid, operator time, hardware allocation, verification, failed attempts, support, appeals and replacement credentials. Then state the denominator. Attempted enrollments, issued credentials and accepted unique eligible participants are not interchangeable business outcomes.
Illustrative cohort arithmetic: 1,000 currency units of included cost divided by 70 independently accepted eligible enrollments is about 14.29 per enrollment. Dividing by 75 admitted cases from the screening example would hide five false acceptances. Token face value, a recipient's realizable proceeds and the issuer's cash expenditure should be recorded separately; none is automatically the other.
This model identifies a cost boundary, not the economic value of a verified person. Repeat enrollment, retention, useful transactions and lawful customer demand require their own evidence. A large enrollment count does not prove a profitable identity network.
6. What can activity tell us about adoption?
Define whether the metric counts addresses, wallets, accounts, controllers, customers or accepted tasks. One user can control several addresses; one address can represent an organization or service used by many people. Automated activity can provide real utility. Calling every bot a fake user is as unsupported as calling every active address a person.
Report the observed metric with its period, chain coverage, deduplication method and exclusions. Separate paid customer use from internal testing, subsidized tasks and uncertain classifications where evidence permits. A correlation between incentives and activity does not by itself establish a founder's intent or prove that all additional usage is artificial.
| Decision | Record in the existing evidence file | Keep unresolved |
|---|---|---|
| Accept a credential | Issuer, exact assertion, validity/status and relying-party policy | Unsupported truth or liability claims |
| Trust historical performance | Controller and version continuity | Unobserved changes in control |
| Approve eligibility | Defined rule, authorized test and independent labels | Cases without reliable ground truth |
| Evaluate screening | False accepts, false rejects, unknowns and appeals | A universal AI-detection probability |
| Budget enrollment | Complete cohort cost and accepted eligible denominator | Credential face value as customer value |
| Describe adoption | Actual counting unit, period and attribution method | Human counts inferred from addresses alone |
Use the existing downloadable reputation record to retain the subject, task, source, version and review references. Attach this matrix to the same controlled file rather than collapsing its questions into an overall trust score. A filled field records a statement, not an independently authenticated conclusion.
Method and limits: documentary review, a proposed defensive evaluation protocol and fictional arithmetic. No identity provider, airdrop cohort, biometric system or network adoption population was tested. No account purchase, screening evasion or unrestricted data collection is recommended.
Frequently Asked Questions
Why was the reputation score removed?
The weights, penalty points and high-assurance labels were not calibrated against independent outcomes. The replacement identifies evidence gaps without supplying a safety or credit rating.
Can reputation travel without a universal score?
Yes. Preserve the source claim, schema, units, cohort and version, then document the receiving interpretation and missing information.
Does a reviewer cluster prove a Sybil attack?
No. Correlated behavior can be an investigation lead; ownership and intent require additional evidence.
Methodology, Limits and Change Log
12 November 2026: replaced the uncalibrated slider, trust bands, arbitrary penalties and dollar-based suitability table with the DN Reputation Portability Record. Preserved task evidence, reviewer independence, continuity, incentives, incidents and contextual authority. Corrected cluster language implying ownership or collusion without ground truth. No agent was tested or ranked.
Corrections should identify the claim, source version and supporting reference. Do not send private keys, credentials or personal identifiers. The record documents evidence; it does not provide financial, credit, cybersecurity or legal authorization.
Machine-readable Reputation Portability Record
Schema urn:dn:reputation-portability:v1, version 1.0.0. Copy the blank JSON below and retain non-sensitive references for the defined review. All unknowns remain explicit. Change basis to self_reported_review and supply a real calendar review date and subject when completing it. The schema checks structure only; it does not fetch or authenticate a reference. No API or live dataset is implied.
| Area key | Meaning |
|---|---|
sourceMeaning | Source registry/service, schema, identifiers, metric units, cohort and revocation handling. |
receivingContext | Destination task, acceptance rule, permissions and mapping limitations. |
taskProvenance | Requested work and defined accepted outcome. |
outcomeValidation | Reviewer, test scope and validation evidence. |
reviewerIndependence | Confirmed, suspected and unknown relationships and incentives. |
identityContinuity | Principal, chain/registry, keys and behavioral-version changes. |
economicExposure | Actual value or consequence and valuation basis. |
recencyAndRelevance | Observation dates and applicability to the current version and task. |
disputesAndRecovery | Failures, incidents, corrections, refunds and pending cases. |
Each area has status (unknown, reference_recorded or issue_recorded), reference and notes. A recorded reference or issue requires a non-empty reference. Unknown means unresolved; it is not an observed zero or a failure finding. Preserve private source documents in their authorized system rather than embedding account numbers or credentials.
Copy blank record JSON
{
"recordType": "reputation-portability",
"schemaVersion": "1.0.0",
"basis": "blank_template",
"reviewedDate": null,
"subject": null,
"areas": {
"sourceMeaning": {
"status": "unknown",
"reference": null,
"notes": null
},
"receivingContext": {
"status": "unknown",
"reference": null,
"notes": null
},
"taskProvenance": {
"status": "unknown",
"reference": null,
"notes": null
},
"outcomeValidation": {
"status": "unknown",
"reference": null,
"notes": null
},
"reviewerIndependence": {
"status": "unknown",
"reference": null,
"notes": null
},
"identityContinuity": {
"status": "unknown",
"reference": null,
"notes": null
},
"economicExposure": {
"status": "unknown",
"reference": null,
"notes": null
},
"recencyAndRelevance": {
"status": "unknown",
"reference": null,
"notes": null
},
"disputesAndRecovery": {
"status": "unknown",
"reference": null,
"notes": null
}
},
"boundary": "This record does not independently verify evidence, measure performance or authorize an action."
}
Documented JSON Schema
{
"$schema": "https://json-schema.org/draft/2020-12/schema",
"$id": "urn:dn:reputation-portability:v1",
"title": "Reputation Portability Record",
"description": "An entered evidence record. Schema validation checks structure, not authenticity, legal permission or physical or financial outcomes.",
"type": "object",
"additionalProperties": false,
"required": [
"recordType",
"schemaVersion",
"basis",
"reviewedDate",
"subject",
"areas",
"boundary"
],
"properties": {
"recordType": {
"const": "reputation-portability"
},
"schemaVersion": {
"const": "1.0.0"
},
"basis": {
"enum": [
"blank_template",
"self_reported_review"
]
},
"reviewedDate": {
"type": [
"string",
"null"
],
"pattern": "^\\d{4}-\\d{2}-\\d{2}$",
"format": "date"
},
"subject": {
"type": [
"string",
"null"
],
"maxLength": 240
},
"areas": {
"type": "object",
"additionalProperties": false,
"required": [
"sourceMeaning",
"receivingContext",
"taskProvenance",
"outcomeValidation",
"reviewerIndependence",
"identityContinuity",
"economicExposure",
"recencyAndRelevance",
"disputesAndRecovery"
],
"properties": {
"sourceMeaning": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"receivingContext": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"taskProvenance": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"outcomeValidation": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"reviewerIndependence": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"identityContinuity": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"economicExposure": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"recencyAndRelevance": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
},
"disputesAndRecovery": {
"type": "object",
"additionalProperties": false,
"required": [
"status",
"reference",
"notes"
],
"properties": {
"status": {
"enum": [
"unknown",
"reference_recorded",
"issue_recorded"
]
},
"reference": {
"type": [
"string",
"null"
],
"maxLength": 1000
},
"notes": {
"type": [
"string",
"null"
],
"maxLength": 2000
}
},
"allOf": [
{
"if": {
"properties": {
"status": {
"enum": [
"reference_recorded",
"issue_recorded"
]
}
}
},
"then": {
"properties": {
"reference": {
"type": "string",
"minLength": 1
}
}
}
}
]
}
}
},
"boundary": {
"const": "This record does not independently verify evidence, measure performance or authorize an action."
}
},
"allOf": [
{
"if": {
"properties": {
"basis": {
"const": "self_reported_review"
}
}
},
"then": {
"properties": {
"reviewedDate": {
"type": "string"
},
"subject": {
"type": "string",
"minLength": 1
}
}
}
}
]
}
Related reading:
AI Agent Identity in 2027: Which Layers Does Your Agent Need?
The Best AI Agent Runtimes of 2027: OpenAI vs Google vs AWS vs Microsoft
How to Verify an AI Trading Bot Before Risking Any Money
AI Context Efficiency 2027: Useful Outcomes per Token
AI Model Routing 2027: Does It Save Money Without Losing Quality?
The Agentic Treasury Benchmark: Payments, Stablecoins and Financial Control




