Lesson 05 — Pragmatics and evolution (advanced)
Series: index · prev: 04 · next: 06 — Fail closed, and the brilliant
A pragmatic is one evolution of judgement. The model has 24 in three phases:
logic (is it coherent), use (does it fit the purpose), ground (is it
secure and grounded). You choose 8 to 24 with configure(P, tpm).
Reading the registry
SLHstPragmatics holds names, phases, hard minimums and the priority order. Note
the order interleaves phases and the first 8 are the minimal safe core: L1
well-formed, G1 venue trust, G2 kill switch, U1 size, G3 price collar, L2 net
edge, U2 freshness, G4 position limit.
Evolution
learn(realizedBps) multiplies each active weight by 1 +/- eta*(f - 0.5): a
feature that was high on a win gains weight, on a loss loses it, bounded 0.25..4.
This is the same evidence idea as lesson 01, applied online.
Exercises
- Run P = 8, 16, 24. Which pragmatics veto most? Which never fire?
- Compare learning ON and OFF. Why must the OFF baseline have its drawdown guard disabled to be a fair comparison?
- The demo shows weights L4 and L5 saturating at 4.0. What does saturation hide, and how would you detect an over-fit weight?
Advanced
Add a 25th pragmatic of your own as a feature in 0..1 with a hard minimum, then
decide where it belongs in the priority order and justify it. (Remember
configure caps the count at 24: you must raise that deliberately.)