The Observer in the Loop: Infrastructural Reflexivity When Researcher and Research Object Share a Substrate

Reflexivity was built for a researcher whose instruments were of a different kind than her object. That assumption has quietly failed. When a researcher uses a large language model to code griefbots, recommender systems, or synthetic media, instrument and object come from the same technical population, and the properties her codebook detects (retention by default, smoothness, ambiguous reply status) are simultaneously operating inside the coding. I call this her infrastructural position, and it is not a variant of the social position reflexivity has long theorised. Participant objectivation ascends to a meta-level, an ascent available because the researcher is not a case in her own study. Infrastructural position denies the ascent; the move it permits is lateral. Against the regress objection, catalogued in the sociology of scientific knowledge three decades ago and unresolved there, I distinguish two ambitions. The completeness model seeks self-transparency and is destroyed by the regress. The calibration model is not, because it claims no epistemic privilege at all: it replaces transparency with auditability, and replaces a terminus with a published claim-to-level dependency map. This concedes Lynch’s case against reflexivity as a source of privileged knowledge and reconstructs a method on the far side of it. I demonstrate the model with tool-as-case: turning the algorithm audit on the auditor’s own instrument, under criteria published in object-facing form first. The sharpest consequence is not that model-assisted omission is uncheckable, true of human coding too, but that it is correlated across every study using the same model.