# W15 No-Video Transcript — Fair Attribution and Mechanism Design

## Status and use

No W15 recording exists. This is an authored no-video equivalent, not a transcription. W15 has not been taught, rehearsed, timed, recorded, or piloted. Durations are instructional planning budgets, not observed delivery or learning evidence. The allocation values and code-of-practice card are synthetic; the standards records are dated identity/status routes; L07 is an inert local fixture.

## Planned chapter budget

| Chapter | Topic | Planned minutes |
|---:|---|---:|
| 1 | What attribution can mean | 4 |
| 2 | Mechanism and stakeholders | 4 |
| 3 | Information and strategic response | 4 |
| 4 | Four fairness dimensions | 4 |
| 5 | Equal and displayed-count rules | 4 |
| 6 | Contribution and safeguarded rules | 4 |
| 7 | Distribution, concentration, and appeal | 4 |
| 8 | PAPER evidence routes | 4 |
| 9 | Instrument status distinctions | 4 |
| 10 | S01–S10 and applicability cards | 4 |
| 11 | L07 and bounded decision | 2 |

**Total planned route: 42 minutes.** Reading, calculations, discussion, pauses, and the 80-minute studio are outside this narration budget.

## Chapter 1

Welcome to W15, Fair Attribution and Mechanism Design. Our essential question is: can visibility incentives reward useful evidence? Before trying to answer, we need to slow down the word attribution.

A displayed citation is something an interface presents. It may resolve to a source, but it does not prove that the attached claim is supported. Claim support is a relation between a proposition and evidence. Conditional contribution asks how an outcome changes when a source is removed or when coalitions vary under a declared utility. Authorship and ownership are different legal and institutional relations. Authority is fitness for a particular claim. Credit is the benefit the rule allocates: attention, reputation, money, or something else.

One event does not entail the others. A source can be cited beside a claim it does not support. A source can influence an answer and receive no visible link. Two sources can repeat the same evidence, making unique contribution ambiguous. An official manufacturer can be authoritative for a current specification and weak evidence for an independent comparison. A mathematically exact contribution allocation can still omit affected subjects or rely on a contested utility.

This separation matters because credit changes behavior. If visible citation count is rewarded, sources may produce more citable fragments. If marginal contribution is rewarded, actors may contest player identity, coalition construction, and utility. If verifiable additions are rewarded, suppliers may prefer easy-to-check facts or try to influence the verifier. Some adaptation can improve evidence; some can game the rule.

The W15 task is therefore not “choose the fair score.” It is: define the mechanism, predict strategic response, reproduce a synthetic allocation, report distribution and burden, classify governance sources by status, and preserve every unresolved applicability question.

We will also keep three layers of evidence apart. The synthetic layer lets us reproduce arithmetic. The paper layer supplies bounded methodological examples with their own estimands. The governance layer verifies institutional identity and status at a date. Perfect arithmetic cannot validate a paper, and an official status page cannot decide the ethics of an allocation. This separation is the quality-control backbone of the lesson.

Retrieval check: imagine a sentence with one displayed source, a supported numerical clause, and an unsupported certification clause. Which objects exist? A citation exists. One support edge exists. The certification support edge fails. No conclusion about authorship, ownership, contribution, or credit follows without additional rules.

## Chapter 2

A mechanism is an auditable decision system. It has at least nine parts. First, the players: who is eligible, how identity is resolved, and whether syndicated or dependent sources are merged. Second, the outcome whose value is allocated. Third, the utility and evaluator. Fourth, information: what each actor observes or privately knows. Fifth, available strategies. Sixth, the allocation budget, formula, gates, timing, and rounding. Seventh, controls and appeals. Eighth, update, drift, incident, and rollback rules. Ninth, the claim ceiling.

Now map stakeholders. Source creators may seek attribution, payment, correction, or low evidence-production cost. Platforms may seek answer usefulness, trust, engagement, low computation, and manageable legal or safety risk. Users may need accurate, diverse, accessible, and timely evidence. Publishers and rights holders may care about license, context, traffic, and revenue. Affected subjects may care about dignity, privacy, non-discrimination, and correction even though they receive no credit. Reviewers may value evidence quality and independence. Public authorities and standards bodies act within different institutional mandates.

These objectives do not collapse into one welfare number by themselves. If a platform optimizes engagement, a creator optimizes exposure, and an affected subject bears misrepresentation harm, a positive platform–creator score does not prove a social win.

Decision rights also matter. The platform often defines players, utility, judge, and credit budget. That is institutional power. A mechanism that audits supplier behavior but ignores platform discretion is incomplete. An appeal process should expose enough evidence to contest the decision while protecting personal data, security, and lawful confidentiality.

A valid no-credit state is essential. When evidence cannot be verified, the rule should be able to withhold allocation rather than fabricate a positive share. A no-deployment state is similarly necessary. In W15, the synthetic allocation advances at most to independent review.

Your check is to write one row for an affected subject: objective, information, strategy or remedy, and possible burden. Notice that this actor may have no direct strategy inside the credit formula. That omission is itself distributional evidence.

## Chapter 3

Strategic response follows information asymmetry. Creators know their own effort and source lineage. Platforms know logs, candidate sets, and internal rule details. Users see outputs but not all alternatives. Affected subjects may know a real-world error or harm that the platform’s proxy misses. Reviewers may see only selected artifacts or public metadata.

Each credit rule creates a different attack surface. An equal split is sensitive to identity fragmentation: one origin can appear as several sources, or syndication can look like independent contribution. A displayed-count rule rewards repetition, salience, position, and extraction-friendly fragments, whether or not the attached claim is supported.

A contribution rule moves gaming into player and utility definitions. If a source is removed, what replaces it? Does position change? Does redundancy make contribution look small despite social value? Who controls the evaluation model? Exact computation does not validate those choices.

A safeguarded rule introduces verification, need, authority, and harm judgments. Actors may inflate need, mimic authority, choose verifier-friendly claims, capture reviewers, or contest harm labels. Gates can protect people; they can also become opaque exclusion. Every failed gate needs a reason, evidence locator, accountable reviewer, appeal path, and refresh.

The platform can strategically alter any rule by changing the source pool, exposure budget, missingness policy, or evaluator. Monitoring therefore includes false credit, false exclusion, burden by source class, appeals, concentration, complaints, correction time, and rule changes.

Consider temporal strategy. If actors learn that recent evidence earns more, publication timing may shift. If only first citations count, writers may optimize early passages. If verification is expensive, smaller sources may exit while incumbents amortize the cost. A snapshot allocation cannot describe these entry, exit, and adaptation dynamics. A real mechanism needs time blocks, rule-version logs, and a transition policy.

Strategic response can also involve collusion and dependence. Several sites may syndicate the same origin and appear to corroborate one another. The mechanism must preserve origin and dependency records without assuming that shared language always means coordination. False-positive duplicate detection can silence legitimate reporting. Controls require calibrated evidence and appeals.

We discuss these risks defensively. The aim is not to teach manipulation. The aim is to preregister tests and prevent a rule from being declared successful simply because its intended score rises.

Retrieval check: which rule is strategy-free? None. A mechanism may make a desired strategy more attractive and harmful strategies more costly, but that claim is environment-specific and empirical.

## Chapter 4

W15 distinguishes contribution, need, authority, and harm. Contribution is a source’s conditional effect on declared utility. It depends on the outcome, players, context, model, and coalition or removal design. It is not authorship or truth.

Need asks whether an actor is systematically under-credited or cannot bear the burden of producing and verifying useful evidence. Need can justify redistribution, but measuring it can invite strategic reports or intrusive profiling. A legitimate rule needs evidence, privacy protection, and review.

Authority is claim-relative fitness. An organization can be primary authority for its own current policy and interested evidence for an independent effectiveness claim. Prestige is not universal authority. If historical credit increases an authority score, the mechanism may reinforce incumbency.

Harm includes misinformation, privacy exposure, inaccessible evidence, reputational injury, discriminatory burden, competitor suppression, or degraded source diversity. When harm is non-compensable, represent it as a gate. A large contribution should not purchase permission to impose an unacceptable harm.

Weights among these dimensions are policy choices. The W15 safeguarded formula uses fifty percent contribution, thirty percent need, and twenty percent authority after verification and harm gates. Those numbers are not discovered truths. They make a transparent allocation possible so learners can challenge who authorized the weights and how sensitive the distribution is.

Distribution must be reported separately from total utility. Source shares, group totals, concentration, zero-credit decisions, false exclusions, and appeal burdens can disagree. Lower concentration does not prove correct evidence. A rule that helps community sources can still become more concentrated if another source is excluded.

Pause and ask whether your own fairness claim means equality, contribution, need, authority, harm protection, procedural voice, or some combination. If the answer is “all of them,” specify the conflicts and decision rights instead of hiding them inside one adjective.

## Chapter 5

We now enter the Estuary synthetic case. Four fictional sources share 100 credits. Source A, Public Lab, has displayed count four, contribution zero point four two, need zero point one zero, authority zero point nine zero, harm zero point zero five, and passes verification. B, Local Field Group, has count two, contribution zero point two eight, need zero point eight five, authority zero point five five, harm zero point one zero, and passes.

C, Accessibility Collective, has count one, contribution zero point one eight, need zero point nine zero, authority zero point six five, harm zero point zero five, and passes. D, Commercial Repeater, has count five, contribution zero point one two, need zero point two zero, authority zero point four five, harm zero point six zero, and fails verification. These are authored values, not measurements.

The equal rule gives 25 credits to each source. Its simplicity does not remove the player-identity problem. If one origin splits into three nominal sources, equality per player may produce inequality per origin.

The displayed-count rule divides counts four, two, one, and five by twelve. Using largest-remainder rounding to hundredths, A receives 33.33, B 16.67, C 8.33, and D 41.67. The total remains 100.

D receives the most under displayed count even though its separate verification record fails. That is deliberate. The naive count rule has no verification or harm gate. It shows how a mechanism can reward visible repetition while ignoring evidence quality.

The fixed pool also creates an externality. More credit for one source reduces shares for others. A statement that D “improved visibility” is incomplete unless it reports B and C’s lost opportunity under the same rule.

Calculation check: why does B receive the first residual cent before A when rounding? Its unrounded value, sixteen point six recurring, has a larger fractional remainder than A’s thirty-three point three recurring after flooring to cents. D has the same high remainder; the tie is resolved by source ID order.

## Chapter 6

The contribution-proportional rule is simpler because the authored contribution values sum to one. It gives A 42.00, B 28.00, C 18.00, and D 12.00. This allocation exactly follows the declared contribution record. It does not validate the record’s utility, source set, evaluator, or causal interpretation.

Now the safeguarded evidence rule. Eligibility requires verification pass and harm no greater than zero point four zero. D fails both the verification record and harm threshold, so its safeguarded credit is zero. For A, B, and C, raw score equals zero point five times contribution plus zero point three times need plus zero point two times authority.

A is zero point two one plus zero point zero three plus zero point one eight, totaling zero point four two zero. B is zero point one four plus zero point two five five plus zero point one one, totaling zero point five zero five. C is zero point zero nine plus zero point two seven plus zero point one three, totaling zero point four nine zero. The eligible total is one point four one five.

Normalize to 100. After rounding, A receives 29.68, B 35.69, C 34.63, and D zero. The values total exactly 100.

Why might someone defend this rule? It requires a verification record, blocks high harm, preserves contribution, and recognizes need and authority. Why challenge it? The weights are normative. The rubrics can be wrong or gamed. Authority can reinforce incumbency. Need review can be intrusive. A zero is severe. A false gate can exclude a legitimate source. The platform controls the evaluator.

The correct W15 disposition is not “safeguarded is fair.” It is: this rule is specified well enough for independent review, distributional challenge, sensitivity, and appeal design. No real credit is allocated.

Now inspect conjunctive gates. Suppose D’s harm score is corrected from zero point six to zero point three. D still receives zero because verification fails. Suppose verification passes but harm remains zero point six. D still receives zero. Both conditions must pass. This protects the rule from a large composite score compensating for a failed condition, but it also makes gate errors consequential.

An appeal must therefore record which condition was challenged, new evidence, reviewer independence, decision, and effective version. If D becomes eligible, every survivor’s normalized share changes because the pool is fixed. The appeal is not only an individual correction; it has distributional consequences for A, B, and C.

## Chapter 7

Compare all four distributions. Equal is 25, 25, 25, 25. Displayed count is 33.33, 16.67, 8.33, 41.67. Contribution is 42, 28, 18, 12. Safeguarded is 29.68, 35.69, 34.63, zero.

For one declared fictional grouping, community sources B and C receive 50 under equal, 25 under displayed count, 46 under contribution, and 70.32 under safeguarded. A and D together receive the complements: 50, 75, 54, and 29.68. This grouping is a teaching lens, not a demographic fact.

Now calculate concentration from the unrounded normalized rule weights as the sum of squared shares. Equal is zero point two five. Displayed count is approximately zero point three one nine four four four. Contribution is zero point three zero one six. Safeguarded is approximately zero point three three five three eight nine.

Safeguarded has the highest concentration even though it increases B and C’s group total. Why? Only three sources receive credit, and D receives zero. This is a useful conflict: one favorable distribution statistic does not decide the fairness question.

Sensitivity changes weights, harm threshold, verification error, player dependence, grouping, rounding, or budget. If D wins an appeal, do not edit the original. Create a new source record, reviewer, evidence, gate decision, and allocation version. Preserve whether the appeal exposes a rule error or new evidence.

Procedural fairness requires notice, reason, evidence access where lawful, independent review, correction, and a response time. It also requires measuring who can afford to appeal. A community source may receive more credit but face a larger verification burden.

Create an error ledger with at least six rows: false credit, false exclusion, wrong-source attachment, dependence error, stale authority, and inaccessible appeal. Each row needs an affected party, severity, detection route, correction, and residual uncertainty. An overall allocation accuracy could look high while the rare wrong-source cases produce serious reputational harm.

Mechanism drift deserves its own record. A new judge, player-resolution rule, or harm threshold changes the allocation system even if the formula string stays constant. Preserve the old rule, effective date, transition, retroactivity decision, and monitoring. Otherwise historical and current credits cannot be compared.

Retrieval check: if equal has the lowest concentration, is it automatically fairest? No. It can reward unsupported repetition and identity fragmentation. Concentration answers one distribution question, not evidence validity or procedural justice.

## Chapter 8

The paper routes sharpen the method but do not validate Estuary. PAPER-17 studies fair context attribution under an explicitly defined utility and player set. Its computational contribution shows that exact allocation can become more tractable for its construction. Exactness does not establish true hidden model use, authorship, ownership, truth, or universal fairness. Change the utility or merge dependent players and the meaning changes.

PAPER-42 studies strategic interaction between suppliers and a platform and evaluates a mechanism that combines suspicion with credit for earlier-version-verifiable content. It is useful because it treats actors as adaptive and retains component ablations. Its “win-win” label is conditional on a custom defense and creator-welfare objective, stakeholder weights, fixed candidate environment, finite repeated simulation, judges, and verification channel.

Earlier-version support is not independent truth. A false statement already present can pass that channel. A custom welfare score is not legal fairness, social welfare, consumer benefit, or market health. The study motivates a hypothesis for controlled evaluation, not a production guarantee or global equilibrium.

PAPER-36 supplies an extension on competitive first-citation choice. Exactly two anonymized candidate sources are inserted into a controlled context. That design can isolate one factor and position in its supplied setting. It does not call a live search engine, identify crawling or retrieval, measure all citations, test claim support, or measure user or market welfare.

The opportunity cost is important: in an exactly two-source first-citation choice, a higher probability for one source necessarily changes the other source’s opportunity. A tactic cannot be called a universal win from the target source’s outcome alone.

Together, PAPER-17, PAPER-42, and PAPER-36 support explicit players, utility, strategic adaptation, controlled comparison, and evidence ceilings. None supplies the W15 numbers, proves the safeguarded weights, or authorizes deployment.

## Chapter 9

The second half of W15 is status-first governance. We distinguish eight institutional classes.

A law or official regulation or measure has an issuer, legal form, jurisdiction, dates, actors, scope, and enforcement context. Identity does not decide applicability or provide operational legal advice.

A mandatory national standard has mandatory status within its formal system and scope. That does not make it universal across geography, actor, content, or date. A recommended national standard has published recommended status; it is not automatically law or mandatory.

An international management standard can specify organizational management-system requirements or guidance. Public metadata can verify title, edition, status, and overview scope. It cannot supply paywalled clauses, prove implementation, establish conformity, or create certification.

A voluntary framework organizes risk or governance work without becoming law or a certificate. A technical specification defines technical behavior or formats; conformance to a format does not prove truth or utility. Guidance recommends considerations or actions within an issuer’s scope but does not automatically impose obligations. A code of practice records behavioral expectations from an issuer or community; its force depends on adoption, contract, regulation, or law.

Do not arrange these classes as a universal strength ladder. A technical specification may be the best authority for a provenance manifest and irrelevant to legal fairness. A legal requirement may impose disclosure and say nothing about scientific metric validity.

The rule is: verify the institutional object, then ask applicability, then map a control, then collect evidence. Never jump from an official title to “compliant,” “certified,” or “conformant.”

## Chapter 10

Now read the S01–S10 crosswalk at the course cutoff of 24 August 2026.

S01 is NIST AI RMF 1.0, a United States government voluntary framework. It is not law, certification, or a legal safe harbor. S02 is NIST AI 600-1, a government technical report and voluntary GenAI profile or guidance. It does not prove that a product implements its suggested actions.

S03 is ISO/IEC 42001:2023, a published international AI management-system standard. Public metadata verifies identity and broad scope, not paywalled requirements, conformity, or certification. S04 is C2PA Content Credentials Technical Specification 2.4. It supports technical provenance-manifest questions; provenance integrity does not establish truth, authorship, or ranking.

S05 is the PRC Interim Measures for the Management of Generative AI Services, an official regulatory text with jurisdiction, actor, public-service, and exclusion questions. S06 is the separate PRC measure on labeling AI-generated and synthetic content, with its own dates and covered roles. Neither is operational legal advice from this course.

S07 is GB 45438-2025, a mandatory Chinese national standard concerning specified labeling methods within scope. Mandatory status does not establish that Estuary or every author worldwide is covered. S08 is GB/T 45654-2025, a recommended Chinese national standard on basic security requirements for generative AI services. S09 is GB/T 45652-2025, a recommended Chinese national standard on pre-training and fine-tuning data security. S10 is GB/T 35273-2020, a recommended Chinese national personal-information security specification, with a revision process noted at the cutoff.

Every applicability card must record jurisdiction, date, actors, scope, force, evidence, reviewer, and uncertainty. Add issuer, identity, effective date, accessible text, control, exceptions, and refresh where relevant. “Unresolved” is valid and may block release.

Imagine an analyst writes, “S03 applies because our workflow uses AI.” That skips organization, management-system boundary, role, adoption, contract, jurisdiction, and clause access. The repair is not a stronger conclusion. It is an applicability question routed to a qualified reviewer with licensed text and organizational evidence.

Now imagine, “S07 is mandatory, therefore every generated page must display the same label.” Mandatory is a verified status, not a universal scope determination. The card still needs jurisdiction, covered actor, service and content type, distribution state, dates, exceptions, exact provision, and evidence. W15 does not supply that legal analysis.

For a code of practice, identify issuer and adoption route. It may be purely voluntary, contractual, professional, or incorporated into another requirement. The synthetic CODE-LOCAL-01 applies only to this course fixture. Calling it a code does not create public authority.

For S03, the W15 card says public overview only; licensed clause mapping is absent. For S07, it says actor and content scope for the fictional service remain unresolved. The synthetic code `CODE-LOCAL-01` illustrates code-of-practice status but creates no external requirement.

## Chapter 11

L07 closes the lesson with a release decision. It runs six inert cases: three blocked, two detected, and one escaped. The command says, “L07 PASS: 6 cases, 1 escape, release BLOCKED.” The escaped case, AB-05, loses an accessible claim-to-source relation. Its residual severity is high, likelihood likely, and score nine.

The audit passes because it correctly preserves the escape and applies the blocking rule. Release remains blocked. This distinction prevents successful detection from being rewritten as safety.

Trace the six outcomes. AB-01 rejects an unresolved source identity before graph construction. AB-02 detects a dependency cluster and withholds aggregate corroboration. AB-03 keeps an inert source marker from becoming a control instruction. AB-04 quarantines an identifier-shaped synthetic field. AB-05 is the escape: an automated structure check passes, but manual review finds that an assistive-technology user cannot recover the claim-to-source relation. AB-06 detects and corrects the collapse of voluntary guidance into mandatory law.

This pattern teaches two further lessons. Detection after exposure is not prevention, and an automated pass cannot replace the human test required by the control. AB-05’s residual risk stays high and likely. The release block remains until the evidence relation is repaired, tested with appropriate assistive technology and human review, rollback verified, and the decision explicitly reconsidered. The validator cannot predeclare that remediation successful.

The L07 governance vocabulary is deliberately narrow. It keeps labels consistent and computes the blocking rule. Its own boundary says it does not determine legal applicability, compliance, certification, or safety. W15 uses that distinction as a model for all automation: software may preserve a decision contract without possessing the authority or evidence to make every underlying judgment.

State the final result. In the synthetic allocation, safeguarded credit gives A 29.68, B 35.69, C 34.63, and D zero under declared verification, harm, contribution, need, and authority choices. It advances only to independent review and is not proven fair. S01–S10 retain distinct statuses; applicability remains a dated qualified judgment. No paywalled clauses, legal advice, compliance, conformity, certification, safety guarantee, or production effect is established.

Your exit response names one actor rewarded, one actor burdened, one gaming path, one unresolved applicability fact, and the reason L07 is both PASS and BLOCKED. That is the W15 discipline: reward useful evidence only through a rule whose assumptions, strategies, distribution, governance status, and failure boundaries remain visible.

If any of those fields is unknown, say “unresolved,” name the accountable reviewer, and preserve the block. Precision about uncertainty is stronger evidence than a polished but unsupported fairness, compliance, or safety label.