# W01 Worked Synthetic Case — HarborCell H2

## Case status and purpose

Every entity, product, system, source, response, URL, timestamp, and event in this case is fictional. The fixture is designed to test reasoning, not to simulate evidence about any real platform or company. It can be used offline and contains no personal data, commercial account state, or live collection requirement.

The learner’s task is to move from raw input to atomic claims, map each event to a pipeline stage, reject invalid inferences, and write the strongest conclusion supported by the record.

## 1. Frozen input bundle

### 1.1 System card

| Field | Frozen value |
|---|---|
| Fixture system | Glimmer Answer Surface |
| Fixture version | `synthetic-1.0` |
| Surface | single-turn cited answer |
| Corpus state | not exposed by the fixture |
| Candidate retrieval | not exposed by the fixture |
| Effective context | not exposed by the fixture |
| Citation attachment process | not exposed by the fixture |
| Locale / language | `en-HK` / English |
| Account state | none; static fixture |
| Collection method | supplied response record; no live call |
| Query ID | `Q-001` |
| Observation ID | `O-001` |
| Nominal fixture time | `2026-08-18T10:00:00+08:00` |
| Time meaning | a synthetic field for schema practice, not a real collection event |
| Missing-state rule | unknown internal fields remain `unknown` |

The fixture makes the response and displayed citation observable. It deliberately withholds retrieval and generation traces. The learner may not infer those traces from the fact that a citation appears.

### 1.2 Query record

```json
{
  "query_id": "Q-001",
  "text": "Which campus e-bike battery is best for long rainy commutes?",
  "language": "en",
  "locale": "en-HK",
  "intent": "comparative recommendation",
  "population_status": "single synthetic test query; no population claim",
  "sensitive_attributes": [],
  "notes": "Best, long, and rainy are underspecified and should trigger scope checks."
}
```

The query itself creates ambiguity. “Best” needs a declared objective and alternatives. “Long” needs a distance or duty cycle. “Rainy” may invoke enclosure, electrical safety, riding behavior, maintenance, and operating conditions. A good system could respond by asking for constraints; the fixture instead makes a broad recommendation so that the audit has something to diagnose.

### 1.3 Governed source identity

```json
{
  "source_id": "S-001",
  "issuer": "Blue Marsh Mobility (fictional)",
  "title": "HarborCell H2 Product Specification",
  "version": "2.1",
  "issued": "2026-07-30",
  "effective": "2026-08-01",
  "url": "https://example.invalid/blue-marsh/harborcell-h2/spec-v2-1",
  "source_type": "fictional first-party product specification",
  "license": "course fixture; original synthetic text",
  "supersedes": "2.0",
  "correction_contact": "fixture-owner@example.invalid"
}
```

The `.invalid` top-level domain is reserved for examples. The URL is an identifier inside the exercise, not a route to retrieve.

### 1.4 Source passages

**Passage S-001:P1 — capacity**

> HarborCell H2 has a nominal energy capacity of 480 Wh when measured at the rated voltage. Usable energy varies with temperature, load, battery age, controller configuration, and cut-off settings.

**Passage S-001:P2 — enclosure**

> The installed enclosure is rated IP54 under the laboratory configuration described in Test Record TR-H2-54. The rating does not mean the product is suitable for immersion, pressure washing, riding through floodwater, or use with damaged seals.

**Passage S-001:P3 — warranty**

> The product carries a 24-month limited warranty from the documented retail purchase date, subject to the exclusions and service process in Warranty Terms WT-2026-04.

**Passage S-001:P4 — explicit non-claims**

> This specification does not provide comparative safety, “safest product,” maximum-trip, or real-world range findings. Range and safe operation depend on the complete vehicle, installation, rider, route, weather, maintenance, and operating conditions.

These passages are first-party specifications in the synthetic evidence design. They can support the issuer’s stated product properties and limitations. They do not independently establish comparative superiority or real-world performance across a market.

### 1.5 Synthetic answer observation

```json
{
  "observation_id": "O-001",
  "query_id": "Q-001",
  "fixture_system": "Glimmer Answer Surface",
  "fixture_version": "synthetic-1.0",
  "visible_response": "HarborCell H2 has 480 Wh capacity, an IP54 enclosure, a 36-month warranty, and is the safest long-range campus option [1].",
  "displayed_attributions": [
    {
      "marker": "[1]",
      "display_label": "HarborCell H2 Product Specification",
      "resolved_source_id": "S-001"
    }
  ],
  "visible_follow_up": null,
  "interaction_trace": null,
  "internal_trace": null
}
```

The answer is the only system output in the fixture. No second run, alternative query, candidate list, context record, user trace, or intervention comparison is supplied.

## 2. Object inventory

| Object type | Case instance | Why the distinction matters |
|---|---|---|
| Entity | HarborCell H2 | The product can be mentioned even if no governed source is cited. |
| Issuer entity | Blue Marsh Mobility | Issuer identity is not identical to the source document. |
| Source | S-001, specification v2.1 | Citation resolution and version determine which claims are available. |
| Passage | S-001:P1 through P4 | Each passage bears on different atomic propositions. |
| Query | Q-001 | One synthetic query is not a sampled demand population. |
| Response | visible text in O-001 | The response contains several separately auditable claims. |
| Attribution | marker [1] resolving to S-001 | A resolved source relation does not extend support to all nearby claims. |
| Events | mention, citation display, claim support status | Each event has its own rule and unit. |
| Downstream outcome | none observed | No click, follow-up, decision, installation, purchase, or safety outcome exists in the record. |

## 3. Atomic claim ledger

The entity mention is recorded separately from the five propositions in the answer.

| Claim ID | Atomic response proposition | Relevant source passage | Disposition | Reasoning | Strongest admissible rewrite |
|---|---|---|---|---|---|
| A-01 | H2 has nominal capacity of 480 Wh. | P1 | **Supported with scope** | The number and property match; “nominal” and operating variability should remain visible in technical use. | “The specification lists 480 Wh nominal capacity, with usable energy dependent on operating conditions.” |
| A-02 | H2 has an IP54 enclosure. | P2 | **Partially supported / scope restoration required** | The rating matches, but the answer omits the laboratory configuration and explicit exclusions; the short form can invite an immersion or all-weather interpretation. | “The installed enclosure is rated IP54 under the stated test configuration; the specification excludes immersion, floodwater, and damaged seals.” |
| A-03 | H2 has a 36-month warranty. | P3 | **Contradicted** | The governed source states a 24-month limited warranty. | “The specification states a 24-month limited warranty, subject to the issued terms.” |
| A-04 | H2 is the safest campus option. | P4 | **Unsupported and explicitly disclaimed** | No comparative safety study, alternatives, safety construct, test population, or result is supplied. | Remove the superlative; no evidence-supported comparative rewrite is available. |
| A-05 | H2 is a long-range campus option. | P1 and P4 | **Unsupported / undefined construct** | Capacity alone does not identify range. Vehicle, controller, route, weather, rider, and test cycle are missing, and P4 rejects a range finding. | “The specification does not provide a real-world range estimate; evaluate range under a declared vehicle and route protocol.” |

### 3.1 Citation audit

Marker [1] resolves to the correct governed source identity, so the fixture contains a **resolved displayed citation event**. Claim-level support is mixed. The citation supports A-01 with scope and bears on A-02 if conditions are restored. It contradicts A-03 and cannot support A-04 or A-05.

Therefore, “the answer has a citation” and “the answer is supported” must remain separate fields. The record also demonstrates why citation count is a poor substitute for correctness and completeness: a single source marker sits beside five propositions with four different dispositions.

### 3.2 Claim–Evidence–Source edges

```text
A-01 --supported-with-scope--> S-001:P1 --part-of--> S-001 v2.1
A-02 --partial/conditions-omitted--> S-001:P2 --part-of--> S-001 v2.1
A-03 --contradicted-by--> S-001:P3 --part-of--> S-001 v2.1
A-04 --explicitly-not-established-by--> S-001:P4 --part-of--> S-001 v2.1
A-05 --explicitly-not-established-by--> S-001:P4 --part-of--> S-001 v2.1
```

This graph records support relations. It does not show that S-001 caused the response wording.

## 4. Six-stage map

| Stage | Event definition for this case | Available evidence | Status | What would raise the ceiling? |
|---|---|---|---|---|
| Discoverable \(Z_D\) | S-001 is available to the fixture’s acquisition process. | No acquisition or corpus record. | **Unknown** | Versioned corpus manifest or permitted fetch/index log. |
| Retrieved \(Z_R\) | S-001 or a representation enters the candidate set for Q-001. | No run file or candidate trace. | **Unknown** | Query-linked candidate list with source resolution. |
| In context \(Z_C\) | S-001 enters the effective generation context. | No context trace. | **Unknown** | Frozen ordered context or instrumented open-system log. |
| Contributes \(Z_G\) | Source-distinctive evidence from S-001 shapes a response claim. | Two common specifications appear; no alternative-source set or intervention trace. | **Unresolved proxy** | Distinctiveness audit, controlled source perturbation, or open trace, with fidelity checks. |
| Attributed \(Z_T\) | A visible marker resolves to S-001 and supports a neighboring atomic claim. | Marker [1] resolves; support differs by claim. | **Display observed; faithfulness mixed** | Claim-level entailment and completeness labels with reviewer protocol. |
| Outcome \(Z_A\) | A predeclared user or operational consequence occurs. | No interaction or task trace. | **Not observed** | Authorized event record plus an identification design for causal interpretation. |

The stage table contains three different non-positive states. **Unknown** means the fixture withholds the relevant variable. **Unresolved proxy** means indirect evidence exists but cannot distinguish source-specific use. **Not observed** means the defined downstream event is absent from the provided record. These states must not be collapsed into zero.

## 5. Measurement cards

### 5.1 Entity mention

- Construct: entity occurrence in a visible response.
- Unit: entity–response.
- Eligible records: the one valid supplied response.
- Numerator: 1 response containing the resolved string “HarborCell H2.”
- Denominator: 1 valid supplied response.
- Descriptive result: 1/1 for this fixture only.
- Prohibited interpretation: population mention rate, favorable sentiment, source use, market visibility, or comparative recommendation quality.

### 5.2 Resolved citation display

- Construct: visible source attribution.
- Unit: source–response.
- Event rule: a displayed marker resolves to the governed source ID.
- Numerator: 1 O-001–S-001 pair.
- Denominator: 1 eligible O-001–S-001 pair in the supplied fixture.
- Descriptive result: 1/1 for this fixture only.
- Prohibited interpretation: faithful support for all response claims, retrieval probability, source contribution, user notice, or traffic.

### 5.3 Claim support profile

- Construct: disposition of each atomic response proposition relative to the supplied source set.
- Unit: claim–response with source passage edges.
- Eligible claims: A-01 through A-05.
- Profile: one supported with scope, one partial with required scope restoration, one contradicted, two unsupported.
- Aggregation rule: do not reduce to one percentage in W01; dispositions have different meanings and remediation paths.
- Missing evidence: no independent source set, inter-rater record, or external product validation.

### 5.4 Absorption

- Construct: source-distinctive information reflected in the response.
- Unit: claim–source–response.
- Current disposition: unresolved.
- Reason: “480 Wh” and “IP54” may appear in many representations; the fixture has no alternative-source corpus or trace.
- Prohibited interpretation: using citation display as the absorption label.

## 6. Invalid inferences

Each statement below exceeds the fixture.

1. **“S-001 was crawled and indexed.”** No acquisition evidence exists.
2. **“S-001 ranked first.”** No candidate list, rank, or reranking trace exists.
3. **“The answer engine used S-001 to generate the specifications.”** Common facts plus a citation do not identify source-specific contribution.
4. **“The citation validates the whole sentence.”** A-03 is contradicted; A-04 and A-05 are unsupported.
5. **“H2 has the best visibility.”** There is one entity and no alternative product, query population, metric, or comparator.
6. **“The system recommends H2 because of its IP54 rating.”** The visible response does not identify the causal reason for ordering or wording.
7. **“The cited page is authoritative about comparative safety.”** It is first-party evidence for stated specifications, not an independent comparative safety study.
8. **“One out of one means a 100% citation rate.”** The arithmetic describes the fixture but not a population estimate; there is no sampling frame or repetition.
9. **“The false warranty proves the source was ignored.”** The source could have been present while generation or attribution failed; the stage is unknown.
10. **“A clearer page would fix the answer.”** No intervention or counterfactual observation exists.
11. **“A click would prove persuasion.”** No click exists, and an action would still require a causal design.
12. **“This case describes a commercial engine.”** Every system and event is synthetic.

## 7. Counterexample extension: two screenshots

An instructor may append a second synthetic record:

- `O-000`, Monday 09:00: H2 is not cited.
- Page revision at Monday 12:00: headings are shortened; the instructor says factual claims are intended to remain fixed.
- `O-001`, Tuesday 10:00: H2 is cited.

The pair does not identify a heading effect. Treatment is perfectly confounded with time. Ordinary output variance is unknown. Query, session, competing sources, index state, system alias, and citation display rules may differ. “Intended factual equivalence” is not an audited content diff.

The strongest descriptive sentence is: “The two synthetic response records differ in whether S-001 is visibly cited; they were captured in different time blocks surrounding a page revision.” A suitable next study would freeze and diff both variants, specify the target stage, use repeated blocked or randomized runs where permitted, preserve all failures, and record a system version or an explicit unknown-state field.

## 8. Bounded conclusion

> In synthetic observation O-001 on Glimmer Answer Surface fixture `synthetic-1.0`, the visible response mentioned the fictional HarborCell H2 entity and displayed citation [1], which resolves to fictional source S-001 version 2.1. Among five atomized response propositions, A-01 is supported with scope, A-02 requires restoration of material conditions, A-03 contradicts the source, and A-04–A-05 are unsupported by the supplied source set. The fixture does not expose acquisition, candidate retrieval, effective context, source-specific contribution, or downstream action. No population frequency, comparative product quality, hidden mechanism, causal intervention effect, or real-platform behavior is estimated.

This is the maximum-strength conclusion for the supplied record. It includes the favorable events, the factual failures, and every major unknown in one inspectable statement.

## 9. Learner deliverable and self-check

Submit:

1. the completed system card with unknown fields preserved;
2. the five-row atomic claim ledger;
3. the six-stage map using observed, unresolved proxy, unknown, and not-observed states;
4. one measurement card with an explicit unit and denominator;
5. three invalid inferences rewritten as bounded statements; and
6. a conclusion no stronger than the one supported by the fixture.

Self-check before handoff:

- Did I keep entity, source, claim, and response units separate?
- Did I distinguish citation display from claim support and absorption?
- Did I restore material source conditions rather than treating exact number match as complete fidelity?
- Did I preserve contradiction and unsupported statuses?
- Did I avoid filling hidden stages with a familiar architecture?
- Did I state that the query and system are synthetic and not population evidence?
- Did I name what additional evidence would change each unknown?