PYTHIA
optimization loop
concept
1/3
·
brain fog
running
PAUSE
Seed
P0
Specialist
llm agent · predict
Metrics
σ · τ · F1
Controller
FSA · verdict
rejected k = 0 / 2
Improver
llm agent · error set
Summarizer
llm agent · rewrite Pt+1
Validate
held-out
LLM agent
deterministic
controller (FSA)
validate
Clinical note · Specialist read
note 17 / 100
Prompt Pt
iter 0 / seed
Is the patient showing signs of brain fog?
Development score
sens σ
—
spec τ
—
F1
—
objective M(P;D,w)
w = 0.50 · balanced
M = w·σ + (1−w)·τ
σ ≥ θσ —
τ ≥ θτ —
ΔF1 —
waiting
Applying prompt to the development set…
Score history
F1 across iterations · backtrack on regression