TokenKit · Decision model exploration
Decision models and flattened expertise
Directions for four ideas. Each card is one way to draw it; pick, mix, or push further.
The decision model tandem
A fast, typed decision layer handles the high-confidence cases. The language model is only woken when confidence drops below a threshold.
Decide fast, write only when needed
Every case gets a fast, typed decision. Above the threshold the system acts; below it, the language model is woken to write.
Most cases clear the bar
The decision layer scores every case. The dense cluster above 0.9 is acted on; the scattered few below it route down to the language model.
Fast for the many, careful for the few
Width reads as share of cases (illustrative, not measured). One threshold splits the stream: most go straight to action, a few get language.
Where the two kinds of system meet
Probabilistic and deterministic systems overlap. The checks, thresholds and recursions live in that overlap, and placing them is the skill.
The skill lives in the overlap
One side proposes, the other verifies. Seeing the overlap, and placing checks, thresholds and recursions inside it, is the scarce skill.
Two bars, one handoff zone
Each kind of system has a stretch it is good at. Where they overlap is where the checks go, so the handoff is designed rather than hoped for.
Where each check sits
A run alternates between kinds of step. The judgment is in where to put the check, the threshold, and the loop back, not in any single step.
The through-line: efficiency gets absorbed
As computation gets cheaper, the relative cost of human coordination rises, and organizations fill the freed-up space with more process.
Each gain climbs into the next cost
Cheaper output does not stay saved. It is spent on more output, which needs more coordinating, which grows the structure that does the coordinating.
Use outruns the efficiency gain
Shape only, not data: as unit cost falls, use climbs faster than the gain, and coordination cost follows it up.
The structure thickens to keep up
More output means more hand-offs between people; the hand-offs become roles, and the roles become layers.
Apprenticeship over prescription
Walk the path first, hit the failures, build judgment, then capture what worked as documentation of practice.
Walk it first, draw the map after
The path is learned by walking it. Failures along the way become judgment, and only then is it written down as markdown for the next person.
Prescription hands over a map
A playbook is a map drawn before anyone walked the road. Apprenticeship walks first; the markdown records what the road taught.
The markdown comes last, every time
Walk the path, hit the failures, build judgment, then write it down. The write-up is an output of the walk and the start of the next apprenticeship.