Traces / focused day
30–40
Fully judged & annotated by me, with tooling carrying capture and formatting.
The honest pitch behind the other three doors. The role asks one person to challenge frontier models, catch where they fail at literary reasoning, and feed that back as structured signal — at volume, without lowering the bar. This is how I keep both: my own AI Air Team handles the routine half — surfacing candidate passages, drafting prompt variants, logging traces — so the human half (the close reading, the judgement on what's genuinely a failure, the metacognitive feedback) gets all of me.
The discipline of the whole system is one rule: tooling never decides what counts as a failure or what good reading looks like. It clears the runway. The human lands the plane.
Five stages from a blank passage to trainable signal. The coloured tag on each stage says who owns it — so you can see exactly where the human judgement lives.
An honest sketch of what this division of labour buys — illustrative figures for a focused day, not a promise. The number that matters is the last one.
That's the whole system. Tooling clears the runway so the human close reading — the part Meridial is actually buying — gets all of me, at a pace that's useful rather than precious. You've now seen all four doors: the materials, the engine, the rubric, and the pipeline behind them.