# Livius bibliography review apparatus

These are the main editorial reviews accompanying the metadata. Internal JSON
filenames refer to the retained review packet, not additional public downloads.
The public CALIBRATION.json retains all source grounding and locators while
omitting private response paths. Raw source bodies are excluded. The later
Aquileia label applies the reviewed byline and stub status; its original label
is preserved in ATLAS_CITATION_CORRECTIONS.json. Historical acceptance is unchanged.

# Livius metadata — Fact-Check and Revise

Main review, 2026-09-09. Fact-Check and Revise v2.8.

## Intake and findings

Venue: Craig's own research site and bibliography downloads; model-assisted
metadata composition is authorized. This review covers the nine candidates in
IDENTITY_DRAFT.json, not a rewrite of the monograph or historical case records.
Sources were fetched and read in this session. Existing independent-review
requests concern separate historical/statistical work and remain pending.

**Correct.** The nine requested pages identify the named ancient texts or the
Aquileia reference article. Their displayed titles, contributor credits and
web-page dates support the proposed metadata. The page numbers in the Appian
URLs identify web parts; the separately displayed ancient sections are retained.
The nine records cover ten current citation IDs, eight cases and 29 original
occurrences. These counts describe the existing joins, not nine unique works.

**Overstatement avoided.** Livius credits Horace White for the Appian translations
and Jona Lendering for notes; those roles do not make Lendering the ancient author.
The Periochae are summaries of Livy's history, not surviving full books 86–90.
The linked introduction distinguishes the epitomator from Livy and credits
Lendering's translation. Aquileia is explicitly a stub. The test-site Periochae
view is also marked as a stub and presents Book 89 in a different order from the
main-site page. Its distinct access record and dates remain intact.

**Unverifiable fields.** Four Appian records do not identify the exact print
edition underlying the online Horace White text. Their underlying_print_edition
fields stay null and are listed in FIELD_EXCEPTIONS.json. The Zosimus page reports
an anonymous translator and only a probable preparer, J. Davis; that qualified
statement must not become a certain translator credit. Its 1814 London imprint
is reported by Livius and corroborated by the linked Tertullian transcription
heading; no scan of a printed title page was inspected.

**Source quality.** ABC 9 describes its translation as adapted from Grayson
(1975) and Glassner (2004). Those are reported modern source editions, not named
ancient authors or proof that both printed books were inspected. The page does
not name its adapter. Ancient authorship, modern adaptation, notes, hosting and
web dates stay separate. Appian's site introduction warns about numbering
differences between editions. No source-independence or funding conclusion is
drawn from these pages. Notices and limitations relevant to their use were read.

**Structure and register.** The old recorded labels mix titles, passage locators
and purpose notes. They remain unchanged. The typed metadata is an additional
view; accurate purpose notes are not declared errors simply because they are
not bibliographic titles. No causal, normative or population claim is introduced.

**Attribution.** Page dates are creation/modification of the web presentation,
not dates of ancient composition or first printed publication. The source's
tentative 346 BCE Sidon date is already covered by the current case's explicit
345/346 calendar-conversion note. Suessula already distinguishes the general
Sullan settlement program from independent confirmation of the local event.
Neither observation warrants a duplicate hold or new statistical decision.

## Grounding and revision

SOURCE_OBSERVATIONS.json binds the nine requested pages and two linked context
pages to four private tool-response captures. The headers, credits, page dates,
numbered sections and relevant limits were actually read. INPUT_BINDINGS.json
and BOUND_OCCURRENCES.json bind the current records and their joins. No source
claim relies on model memory. The revision is the separate typed metadata in
IDENTITY_DRAFT.json; original IDs, URLs, purpose notes and occurrences are retained.

The source URLs and precise locators travel with each candidate. Private response
captures are not raw HTML archives. The first public Save Page Now request timed
out; no archive URL was confirmed, and remote completion is unknown. The request
was not repeated. ARCHIVE_STATUS.json keeps that limitation explicit: current
access verification is not durable public archival verification. These findings
and self-audits accompany any later metadata publication.

## Required self-audit

| Required check | Disposition | Evidence or reason |
|---|---|---|
| Grounding | Pass | Nine original pages and two linked context pages were fetched and read; bindings record the text actually consulted. |
| Required vs. offered | Pass | Typed title improvements do not classify faithful existing purpose notes as false. |
| Claims | Pass | Historical narratives and scores are unchanged; no underlying edition is guessed. |
| Deletions | Pass | Both Periochae access records and all 29 occurrences remain. |
| Criticisms | Pass | No external reviewer finding or replacement text was adopted. |
| Internal consistency | Pass | Online component identity, ancient section numbering and print-edition identity remain distinct. |
| Register discipline | Pass | Descriptive metadata does not adopt the source introductions' evaluative praise. |
| Moral framework register | Pass | No new normative claim is present. |
| Political-spectrum contamination | Pass | No political placement label is introduced. |
| Attribution posture | Pass | Ancient author, translator, notes, adaptation and host are separate fields. |
| Quotations | Pass | Titles and short credits follow the consulted source strings; paraphrases are not enclosed as quotations. |
| Causal reasoning | Pass | No new historical cause or motive is inferred from a title, date or contributor credit. |
| Categorical vs. dimensional | Pass | Identity status does not imply complete historical evidence or binary refugee motives. |
| Denotation/connotation and convenient layer | Pass | A web part is not treated as an ancient section; a host credit is not authorship. |
| Proper-name variants | Pass | Supplied initials, full names and probable/anonymous credits remain explicit. |
| Engagement posture | Pass | Existing historical argument and source-use notes remain intact. |
| Confidence calibration | Finding retained | Four underlying print editions remain unresolved; the Zosimus preparer stays probable. |
| Fetched-source disclosures | Pass with limits | Page notices, translator/adaptation caveats and stub status were read; no claim of source independence or full printed-edition inspection is made. |
| Citation resolution | Pass at candidate scope | Nine stable work IDs retain their original links and bindings; historical passage acceptance is unchanged. |
| Adopted reviewer text | Pass | No reviewer-supplied prose is present. |
| Editorial synthesis (SYNTH) | Pass | No new historical join is made; access-view relations are explicit editorial metadata judgments. |


# Livius metadata — LLM Artifact Tells

Main review, 2026-09-09. LLM Artifact Tells v2.7. Input:
IDENTITY_DRAFT.json and its required first-round companion notes. The metadata
is model-assisted work for Craig's own research site; this pass does not claim
human authorship or seek to evade a detector.

## Findings, in tier order

| Instance or scope | Finding | Tier | Class | Disposition |
|---|---|---|---|---|
| Candidate metadata and first-round notes | No hard artifact was found by the recorded searches; no DOI or ISBN was introduced. | 1 | Hard artifacts | No correction required; TELL_CHECKS.json records actual searches. |
| Dates, contributor roles and edition limits | No significance claim or source-count inflation is appended to the metadata. | 2 | Source integrity | Keep qualified and anonymous credits; leave four unidentified print editions open. |
| Grounding and archive-status notes | Process language explains the required review apparatus and its limits. | 2–3 | Context-dependent | Keep as companion apparatus, not historical case prose. |
| Repeated JSON keys and distinct role lists | Repetition follows the data schema. It is not a prose-style defect. | 3–4 | Context-dependent | Keep stable field names and separate roles. |
| Short metadata notes | Contrasts distinguish genuinely different things: web parts/ancient sections and web dates/print dates. | 4 | Context-dependent | No decorative reversal or rhythm-only rewrite required. |
| Titles and names | No ornamental intensifier or umbrella phrase needs removal. | 5 | Weak stylistic signals | Preserve exact supplied metadata. |

REVISED_IDENTITY_CANDIDATES.json is byte-identical to the draft. No stylistic
rewrite is needed. Source-integrity judgments remain in Fact-Check and Calibration;
this pass does not turn the historical claims into accepted findings. Raw
tool-response captures remain private research inputs rather than published prose.

## Required self-audit

| Required check | Disposition | Evidence or reason |
|---|---|---|
| Checklist vs. heuristic | Pass | Exact titles and schema repetition were judged in context rather than removed by vocabulary matching. |
| False positives | Pass | Real distinctions in contributor role, dates and numbering were retained. |
| Hard artifacts | Pass | Required string searches were run; no new DOI/ISBN or malformed metadata access date was found. |
| Tier coverage | Pass | Tiers 1–5 were examined in order, as reported above. |
| Decay weighting | Pass | Mechanical artifacts were checked first; weak stylistic signals did not determine provenance. |
| Guideline vs. style sort | Pass | Source-integrity issues remain substantive; optional style changes were not made mandatory. |
| Introduced tells | Pass | No rewrite was introduced; revised candidates equal the input bytes. |
| Writer's voice | Pass | Existing case and monograph prose were untouched. |
| Register homogenization | Pass | No essay was flattened into metadata or review prose. |
| Quality tradeoffs | Pass | No cosmetic edit weakens the specific role or date information. |
| Editorial synthesis (SYNTH) | Pass | No connecting historical claim was added. |
| Venue check | Pass | Craig's own research site accepts the authorized model-assisted metadata work. |
| Provenance honesty | Pass | The packet remains identified as main-agent, model-assisted review; no independent review or human-authorship claim is made. |


# Livius metadata — Calibration Review

Main review, 2026-09-09. Calibration Review v1.2, against
REVISED_IDENTITY_CANDIDATES.json and the source text recorded in
SOURCE_OBSERVATIONS.json. No independent reviewer was launched.

## Full audit table and coverage

The full table is [CALIBRATION.json](CALIBRATION.json). Every row has the claim
location, load-bearing terms, source URLs, exact grounding passages and locators,
what the source licenses, materiality, verdict and proposed wording. Its nine
Matched rows contain exact source spans; its four Uncheckable rows explicitly
identify the missing edition evidence instead of inventing a quotation.

All nine work/component groups were audited for titles, supplied contributor
roles, web-page dates, component/section distinctions and stated metadata limits.
The four unfilled Appian print-edition fields were audited separately. Counts:
**nine Matched, four Uncheckable, zero Overclaimed, zero Underclaimed and zero
Unsupported-but-sound**. All are detail-level claims about bibliography metadata.

Operational access timestamps, current-case joins and preservation counts are
checked separately against the local records. This pass does not certify the
ancient accounts' truth, independently corroborate events, re-adjudicate case
prose or scores, or assert that a public archive capture succeeded. The linked
Periochae introduction and Tertullian heading were read for their stated
metadata context; no complete printed edition was inspected.

## Concentration and reverse check

The gaps are confined to the exact printed editions behind four Appian web
components. Contributor names alone do not identify an edition. FIELD_EXCEPTIONS.json
retains those four null fields. The Zosimus anonymous/probable attribution and
the reported, rather than scan-verified, imprint remain qualified. No wording
change is required by this round.

The reverse check actively looked for underclaims in explicit title, translator,
notes and page-date statements. None was found. Clear credits remain clear;
uncertain or unnamed credits do not become certain names. No Unsupported-but-sound
verdict was used, so no re-cite flag of that type is required.

## Required review self-audit

| Required check | Disposition | Evidence or reason |
|---|---|---|
| Direction of changes | Pass | Nine groups are Matched and four fields remain Uncheckable; no blanket softening was applied. |
| Lane discipline | Pass | The round assesses source-to-metadata alignment, not prose polish or case classification. |
| Deletion vs. flagging | Pass | All nine records and the four missing-edition fields remain present. |
| Mandatory re-cite flags | Pass | No Unsupported-but-sound verdict was assigned; unresolved fields travel in the candidate limits and FIELD_EXCEPTIONS.json. |
| Verbatim grounding | Pass | Every Matched row carries exact source spans with locators; local substring checks verified the copied spans. |
| Coverage disclosure | Pass | All candidate metadata groups and four edition fields are listed; historical and operational exclusions are explicit. |