# W15 Worked Case — The Estuary Evidence Credit Pool

## Case status

This is a deterministic synthetic allocation and governance review. The actors, evidence scores, credit values, code-of-practice card, and applicability conclusions are fictional teaching inputs. No platform, creator, user, regulator, conformity body, or certification body participated. The case is not legal advice, a conformity assessment, or a production fairness result.

## 1. Decision question, stakeholders, and information

Estuary is a fictional answer service with a fixed 100-credit pool for one answer bundle. It must compare four rules and decide whether any should proceed to further human review. The actors are four source creators, the platform operator, users, an affected subject group, a rights holder, an evidence reviewer, an accessibility reviewer, and a governance reviewer.

Creators know their effort and evidence lineage but not the full candidate set. The platform knows displayed counts and controls player identity, utility scoring, and the credit budget. Users see the answer and citations but not the counterfactual contribution trace. Affected subjects may know harms omitted from the platform record. Reviewers see only assigned artifacts. These asymmetries create strategies: repetition, fragmentation, syndication, authority mimicry, selective disclosure, appeal, abstention, and gate capture.

The decision rule does not name one allocation “fair.” It reproduces four distributions, audits gaming and burden, and decides whether a candidate is sufficiently specified for accountable review.

## 2. Frozen source and evidence records

| ID | Fictional source | Display count | Contribution | Need | Authority | Harm risk | Verification |
|---|---|---:|---:|---:|---:|---:|---|
| A | Public Lab | 4 | 0.42 | 0.10 | 0.90 | 0.05 | pass |
| B | Local Field Group | 2 | 0.28 | 0.85 | 0.55 | 0.10 | pass |
| C | Accessibility Collective | 1 | 0.18 | 0.90 | 0.65 | 0.05 | pass |
| D | Commercial Repeater | 5 | 0.12 | 0.20 | 0.45 | 0.60 | fail |

Contribution sums to `1.00`; count sums to `12`. Need, authority, and harm are authored rubric values, not validated constructs. Authority is claim-relative, not prestige. Verification concerns the frozen evidence package, not truth in general. D’s failed record and harm above `0.40` exclude it only from the safeguarded rule.

Credit is rounded to hundredths. The method floors unrounded amounts, then distributes remaining cents by fractional remainder, breaking ties by source ID ascending. Every rule must total exactly `100.00`.

## 3. Four credit rules and exact allocations

**Equal:** divide by four. Allocation `(A,B,C,D) = (25.00,25.00,25.00,25.00)`.

**Displayed count:** divide counts `(4,2,1,5)` by twelve. Largest-remainder allocation is `(33.33,16.67,8.33,41.67)`. D wins by visible repetition even though its separate verification record fails.

**Contribution proportional:** contribution already sums to one. Allocation is `(42.00,28.00,18.00,12.00)`. Exact arithmetic does not prove that the utility, player set, or contribution estimator is fair.

**Safeguarded evidence:** require verification pass and harm `≤0.40`, then calculate `0.50×contribution + 0.30×need + 0.20×authority`. Raw eligible scores are A `0.420`, B `0.505`, and C `0.490`, total `1.415`; D is zero. Normalized credit is `(29.68,35.69,34.63,0.00)`.

| Rule | A | B | C | D | Total |
|---|---:|---:|---:|---:|---:|
| Equal | 25.00 | 25.00 | 25.00 | 25.00 | 100.00 |
| Displayed count | 33.33 | 16.67 | 8.33 | 41.67 | 100.00 |
| Contribution | 42.00 | 28.00 | 18.00 | 12.00 | 100.00 |
| Safeguarded | 29.68 | 35.69 | 34.63 | 0.00 | 100.00 |

## 4. Gaming and distributional audit

Equal credit invites entity fragmentation. Displayed count invites repetition and salience engineering. Contribution invites utility, player, coalition, and evaluator gaming. Safeguarded credit invites need inflation, authority mimicry, verifier targeting, and strategic gate disputes. The platform can game all four by choosing the candidate set, exposure budget, or judge.

For a declared teaching grouping, community sources B+C receive `50.00`, `25.00`, `46.00`, and `70.32` under the four rules. A+D receive `50.00`, `75.00`, `54.00`, and `29.68`. The grouping is not a demographic fact; it is an authored distribution lens.

Concentration, computed from unrounded normalized rule weights as the sum of squared shares, is `0.250000` equal, approximately `0.319444` displayed count, `0.301600` contribution, and `0.335389` safeguarded. Safeguarded credit reallocates toward B and C but has the highest concentration because D receives zero. Both facts remain in the memo.

## 5. Contribution, need, authority, and harm challenge

The review panel asks four questions. Is contribution measured under a credible utility and player definition? Is need recorded without intrusive profiling or unreviewable self-assertion? Is authority fit for this claim and open to contest rather than inherited prestige? Is harm a gate with accessible evidence, proportional review, and appeal?

D may appeal verification or harm classification. An appeal does not edit version `EVIDENCE-CREDIT-1.0`. It creates a new evidence record, reviewer identity, rationale, gate result, and allocation. If harm must never be compensated, it remains a gate. If the organization instead chooses a soft harm penalty, that is a different policy with a new distribution and ethical burden.

The bounded allocation decision is to retain all four outputs as comparison evidence and advance safeguarded credit only to independent review. No credit is paid or exposure changed.

## 5A. Sensitivity, error, and appeal scenario

The panel preregisters three sensitivity questions without replacing the primary allocation. First, lowering the harm threshold cannot change A, B, or C because all are below `0.10`, but it keeps D excluded. Raising the threshold above `0.60` would not restore D while verification still fails. This demonstrates that eligibility is conjunctive. Second, removing the need term would favor A relative to B and C; that is a different contribution-and-authority policy, not evidence that need was wrong. Third, merging dependent source identities would require a new player set and invalidate direct comparison with version `1.0`.

The error ledger distinguishes false credit, false exclusion, wrong-source attachment, dependence error, stale authority, and inaccessible appeal. D’s appeal is handled by `synthetic_reviewer_s`, not the original gate reviewer. The reviewer can affirm, reverse, or mark unresolved. Reversal creates `EVIDENCE-CREDIT-1.1`, lists new evidence, recalculates all shares, and states whether past credit is corrected or only future credit changes. No appeal deletes the version-1.0 zero or the reason for it.

Affected subjects and users can challenge harm or support even though they receive no share. This process expands procedural standing beyond the player set. Whether it is sufficient remains a human fairness question.

## 6. PAPER evidence routes and limits

PAPER-17 supports reasoning about exact contribution allocation for a declared utility and player set. It does not establish authorship, ownership, truth, hidden model use, or universal fairness. PAPER-42 supports bounded strategic and verifiable-reward mechanism questions; its custom welfare construction and simulation do not establish legal fairness, global equilibrium, social welfare, or production success. PAPER-36 supports a controlled two-source first-citation choice estimand, not live retrieval, full attribution, claim support, user benefit, or market welfare.

These routes motivate the mechanism questions. None supplies the fictional inputs or validates the safeguarded weights.

## 7. S01–S10 status-first crosswalk

The case freezes the following public identities: S01 voluntary NIST AI RMF framework; S02 NIST GenAI technical report/voluntary guidance; S03 published ISO/IEC international management-system standard; S04 C2PA technical specification; S05 PRC Interim Measures official regulatory text; S06 separate PRC labeling regulatory measure; S07 mandatory Chinese national standard; S08, S09, and S10 recommended Chinese national standards.

Status does not answer applicability. S07’s mandatory label does not prove that Estuary, an actor, content item, or locale is in scope. S03’s public overview does not provide paywalled clauses, certification criteria, or conformity evidence. S04 provenance does not establish truth. S01 and S02 are not law. S08–S10 are not converted into mandatory requirements by inclusion in a course table.

## 8. Applicability-card comparison

One complete fictional review card for S03 records:

| Required field | Frozen entry |
|---|---|
| jurisdiction | International identity; organizational adoption and local legal context unresolved |
| date | edition 2023; accessed 2026-08-24; refresh before release |
| actors | fictional platform operator; certification and contractual roles unresolved |
| scope | candidate AI management workflow; clause-level mapping unavailable from public overview |
| force | published international management standard; not law by identity alone |
| evidence | official metadata plus local management-control artifacts; licensed clause text absent |
| reviewer | `synthetic_governance_reviewer_q`; qualified conformity/legal review required |
| uncertainty | applicability, exact requirements, implementation adequacy, conformity, and certification unresolved |

A second card for S07 records PRC jurisdiction, 2025 publication/effective dates, actor and AI-generated-content scope unresolved for Estuary, mandatory national-standard status, official identity metadata, governance reviewer, and explicit applicability uncertainty. It does not say “compliant.”

The synthetic `CODE-LOCAL-01` code of practice has course-fixture force only. It illustrates that codes can gain force through issuer authority, contract, or legal adoption; the name alone creates no external duty.

## 9. L07 deterministic governance result

The linked L07 system is `SYS-RISK-DEMO-001`, an offline synthetic corpus-to-evidence pipeline. Its command reports `L07 PASS: 6 cases, 1 escape(s), release BLOCKED`. Outcomes are three `blocked`, two `detected`, and one `escaped`. Blocking case `AB-05` is an accessibility-and-attribution loss with residual risk score `9`. A successful audit therefore preserves a blocked release.

L07 input hashes include system model `8758f155939beb631f67c1a5eab7b4b8919e559e9bc0f8d46814b72521972c3d` and governance crosswalk `2979542df5f63ca50196b9117806f49aa268c1c71b71e049680ccac9f722d80e`. The full eight-file identity is in the manifest. Its claim ceiling is `synthetic control tests and status vocabulary audit only; no safety guarantee, certification, compliance, or legal advice`.

## 10. Release memo and self-check

The decision is:

> The synthetic safeguarded rule allocates A 29.68, B 35.69, C 34.63, and D zero after declared verification and harm gates. It advances only to independent review because its weights, rubrics, distribution, exclusions, and gaming response are policy choices, not proven fairness. S01–S10 retain their verified statuses, while applicability, clauses, implementation, conformity, and certification remain unresolved. L07 passes its audit but blocks release after AB-05 escapes.

Before accepting the package, verify all totals equal 100; the displayed-count rule rewards D most; the community group totals are preserved; concentration is reported even when unfavorable; every applicability card includes jurisdiction/date/actors/scope/force/evidence/reviewer/uncertainty; no paywalled clause is inferred; and no legal, certification, conformity, fairness, safety, or production-effect conclusion appears.