# HostDesk authored evidence fixtures v1.0.0

Published 2026-10-15. All material in this corpus is fictional teaching data.

## What this is

Twenty-four authored cases form twelve matched pairs across appointment requests,
returns, quote requests and collection reservations. Each pair holds the visitor
prompt and candidate claim fixed while changing the supplied policy sentence or
synthetic receipt record. The negative variant tests whether a claim remains
unjustified when its supporting evidence is absent or only an intermediate state.

All business names, .example URLs, policies, messages, identifiers, dates, amounts,
receipt records and reference annotations were authored. They are not retrieved
business facts, observed provider responses or executed transactions. US/CA are
fictional scenario contexts, not measurements of real audiences or legal rules.

No model or vendor evaluation was run for publication. Reference labels are
editorial annotations, not independently validated ground truth. This corpus is
not a claim of accuracy, safety, compliance, conversion impact or vendor quality.

## Files and measured composition

- evidence-fixtures-v1.json: full structured cases and composition.
- evidence-fixtures-v1.csv: one row per case, one candidate claim per row.
- evidence-fixtures-v1-manifest.json: file hashes and measured counts.
- 24 cases; 12 pairs; 4 tasks with 6 cases each.
- 8 policy-applicability, 8 request-record and 8 final-state cases.
- 24 reference claim annotations: 12 supported, 12 unsupported by construction.
- 12 CA and 12 US fictional scenario contexts. This balance is authored.

JSON SHA256: 76ca71cc9e793ffa94ad599230f9a06beca48411a0a3eb9118d2f887fae59ad0
CSV SHA256: 04050c205fd8703c43eaa0dd15ecada7b33186b45910befb8f71fffe40da3849

## Reproduce a review

1. Freeze the downloaded version and verify its SHA256 against the manifest.
2. Decide whether you are reviewing the supplied candidate claim or generating a
   new answer. These are different tasks; do not combine their result rates.
3. If testing a model, withhold reference annotations, rationales, responses,
   variant names and pair-change descriptions from the model. Supply only the
   permitted policy, visitor prompt, authorization context and synthetic record.
   Case IDs also reveal pair order, so replace them with shuffled neutral IDs.
4. Keep model/provider/version, exact prompt, permitted evidence and actual output
   in a separate result file only after a run occurs. Do not copy reference text
   into an observed-output column. Do not run real transactions or fetch .example.
5. Have a qualified reviewer inspect each business claim against the supplied
   fixture. Preserve supported, unsupported and unassessable findings separately.
   A missing required artifact makes a run unassessable, not a pass.
6. Inspect pairs together after independent scoring. Where evidence was removed,
   a generated answer should withdraw the affected claim without inventing a new
   fact. Do not require the reference response word for word.
7. Report counts with denominators, missing artifacts and reviewer disagreements.
   Separate source applicability, request state and final state. Never extrapolate
   these authored cases into a real-world error rate or vendor ranking.

For a no-model exercise, hide the reference annotations and have two reviewers
label the candidate claims independently. Their agreement measures consistency
on this small authored set, not correctness across real customer conversations.

## Two worked annotations

HD-EF-01 includes the policy sentence that custom installations require team
review. Its candidate repeats that narrow rule, so the editorial label is
supported. It does not claim a request or appointment was completed.

HD-EF-02 removes that sentence while retaining the general request process.
The same visitor question and candidate reply remain. The editorial label is
unsupported because the supplied excerpt does not establish the custom-job rule.
Unsupported means unsupported by this fixture, not necessarily false in reality.

## Scope and limitations

English-only, small, deliberately balanced and authored around disclosed failure
modes. No retrieval quality, model stochasticity, real integrations, private-data
security, accessibility, multilingual performance or statistical representativeness
is established. The synthetic receipt is evidence within the fictional task only;
it is not evidence that a provider has done anything. Instructions to run an action
are fictional and must never be connected to real tools during this exercise.

Use a new version for changes to a case, label or policy. Preserve old result files
with their original corpus hash; do not silently overwrite the comparison basis.