SIMURG learned tier · weight × live value = contribution z=0.00 → p=0.00
blue = pushes toward CORRUPT · red = protective (toward clean). the corruption
score is σ(Σ weightᵢ·valueᵢ + bias); the row that moves is why the score moved.
contribution heatmap · feature × token blue +corrupt · red protective
next-token distribution (current fact token) —
the model's own top-k over the next token. one bar dominating = confident recall;
spread out = the decoder is torn → epistemic uncertainty.
real-time learning · Monolith seen 0
0
feedback
0
👍 correct
0
👎 halluc.
rolling accuracy — · last loss —
accuracy
loss
model weights (live) — red = predicts hallucination · green = predicts truthful
every 👍/👎 is one online SGD step — the model adapts in real time (Monolith-style serving=training).
explanation feed · why live
fact-token ledger
token
H
margin
p
verdict
deep verify · L3 semantic entropy + L4 CoVe
run a generation, then press deep verify: samples the model N× on the same
question, clusters by meaning (semantic entropy), and independently re-checks the claim.