An RL-ish agent skill that pushes back on your reasoning and learns which challenges work for you.
An RL-ish agent skill that pushes back on your reasoning and learns which challenges work for you.
Challenger is an RL-ish agent skill: it pushes back on your reasoning and learns from your hit / miss feedback which kinds of challenge work best for you.
The mechanism: a local contextual bandit scores proposed candidates over seven cognitive axes, selects one with Thompson-style exploration noise, and updates a linear reward model. State is per-axis-value weights and sample counts on disk.
| axis | what it measures |
|---|---|
| granularity | Scale of what is challenged. word · claim · argument · frame · purpose |
| direction | Where it points: into their reasoning, out to missing context, lateral alternate framing, orthogonal against the premise, or temporal (origin / consequence). inward · outward · lateral · orthogonal · temporal |
| distance | How far from their current position. adjacent · reframe · inversion |
| form | Surface delivery. question · counterclaim · analogy · counterexample · thought_experiment · quiet_prompt |
| demand | Cognitive move asked of the user. defend · specify · compare · predict · commit · imagine |
| meta |
Content (none) vs their thinking about it — process, motivation, pattern, or why they are thinking about it at all.
none · process · motivation · pattern · activity
|
| register | Epistemic mode. factual · analytical · analogical · speculative · imaginative |
Each turn: query under-sampled axis values, generate three candidates that differ on multiple axes (at least one forced into exploration), self-tag, call bandit for selection. The score is the sum of learned weights on the tagged axis values, plus exploration noise scaled by 1 / sqrt(1 + count). Only the winner is displayed to the user.
challenger-hit and challenger-miss apply a gradient step to every axis value of the last challenge. The prior is flat per user and the state stays local.