BasisMacro-F1 of the unsupervised IOHunter-style detector re-implemented here, on the 6 published IO benchmark datasets, against the number the paper reports. A claim about this copy of the method, never a Kenya result.

5 of 6 datasets land within 0.2 Macro-F1 of the number the paper reports. The exception is Iran, where the benchmark disagrees with itself.

iran: the benchmark authors' own code, run on their own release, gives 71.31 where the paper says 60.83. Ours is 71.43, which is +0.12 from the reference code. Iran is judged against the reference implementation's own output (71.31), not the published 60.83, because the benchmark disagrees with itself.

Source: docs/analysis/v2-findings.md / Measured 2026-09-11

Basis88 posts drawn fresh, labelled by a human blind to both methods, and corpus-weighted. Precision: of the posts the filter calls Kenyan, the share that are. Recall: of the Kenyan posts, the share the filter finds.

The learned classifier finds about 83% of Kenyan posts at about 89% precision. The keyword gate has precision 0.973 and recall 0.516, so it misses about 48% of Kenyan posts.

Human labels, drawn fresh, judged blind, corpus-weighted. The honest framing is "the filter finds about 83% of Kenyan posts at about 89% precision", not "X% of the corpus is Kenyan".

Source: analysis/investigations/2026-09-12-relevance-classifier/findings.md / Measured 2026-09-13

BasisOne frozen snapshot, not the live run. Each method's communities were read blind by a model reader at a relevance floor of 0.5; cases are communities read, and the counts are how many that reader called political or Kenya-relevant. This describes what the method surfaced. It is not a rate: there is no denominator.

Neither method surfaced anything the reader called an influence operation. Snapshot 2026-09-05-promotion-off__emb20260912.

Model verdicts, 17 cases per method, read blind. A description of the method's output, never a prevalence rate - there is no denominator. The latest daily run (10 Oct 2026, 21:20 EAT) produced a listing. The listing is account level, so it is not published and its length is not shown.

Source: analysis/investigations/2026-09-12-component-ranking/findings.md / Measured 2026-09-12

BasisEach is a measurement of this project's own method failing, so no collection artifact can flatter it.

1. A high self-amplification score does not separate engagement pods from influence operations. Confirmed operations self-amplify more than the pods this project's detector found, so the filter was dropped.

   effect size d = +0.78, confirmed operations above pods

   Standardised mean difference of the self-amplification score, confirmed operations against the pods found on this corpus. Group sizes are not recorded in the cited document.

   Source: docs/OBJECTIVES.md (A3)

2. At a cosine cut of 0.85 in this encoder, most admitted post pairs are unrelated Sheng replies, not copies.

   59 of 100 sampled pairs unrelated; 26 same message; 15 same topic

   100 cross-author post pairs at or above 0.85 drawn from 279,060 behind the top 500 and labelled blind by a reader. Of all 279,060 pairs, 3% were near-copies and the median pair shared no words.

   Source: docs/analysis/v2-findings.md (section 4) / Checked 2026-09-11

3. A raw toxicity series over this corpus is not publishable: it moved because the collector changed what it collected, not because the discourse changed.

   raw series +78%; composition-standardised series flat then falling

   Baseline scope, July to August 2026. The baseline mix moved from 70.7% search and 12.1% replies to 14.3% search and 72.6% replies. The standardised series holds the mix at the earlier reference week.

   Source: docs/analysis/2026-09-13-publishable-statistics.md

Source: docs/analysis/2026-09-13-publishable-statistics.md / Measured 2026-09-13

Each of these is a number or a claim this project could produce and has decided it cannot defend. They are listed so the gap is visible, not so a reader fills it with an assumption.

• Any prevalence rate

  Nothing in the collector samples the discourse at random, so no rate over it has a denominator. A control arm is needed first.

• Any account, handle or per-account claim

  Detection says accounts act together; neither it nor a reader's judgement establishes that a named account is inauthentic.

• Counts of coordinated accounts or communities from the live run

  The listing length is a triage budget and the community count depends on a clustering resolution and seed. Neither is a finding.

• A toxicity series over time

  The raw series moved with the collector's own target list. Only a composition-standardised or within-partition rate may be shown, and none is shown yet.

• Cross-channel corroboration of clusters

  The earlier statement about it is false and was retired.