Measure long reasoning spans as a board-visible cost
A dwell-time style readout for long reasoning spans, visible on the board so attention cost is not invisible.
Pain. Attention spent on long, uninterrupted model runs is easy to under-report in telemetry: merges still happen, but dwell and context load stay invisible. Interface-level studies need proxies for sustained cognitive engagement beyond token counts.
What changed. A sustained-reasoning window metric with board readout marks extended reasoning intervals as observable series rather than anecdotal operator memory.
Why it matters. Identity and attention constructs in the research program assume operators can be overloaded without declaring it. Instrumented dwell makes “long run” a row-level fact for correlation with defects and halts.
How to try it
- Run assignments that legitimately hold context across a long reasoning span and inspect the readout on the board.
- Compare alongside session reports for the same ASN when exporting data.
Shipped in 4b87f81 — feat(DP-88): sustained-reasoning window metric and Board readout (#43)