λ

LLM System Prompt Constraint Lab

Edit text to modify system context tokens
Active Constraint Rules & Logit Modifiers

Sampling Hyperparameters

0.70

Higher temp flattens logit differences, increasing entropy & risk.

0.90

Truncates tail probabilities below cumulative threshold $p$.

2.50x

Scales direct logit penalties for hallucination-prone tokens.

Compliance Score
0.88 / 1.00

Direct instruction adherence index

Token Entropy ($H$)
1.34 nats

Logit distribution uncertainty

Hallucination Risk Index
0.12 low

Out-of-context token divergence

Top Token Probability
42.0%

Token: " Context"

Next-Token Logit & Probability Distribution

Live softmax shift under applied prompt constraints

Adjusted Raw

Attention Allocation Matrix

How system tokens weight candidate output tokens

Prompt Directive Logit Shift Table

Numerical breakdown of raw logits vs. constrained post-softmax probabilities

State: Fully Computed
Candidate Token Raw Logit Constraint Delta Final Logit ($z_i$) Softmax Prob ($P$) Sampling Mask