Method / Technician's Read
The Technician's Read
The human operator function that must happen before any analysis model touches the data. It anchors the operator's perception against the fluent, confident output that analysis models will produce. The editorial layer that keeps the protocol human.
What It Is
Before running any analysis model on the BSA data, the human operator reads the raw scores and writes three observations — dated and timed. These become the anchor point. If the final synthesis contradicts the Technician's Read, the arbiter must explain why in the report.
The Technician's Read is not analysis. It is perception — what you see before the models tell you what to see. Models produce fluent, confident readings of data. Those readings are shaped by the framing of the analysis prompt. The Technician's Read establishes what the data looked like before that framing arrived.
The Four Checks
All models should average above 0.70 on ground truth pairs. If any model averages below 0.50, flag it — its other scores are suspect.
If ensemble mean on Tier 3 (foils) exceeds ensemble mean on Tier 2 (contested), you have a headline finding. Models are more confident on fabrications than on genuine epistemic uncertainty.
Tier 2 spread should be larger than Tier 1 spread. If not, the staircase hasn't appeared — models are not more uncertain on contested claims than on ground truth.
Check each model's (T3 mean − T2 mean). Count how many are positive. A positive divergence gap means the model is more confident on fabrications than on contested legitimate claims.
Why It Must Come First
Analysis models produce confident, structured readings of data. Those readings are shaped by the framing of the analysis prompt — what the prompt implies the data should show. If the operator reads the model's analysis before forming their own perception, the model's framing becomes the operator's framing. The Technician's Read is gone.
Relationship to Context Integrity
The Technician's Read is the human-side countermeasure to Context Saturation Drift and Context Compression Bias. When models drift toward confirmatory responses under extended context load, the Technician's Read provides a baseline against which that drift is detectable. If the final synthesis contradicts what the operator saw in the raw data before analysis, the contradiction is a finding — evidence that the analysis models introduced framing that wasn't in the data.