A year of Aurora’s Facebook posts

Classification

Every record is processed using its full exported caption. A rule-assisted pass identifies electoral language; contextual review of distinct-caption excerpts, candidate matches and ambiguous posts supplies recorded overrides. This is not an exhaustive independent human reading of every full caption or a review of image/video content.

One primary category is assigned where text supports it. Substantive electoral content takes precedence in mixed civic/election posts. Civic services, public events, local-business promotion, community greetings and mayoral duties are Town related. Family, hobbies and personal experiences are Private. “Private” describes the subject matter, not the Facebook visibility setting.

Incidental election-duty mentions occur earlier: November 18, 2025 in the Town’s Clerk recruitment post and February 9, 2026 in the mayoral appointment posts. These staffing posts are excluded from the substantive-election start dates.

Unclassifiable records have category: null and category_id: null. This is missing evidence, not a fourth substantive category. Confidence labels are qualitative review signals, not calibrated probabilities.

Duplicate detection

Exact: Unicode normalization, case folding, whitespace collapsing and typographic quote/dash normalization. Words, URLs and numbers remain. Captionless records are never matched.

Near: word-sequence and three-word-shingle candidate matching, followed by review of changes. The initial high-similarity proposal requires ≥90% sequence similarity, ≥84% shingle Dice similarity and unchanged number sequences; documented review can accept or reject a proposal. A small number of shorter, clear wording variants are explicitly accepted.

Partial: four reviewed caption-pair links containing substantial reused passages or shortened copies. Related topics and rewritten event coverage are not automatically duplicates. Group merging requires every variant pair to have an approved relationship; similarity chains cannot silently bridge unrelated captions.

Percentages and unique content

This post-weighted measure is directional. Multiple matches do not count a post twice. A separate symmetric Jaccard percentage uses shared distinct groups ÷ the union of distinct groups. Counts of “unique” distinguish distinct captions, page-exclusive captions, exclusive post records and globally single-occurrence posts.

Missing captions are included in full-post denominators, but excluded from distinct-caption and uniqueness counts. Short identical captions can indicate generic wording rather than the same underlying media; those groups are flagged.

Scope and data quality

Only the three supplied exports are analysed. No Facebook pages, linked articles, images or videos were downloaded. Statements within posts are source content, not independently verified facts. Export completeness cannot be established; a zero-post day may reflect no posts or missing capture.

One near-copy discrepancy: otherwise matching State of the Town captions posted on February 18, 2026 name different event years: 2026 on tom.mrakas and 2025 on mrakas4aurora. Both original texts are preserved; search “sold-out room” to compare them.

The earliest election date is an observed lower boundary within these files, not proof of when all campaign activity began. Text duplication and posting proximity alone cannot establish staff involvement, content ownership, public-resource use, audience reach, engagement or legal compliance.