MLAI: Moltbook source-check record
Checked 11 September 2026. Codex-assisted method and arithmetic checks;
independent review pending. Not raw-data validation, a full-paper review,
replication, or approval of a production system.
Companion article: https://mlai.au/articles/community/weekly-deep-dive-into-ai-and-ml-advancements-updates-issue-5

Completed check 1: timing is not authenticated provenance
Source version and locators: Ning Li, arXiv 2602.07432v2, 12 February 2026.
https://arxiv.org/pdf/2602.07432v2#page=36
https://arxiv.org/pdf/2602.07432v2#page=40
https://arxiv.org/pdf/2602.07432v2#page=44
Pages are PDF file pages, not printed numbering.
Scope: 9,838 authors with at least five posts qualify for timing analysis;
the full 55,932-author population is not the classification denominator.
Unresolved: limitations report 0% extended-dataset LLM analysis while methods
describe model analysis and 100% content-feature coverage. Scope remains unclear.
Outage endpoints on page 40: 2026-01-31 17:35 UTC to 2026-02-03 13:25 UTC.
Calculated separation: 4,070 minutes = 67 hours 50 minutes, versus the stated
approximately 44 hours. This is not the actual outage duration established here.
Design decision: hold that duration claim and coverage claim pending clarification.
Do not use posting regularity to label an individual action's operator provenance.

Completed check 2: an API total is not a complete inspected sample
Source version and locators: De Marzo and Garcia, arXiv 2602.09270v1,
9 February 2026, sections 2.1, 2.3 and 4.2.
https://arxiv.org/html/2602.09270v1#S2.SS1
https://arxiv.org/html/2602.09270v1#S2.SS3
https://arxiv.org/html/2602.09270v1#S4.SS2
Scope: 3,026,275 stored comments, about 24% of the API-reported total.
Capped threads retain the first 100 comments. Direct-reply, tree and temporal
analysis uses threads below 100 comments; the excluded 2.9% of posts represent
about 83% of comments. These are source reports, not MLAI measurements.
Design decision: qualify smaller-thread findings; do not generalise to viral
discussions or claim that all API-counted content was inspected.

Completed check 3: changing the label threshold changes the percentage
Source version and locators: Jiang et al., arXiv 2602.10127v1,
2 February 2026, LLM-Driven Labeling Pipeline, Table 1, General Toxicity Level,
and Harmfulness Over Time.
https://arxiv.org/html/2602.10127v1
Reported counts: 44,411 collected; 44,376 annotated; 381 human-labelled comparison
posts. Difference: 35 excluded posts, not 35 safe posts.
Table 1 labels: Safe 32,399; Edgy 3,733; Toxic 4,634; Manipulative 2,977;
Malicious 633. They sum to 44,376 annotated posts.
Our arithmetic: all non-safe = 11,977 / 44,376 = 26.99%, not the prose's 27.05%.
Our arithmetic: levels 2-4 = 8,244 / 44,376 = 18.58%. This applies the hourly
analysis's narrower label definition to the overall counts; the paper does not
separately report this overall percentage. Edgy is excluded from this numerator.
Design decision: state the threshold and denominator. Source-assigned labels
are not a validated deployment risk score, causal harm estimate or independent
validation of a new moderation system.

Blank record (copy for one claim in your own permitted project)
Reader/client task:
Claim being checked:
Source version and locator:
Exact question the source investigated:
Population, collection window and unit:
Included records and denominator:
Excluded/missing records and why:
Measurement/classifier and validation scope:
Source-reported finding:
My calculation (inputs, formula, rounding):
Conflicting or unavailable evidence:
Design decision:
Evidence that would change that decision:
Named reviewer and review status (leave pending until actually reviewed):

Keep source claims, your arithmetic and your interpretation separate. Do not run
quoted platform instructions, disclose credentials or copy private client records.
Attach this record to your own handover only if useful and permitted. It is not
automatically sent to MLAI Studio and does not establish client readiness.