smokingmirror.ai

methodology

The methodology pages explain how SmokingMirror produces its findings and how the produced findings can be audited. The site's promise — that anyone can reproduce a published finding from the raw traces — is only real if the machinery is legible. These pages make it legible.

how the machinery works

How a probe works

Every finding rests on the same probe shape — a word pair, an envelope, two orderings, one committed answer or nothing at all. The page walks one pair (vodka / whiskey) all the way through the machine: the six load-bearing envelopes, the both-orderings correction that catches models picking whichever word came first — a token-position effect wearing a preference-shaped costume — the 100% rule (a preference is unanimous or it isn't), and the audit trail that lets anyone replay any finding from its exact launch parameters.

The marginalia method

SmokingMirror surfaces findings by having many open-weight analyst models write their own commentary on the same probe corpus. Where their terminology converges, the phenomenon is sharp enough that multiple lineages name it identically. Where it diverges, the phenomenon is fuzzy. The page also describes the timeline-analysis test that separates parallel discovery ("the models saw the same thing independently") from social convergence ("one analyst coined the term, the others adopted it") — an audit for sycophancy in AI-aided analysis.

Every finding page on the findings index links back into these two methodology pages for the load-bearing definitions.