# Oxford bibliography review apparatus

These main editorial reviews accompany 21 metadata records and six source-label
corrections. Internal JSON filenames refer to the retained review packet.
The calibration table omits private response paths. Raw source bodies are
excluded. Original source occurrences and historical judgments are unchanged.
The sealed review records describe the review before local application; the
current view and publication descriptor establish application to this build.

# Oxford metadata — Fact-Check and Revise

Main review, 9 September 2026; Fact-Check and Revise v2.8. The venue is
Craig's own research site and accompanying downloads. Model-assisted metadata
composition is authorized. The 21 identity candidates and six proposed source
labels are the subject; historical case judgments and scores are outside scope.

## Findings and disposition

The inspected publisher records identify two books, eleven book chapters,
seven modern reference articles and one journal article. Chapter titles, contributor names, containing
works, reported dates and displayed identifiers distinguish the linked
components from their original broad source labels. The Inventory's editor
roles are supported by its author-institution repository. That repository's
Hermann and Oxford's Herman spelling remain separate supplied variants.

Six current labels misidentify Oxford Classical Dictionary articles as Oxford
History of the Archaic Greek World chapters. The proposed corrections name the
actual article, credited authors and dictionary, while retaining each existing
URL and citation ID. The dictionary articles remain modern secondary sources;
an inherited inventory phrase such as Primary source does not change that.
Other original purpose labels remain provenance, not rewritten work titles.

The Roman source is Forsythe's chapter 8, not the book's general source chapter.
Its chapter byline names Gary Forsythe. Oxford also displays John Connelly as
an editor of the containing book. The UC Press paperback listing names Forsythe
as author; it is a different format and does not establish or disprove the
additional editor credit. The preferred editor field remains null. The 2006
paperback date/ISBN is not substituted for the linked 2005 edition metadata.

The Wales chapter has conflicting final pages: Oxford 82–108 and Crossref
82-107. Both attributed ranges remain, with no chosen preferred range. The
Panama chapter has no displayed page range; its registry request was limited.
These two page fields and the editor field are three explicit gaps. They are
not an exhaustive census of every omitted imprint or physical-edition detail.

Publication dates retain their source roles. The Metapontion chapter's 2024
publisher date and 2025 print availability are distinct. Samoa's and Owolabi's
publisher dates differ from the registered print dates; both roles remain
labelled. Cameron's registered 2019 online date is separate from the 2018
publisher/print date. The Valencia article's 2023 online publication is separate
from its 2024 issue. Oxford's own search guide says a book's Published label
may be its earliest print or online date; it is not uniformly first-online.
No print scan was newly inspected.

The Valencia article studies Valencia. Its use for the case covering Iberia
beyond Valencia requires the relevant general-context passages; its estimates
are not transferred to the wider case. The Samoa link is chapter 2, while its
contents list Chinese Workers and Settler Violence as chapter 6. Exact labour
passages remain a separate research question. Neither observation creates a
new global hold or changes the accepted historical judgment.

## Grounding and access limits

All 21 direct page requests returned HTTP 403. Those responses are not source
text. Title-based publisher searches supplied usable indexed metadata, and
seven public browser pages supplied the missing chapter metadata. Browser
transcripts record selected visible fields, not raw HTML or full chapter
captures. The Valencia article was available in full in the browser, but only
metadata, abstract and relevant disclosures were inspected. Its acknowledgments
name the Bank of Spain, MCIN/AEI, Generalitat Valenciana and Norwegian Research
Council; the data-availability statement offers data/methods on request. Its
rights notice identifies CC BY 4.0. No funding-based reliability judgment is
inferred. No conflict statement was located by the bounded search; that is not
a certification of no conflicts.

Thirteen of 22 DOI registry requests succeeded; nine were rate-limited, including
the Roman container query. Eight selected identities therefore retain publisher
identifier evidence without a completed registry comparison. Registered short
titles and containing-book punctuation are not substituted for fuller publisher
titles. The failed requests are terminal and not automatically retried. No
public archival URL is confirmed. Private response files and observed-field
transcripts preserve the evidence actually available, with their limits.

The revised candidates retain the draft's source-qualified fields and nulls.
The main review accepts their metadata identity scope and the six label repairs;
publication, citation-resolution checks and independent historical reviews
remain separate. No independent reviewer has been launched.

## Required self-audit

| Required check | Disposition | Evidence or reason |
|---|---|---|
| Grounding | Pass with limits | Publisher responses, seven browser metadata pages, institutional/publisher counterparts and 13 registry records support supplied fields. |
| Required vs. offered | Pass | Six wrong-container labels require repair; faithful historical purpose notes remain. |
| Claims | Pass | Observed titles and credits are stated directly; three preferred fields remain unresolved. |
| Deletions | Pass | All 21 original identities and occurrence records remain. |
| Criticisms | Pass | No outside reviewer finding or replacement prose was adopted. |
| Internal consistency | Pass | Chapter, book, dictionary and journal types and date roles remain distinct. |
| Register discipline | Pass | Publisher promotional abstracts are not adopted as evaluation. |
| Moral framework register | Pass | No normative historical conclusion is introduced. |
| Political-spectrum contamination | Pass | No political placement label is introduced. |
| Attribution posture | Pass | Publisher/registry dates, page ranges and the unconfirmed editor credit retain attribution. |
| Quotations | Pass | Supplied titles and credits follow inspected metadata; summaries of limits are unquoted. |
| Causal reasoning | Pass | No causal historical claim or funding-to-validity inference is added. |
| Categorical vs. dimensional | Pass | Metadata types do not assign historical scores or binary flight status. |
| Denotation/connotation and convenient layer | Pass | Dictionary articles are not ancient sources; a platform name is not a personal author. |
| Proper-name variants | Pass | Supplied initials and Herman/Hermann variants remain explicit. |
| Engagement posture | Pass | Existing case arguments and qualifications remain intact. |
| Confidence calibration | Gaps retained | Editor attribution and two page fields stay unresolved; access/registry limits are disclosed. |
| Fetched-source disclosures | Pass with limits | Article acknowledgments, data availability and rights plus platform date/affiliate notices were inspected. |
| Citation resolution | Pass at metadata scope | Proposed repairs retain exact citation IDs and URLs; live application still requires checks. |
| Adopted reviewer text | Pass | No reviewer-supplied text is present. |
| Editorial synthesis (SYNTH) | Pass | Work/component and date-role distinctions are bounded metadata findings, not historical adjudications. |


# Oxford metadata — LLM Artifact Tells

Main review, 9 September 2026; LLM Artifact Tells v2.7. The input is
REVISED_IDENTITY_CANDIDATES.json, the first-round report and proposed label
corrections. Craig's own site/downloads permit this model-assisted work.

## Findings and tier coverage

| Scope | Finding | Tier | Disposition |
|---|---|---|---|
| Candidate fields, labels and first-round prose | Actual marker, encoding, placeholder and tracking searches found no hard residue. | 1 | No repair required; TELL_CHECKS.json records the search. |
| Source identity and significance | Personal names, component identities and source-role qualifications are grounded. Nulls and registration limits are retained. | 2 | No invented identity or historical synthesis introduced. |
| Repeated data structure | Repeated limits and fields preserve provenance across separate access records. | 3 | Retain useful schema repetition. |
| Vocabulary | Colonization and settlement occur in actual work titles and existing purpose labels. | 4 | No vocabulary-only alteration. |
| Punctuation and initials | Source initials, ranges, dates and title subtitles serve bibliographic functions. | 5 | Preserve meaningful distinctions. |

All five tiers were considered in order. This round does not rewrite the
candidates. It does not flatten Craig's historical prose, invent a uniform
publication date, fill a missing range, or present clean wording as human
authorship. Eight publisher-observed identifiers lack a completed registry
comparison because of rate limits. Identifier syntax alone is not independent
source verification. Public apparatus must retain grounding and omit private
response paths. Private web-reader citation traces are outside this prose pass.

## Required self-audit

| Required check | Disposition | Evidence or reason |
|---|---|---|
| Checklist vs. heuristic | Pass | Actual hard-artifact checks precede contextual prose judgments. |
| False positives | Pass | Data-schema repetition and descriptive title terms remain. |
| Hard artifacts | Pass with limits | Marker/encoding/date/ISBN checks are recorded; DOI registration limits are explicit. |
| Tier coverage | Pass | Tiers 1–5 are covered above. |
| Decay weighting | Pass | Objective artifact evidence outranks familiarity of wording. |
| Guideline vs. style sort | Pass | Wrong-container labels and uncertain fields are source issues, not taste. |
| Introduced tells | Pass | No candidate rewriting in this round. |
| Writer's voice | Pass | Case narratives and manuscript remain unchanged. |
| Register homogenization | Pass | Original purpose labels remain provenance; metadata roles stay distinct. |
| Quality tradeoffs | Pass | Explicit title, date and pagination detail remains readable. |
| Editorial synthesis (SYNTH) | Pass | No causal or significance connector was added. |
| Venue check | Pass | The venue and model-assisted authorship are already authorized. |
| Provenance honesty | Pass | A clean metadata pass does not imply human authorship, complete chapter reading or full DOI registry success. |


# Oxford metadata — Calibration Review

Main review, 9 September 2026; Calibration Review v1.2. The final candidate
metadata and six proposed labels were checked after the Tells pass. No
independent reviewer was launched.

## Coverage and findings

CALIBRATION.json contains 26 groups with claim locations, load-bearing terms,
source URLs, exact grounding strings, locators, licensed scope, materiality,
verdict and proposed wording. Twenty-one identity groups cover all candidates
and the six label repairs. Three separate groups retain unresolved preferred
fields. Two groups cover the article disclosures and Oxford's date-label guide.
The verdicts are **23 Matched and 3 Uncheckable**, with no Overclaimed,
Underclaimed or Unsupported-but-sound classifications.

Grounding includes supplied names, titles, DOI strings, reported date labels,
page ranges and ISBNs. The table records inspected registry fields separately
where retrieved. Fuller publisher titles are retained when registered titles
omit subtitles. Browser evidence is explicitly an observed-field transcription;
it is not a raw HTML capture or complete chapter reading. Thirteen registry
responses support comparisons; nine requests, including the Roman container,
were limited. A supplied DOI is not promoted to a successful registry request.

All findings concern metadata or source-use disclosures. Operational access
times, checksums and occurrence preservation are checked separately. Ancient
events, statistical estimates, source independence, complete printed editions
and historical claims embedded in original labels are outside this audit.
The Valencia and Samoa scope questions remain questions, not adjudications.

## Concentration and reverse check

The three Uncheckable fields are the Roman container's preferred editor credit,
the Wales chapter's preferred page range, and the Panama chapter's missing
range. They remain null. Oxford's extra editor label is not declared false from
a different-format catalogue's omission. Conflicting final pages are retained
without choosing the convenient one. Neither rate limiting nor missing display
is evidence that a DOI or page range does not exist.

The reverse check tested whether explicit names, chapter titles or source-labelled
dates had been weakened merely because unrelated details remain open. They had
not. The 21 identities remain definite at the stated metadata scope. The six
dictionary labels are corrected directly. Source labels and registered print/
online roles remain distinct, without a blanket assumption that every Published
date is first-online. No candidate rewrite is required by calibration.

## Required review self-audit

| Required check | Disposition | Evidence or reason |
|---|---|---|
| Direction of changes | Pass | 23 groups are Matched; three preferred fields remain unresolved without blanket softening. |
| Lane discipline | Pass | The audit checks source-to-metadata confidence, not historical classification or prose taste. |
| Deletion vs. flagging | Pass | All 21 original records remain; preferred-field gaps are visible. |
| Mandatory re-cite flags | Pass | No Unsupported-but-sound verdict is used; explicit nulls and FIELD_EXCEPTIONS.json carry the actual gaps. |
| Verbatim grounding | Pass with stated method | Exact names, titles, date/page/ISBN strings and registry field values are bound to recorded publisher responses or identified browser transcripts. |
| Coverage disclosure | Pass | All 26 groups and operational/historical/physical-edition exclusions are explicit. |