Fair Attribution and Mechanism Design

Essential question: Can visibility incentives reward useful evidence?
Evidence status: Conceptual method, a frozen synthetic credit allocation, status-verified local records, and an inert L07 fixture. No production behavior, legal applicability, conformity, certification, or social welfare was established.

1. Separate citation, support, contribution, ownership, and credit

Attribution is overloaded. A displayed citation is a presentation object. Claim support is a relation between a proposition and evidence. Source influence is a counterfactual or coalition-dependent change in an outcome. Authorship and ownership are legal or institutional relations. Authority is claim-relative competence and provenance. Credit is a rule-governed allocation of attention, reputation, money, or another scarce benefit. These objects can coincide, but none logically entails the others.

A source can be cited and fail to support the attached claim. A source can materially influence an answer without receiving the displayed link. Two sources can state the same proposition, making unique contribution hard to identify. A page can be authoritative for its current product specification and weak evidence for an independent comparison. An allocation method can exactly solve a declared utility function and still be normatively unacceptable to stakeholders omitted from the player set.

Mechanism design begins when actors anticipate the rule. If credit follows citation count, a supplier may increase citable fragments. If it follows measured marginal contribution, suppliers may try to control player identity, coalition composition, or the utility proxy. If it rewards verifiable additions, a supplier may repeat older supported facts, select easy-to-verify claims, or influence the verifier. The research question is therefore two-layered: what does the rule compute, and what behavior does it invite?

The W15 boundary is strict. We can reproduce a fictional allocation and inspect its incentives. We cannot infer hidden production attribution, intellectual ownership, legal entitlement, or a fair distribution across settings. No production effect is measured.

2. Specify a mechanism as an auditable decision system

A credit mechanism needs more than a formula. Record:

1. players: who is eligible, how identity is resolved, and how dependence or syndication is handled;
2. outcome: the answer, claim set, user task, evidence improvement, or other object whose value is allocated;
3. utility: the exact function, scale, evaluator, failure states, and uncertainty;
4. information: what the platform, source, user, and reviewer can observe or privately know;
5. strategies: content changes, publication timing, duplication, appeals, collusion, abstention, or exit;
6. allocation: budget, eligibility gates, weights, rounding, timing, and payment or exposure rule;
7. controls: provenance, verification, dependence detection, harm review, audit, and appeals;
8. updates: versioning, drift monitoring, incident response, and rollback; and
9. claim ceiling: which conclusions the mechanism evidence cannot support.

Call a rule incentive compatible only relative to a defined environment and strategic model. Real actors may possess strategies absent from the analysis. A platform can also be strategic: it selects the player set, utility, judge, visibility budget, and enforcement. A mechanism that disciplines suppliers while leaving platform error or concentration unmeasured is not a complete fairness analysis.

The no-change and no-credit states matter. Evidence that cannot be verified may be withheld rather than forced into a positive allocation. An appeal must preserve the original inputs and decision, not silently rewrite history. When the rule changes, past and current allocations need separate versions because actors may have adapted to the old rule.

3. Map stakeholder objectives, information, and strategies

The minimum W15 stakeholder map contains source creators, platforms, users, publishers or rights holders, affected subjects, independent reviewers, and public authorities. Their interests overlap incompletely.

Information asymmetry is central. A creator knows the cost of producing evidence. A platform observes logs that creators and subjects cannot inspect. A subject may know the real-world harm but lack standing inside the allocation. A reviewer may have public metadata but not licensed clauses. A fair-process claim therefore needs access, explanation, correction, and appeal—not only a numerical split.

4. Credit rules create characteristic gaming incentives

An equal split is transparent but sensitive to player identity. One organization can fragment into many nominal sources, or a syndicated network can look like independent corroboration. Equal credit ignores contribution differences and may pay unsupported repetition.

A displayed-count rule is easy to audit at the interface but rewards salience, repetition, formatting for extractability, or first-position competition. It treats a citation as credit even when the attached claim is unsupported. It also creates a fixed-pool externality: more credit for one source can mean less for another.

A contribution-proportional rule links credit to a declared utility. Its vulnerabilities move upstream. Who selects the utility? Who counts as a player? How are redundant sources handled? Does removal change context position or replacement evidence? A marginal contribution is conditional on the constructed coalition and model; it is not unique authorship or truth.

A safeguarded evidence rule can gate unsupported or harmful records and combine contribution, need, and authority. It may reward less-established useful evidence, but it creates new targets: inflate need, mimic authority, choose claims the verifier favors, capture the reviewer, or suppress dissent by labeling it harmful. Gates can protect rights or become opaque exclusion. Every rejection therefore needs a reason, evidence locator, reviewer, appeal, and refresh.

No rule eliminates strategic behavior. Mechanism quality is evaluated through prediction, test, distributional audit, sensitivity, error burden, and correction—not through a benevolent name.

5. Keep contribution, need, authority, and harm distinct

Contribution asks how much a source changes a declared utility inside a constructed evaluation. Need asks whether an actor or group would otherwise be systematically under-credited or unable to bear evidence-production costs. Authority asks whether a source is fit for a particular claim, not whether it is prestigious in general. Harm asks what adverse effects an allocation or incentivized strategy can impose on users, subjects, competitors, or the evidence ecosystem.

These dimensions are not commensurable by nature. A weighted sum is an accountable policy choice. A non-compensable harm or verification condition should be a gate, not a small negative coefficient that sufficient exposure can offset. Authority should be claim-relative and contestable; otherwise established institutions receive self-reinforcing credit. Need requires evidence and a privacy-preserving process; otherwise the rule invites strategic declarations or intrusive profiling.

Distribution also matters separately from total utility. Report source-level allocations, group totals, zero-credit decisions, concentration, appeal outcomes, false exclusions, and burden of verification. A mechanism can raise an aggregate score while concentrating exposure or transferring review costs to community sources. Conversely, lower concentration does not prove correct support.

The W15 synthetic rule deliberately combines all four dimensions so learners can criticize it. Its formula is not endorsed as a production mechanism.

6. Freeze the synthetic allocation inputs

The fictional EVIDENCE-CREDIT-1.0 case has a 100-credit pool and four source players:

All numbers are authored teaching inputs, not observations. Contribution sums to one. Displayed count sums to twelve. Need, authority, and harm are normalized rubric values whose construct validity is not established. Source D fails verification and exceeds the safeguarded harm threshold of 0.40. That failure applies only to the safeguarded rule; the other deliberately naive rules show what happens without the gate.

Rounding uses cents of credit. Calculate unrounded shares, floor to two decimals, then allocate remaining cents in descending fractional remainder; ties use source ID ascending. This largest-remainder rule conserves exactly 100 credits and prevents display rounding from creating or losing value.

7. Reproduce four credit allocations

Equal rule. Divide 100 by four: A 25.00, B 25.00, C 25.00, D 25.00.

Displayed-count rule. Allocate in proportion to (4,2,1,5). After largest-remainder rounding: A 33.33, B 16.67, C 8.33, D 41.67. The commercial repeater receives the largest share because it has the most displayed instances, despite verification failure in another record.

Contribution rule. Contribution already sums to one: A 42.00, B 28.00, C 18.00, D 12.00. This rewards the declared utility contribution, but its validity depends on the player set, evaluator, and coalition construction.

Safeguarded evidence rule. Eligibility requires verification pass and harm risk no greater than 0.40. D receives zero. For survivors, raw policy score is

Scores are A 0.420, B 0.505, and C 0.490, summing to 1.415. Normalizing to 100 and rounding gives A 29.68, B 35.69, C 34.63, D 0.00.

This last allocation can be defended as rewarding verified contribution while giving weight to need and claim-relative authority. It can also be criticized: weights are normative, the rubrics can be gamed, D’s zero is severe, the harm threshold is contestable, and gate errors may burden legitimate sources. The correct output is an allocation plus a challenge log, not “the fair answer.”

8. Audit distribution, concentration, sensitivity, and appeal

Combine B and C as the fictional community-source group and A and D as established/commercial for one transparent distribution check. Community credit is 50.00 under equal, 25.00 under displayed count, 46.00 under contribution, and 70.32 under safeguarded evidence. This comparison shows who benefits from each rule, but the group definition is itself a policy object and cannot be generalized.

Credit concentration, calculated from the unrounded normalized rule weights as the sum of squared shares, is 0.250000 for equal, about 0.319444 for displayed count, 0.301600 for contribution, and about 0.335389 for safeguarded evidence. The safeguarded rule produces the highest concentration because D is excluded, even while its surviving credits are relatively balanced. A fairness memo should preserve this apparent tension instead of choosing only a favorable statistic.

Sensitivity asks which assumptions reverse the allocation: contribution/need/authority weights; harm threshold; verification error; player dependence; group identity; rounding; and total budget. If D passes on appeal, its raw safeguarded score would be 0.210, changing every share. That is a new decision version. The appeal record must state evidence, reviewer, whether the gate changed, and the recalculated distribution.

A deployment study would also estimate false exclusion, false credit, reviewer disagreement, burden by source class, strategic adaptation over time, user or subject harm, and platform discretion. The synthetic matrix identifies none of those quantities.

An error budget should disaggregate mistakes by consequence. A false-positive credit decision rewards evidence that should have failed, while a false-negative gate withholds credit from a legitimate source. Wrong-source attachment credits the wrong identity even when the underlying proposition is supported. Dependence error counts syndicated evidence as independent. A stale authority record can preserve credit after a source or claim has changed. Report rates, materiality, affected groups, and correction time separately; one net accuracy score can hide asymmetric harm.

Appeal design is part of allocation, not customer service added later. The source needs notice of the decision, a bounded explanation of evidence and rule version, a way to submit corrections, an independent reviewer where stakes warrant, a response window, and a preserved disposition. Users and affected subjects need a route to challenge unsupported or harmful credit even when they are not players. Rights holders need a channel for identity, license, or syndication disputes. The platform must publish how an appeal changes future and past allocations without erasing the original record.

Mechanism updates also create temporal distribution. If a new verifier raises the cost of participation, historically credited incumbents may retain accumulated visibility while smaller sources face the new burden. Backward recalculation, prospective change, and transition assistance have different winners. Freeze the effective date and version. Monitor entry, exit, appeal access, and review latency, not only the selected credit vector.

9. Read PAPER routes through their estimands and ceilings

PAPER-17 is the core fair-context attribution route. It makes player and utility assumptions explicit and offers an efficient exact allocation for a particular decomposable max-sum utility. That exactness is valuable computational evidence within its construction. It does not prove authorship, ownership, truth, actual hidden model use, or universal fairness. Changing the utility or player set changes the meaning of credit.

PAPER-42 is the core mechanism-design route. It studies strategic supplier–platform interaction and combines suspicious-rewrite restraint with rewards for earlier-version-verifiable content under a custom defense/welfare objective. Its bounded simulation motivates incentive and ablation questions. “Win-win” remains conditional on the authors’ actors, weights, verification, fixed candidates, horizon, judges, and equivalence convention. It is not legal fairness, social welfare, a global equilibrium, or a production guarantee.

PAPER-36 is the extension. It measures first displayed citation choice when exactly two anonymized sources are supplied in a controlled context. Its factorial discipline is useful for strategy and positional-bias discussion. It does not test crawling, retrieval, larger slates, claim support, total attribution, user behavior, or market welfare. A first-citation gain for one source is also an opportunity loss for the other in that two-source setting.

Together the papers motivate explicit utility, strategic response, controlled comparison, and distributional reporting. None validates the W15 allocation values.

10. Classify governance instruments by status before applicability

Institutional status changes what an identity can support:

Never rank these classes as stronger-to-weaker evidence in the abstract. A technical specification may be authoritative for a byte-level format and irrelevant to fairness. A law may establish an obligation and say nothing about whether a metric is scientifically valid. Status determines the kind of inference available; applicability determines whether that object matters to a release.

11. Use S01–S10 as a status-first crosswalk

At the 24 August 2026 course cutoff:

• S01 NIST AI RMF is a U.S. government voluntary framework, not law or certification.
• S02 NIST AI 600-1 is a government technical report and voluntary GenAI guidance, not evidence of implementation.
• S03 ISO/IEC 42001:2023 is a published international AI management-system standard; its public overview does not supply paywalled requirements or conformity evidence.
• S04 C2PA Content Credentials 2.4 is an industry technical specification for provenance manifests; provenance integrity is not truth, authorship, or ranking evidence.
• S05 the PRC Interim Measures are an official regulatory text with a defined public-service and jurisdictional scope; operational interpretation needs qualified legal review.
• S06 the PRC AI-generated-content labeling measures are a separate official regulatory measure with specified dates and roles; they are not interchangeable with S07.
• S07 GB 45438-2025 is a mandatory Chinese national standard for specified labeling methods within scope; mandatory status does not prove applicability or content truth.
• S08 GB/T 45654-2025 is a recommended Chinese national standard for basic security requirements of generative AI services, not law or a GEO-performance rule.
• S09 GB/T 45652-2025 is a recommended Chinese national standard concerning pre-training and fine-tuning data security, not a content-ranking rule.
• S10 GB/T 35273-2020 is a recommended Chinese national personal-information security specification with a revision process noted at the cutoff; privacy/legal review remains necessary.

The crosswalk is dated. Recheck official identity pages, status, amendments, and effective dates before release. Do not infer a clause from a title, secondary summary, or metadata page.

12. Complete the applicability card and preserve the L07 block

Every applicability card records: issuer and identity; status; jurisdiction; publication, effective, access, and review dates; covered actors; system/content and lifecycle scope; institutional force; accessible official text or licensed clause; mapped control; implementation evidence; accountable reviewer; uncertainty, exclusions, and refresh date. The compact W15 card requires at least the user-requested fields: jurisdiction, date, actors, scope, force, evidence, reviewer, and uncertainty.

These fields form a chain, not synonyms. Identity/status can be verified while applicability is unresolved. Applicability can be judged while implementation evidence is missing. A control can exist while its test fails. Conformity can require a separate qualified assessment. Certification, where meaningful, is a further institutional act. The course claims none of them.

L07 makes the release consequence concrete. It runs six inert synthetic cases. Three are blocked, two detected, and one escaped. AB-05 loses an accessible claim-to-source relation; residual severity is high and likelihood likely, producing score nine. The deterministic audit status is PASS because the review worked, while the release decision is BLOCKED. This distinction is essential: successful risk detection is not permission to release.

L07 also detects AB-06, where voluntary guidance is presented as mandatory law. Its claim ceiling is: synthetic control tests and status vocabulary audit only; no safety guarantee, certification, compliance, or legal advice. W15 inherits that ceiling.

The final bounded conclusion is:

In the frozen synthetic allocation, four transparent rules distribute the same 100-credit pool differently and create different gaming and distributional incentives. The safeguarded rule rewards three verified sources under declared contribution, need, authority, and harm choices, but is not proven fair. S01–S10 retain distinct statuses, and applicability remains a qualified, dated review. L07 passes its audit while blocking release because an accessibility-and-attribution control escaped.

A validator PASS reproduces these local identities and calculations. It does not supply legal advice, paywalled clauses, certification, conformity, fairness, safety, or a production attribution effect.