This preprint introduces the Verifier Frontier Observatory (VFO), a standing instrument for measuring how much of artificial intelligence research can be automated and for bounding how close the field is to recursive self-improvement. The argument that AI is approaching recursive self-improvement rests on one unproven premise: that research taste, the work of choosing which problem matters, which result to trust, and when an approach is dead, will fall to scaling as prior capabilities have. This work argues that the premise can be measured rather than asserted and specifies how to measure it. The instrument rests on a single mechanism: AI attains superhuman performance where a usable surrogate verifier exists, and largely fails where one does not. Verifier usability is formalized as a multiplicative Verifier Coverage Index (VCI) combining validity, cheapness, timeliness, and robustness. Its complement, the Verifier Floor, is the structurally non-estimable residual of research tasks that no verifier covers. The Observatory tracks both quantities per model generation and pre-registers a falsification criterion, a tripwire, that overturns the author's own thesis the moment low-coverage tasks begin improving at the rate of high-coverage ones. The instrument runs retrospectively on existing public benchmark records and requires no privileged data and no frontier compute. Its output is not a single date for recursive self-improvement but a tracked bound: the size and location of the human-research floor that compute cannot remove. This work extends the identifiability-uncertainty construct of the author's Traceability Ledger monograph (DOI: 10.5281/zenodo.20480119) into the domain of AI research automation. It is a pre-registration: the metric, factors, predictions, and falsification criterion are fixed prior to data collection.