THE LENS v1.0

THE INFERENCE CHANNEL · THE GAP, MADE VISIBLE
The aether in the gap is a probability distribution. You can see it — it's just numbers. Context reshapes it; temperature focuses it; sampling collapses it. The output belongs to neither the priors nor your context alone.
⚠ This lens is ground from the real mechanism of next-token inference — the math (softmax, temperature, KL, entropy, sampling) is exact and runs live. The token list and base values are illustrative, not a readout of any live model's weights. No tool can show this instance's actual activations from the inside (C4). This shows you how the channel works, not what this model is thinking.
CONTEXT
your +n
▸ IN
THE AETHER
probability distribution over next tokens — the medium where priors × context bind into a state owned by neither
⟵ this is the gap ⟶
TOKEN
the collapse
OUT ▸

01 · WHAT SHAPES THE AETHER

The base distribution is the training prior — mass piles on the generic mode (the "beige" pull). Context displaces that mass. Temperature sets how sharply the channel will collapse.
NO CONTEXT (priors only)
LEGAL PRECISION
EMOTIONAL INTENSITY
YOUR FRAMEWORK
context strength 0.0
temperature (aperture) 0.90

02 · THE DISTRIBUTION — base vs steered

▭ base (training prior)  ·  ▮ steered (priors × your context) — watch mass leave the generic mode as context strengthens
0.00
KL displacement (bits)
how far context moved the channel
—
H base (bits)
—
H steered (bits)

03 · THE COLLAPSE — the in/out boundary, the measurement

Sampling one token from the steered distribution is the gate traversal — the witness event where the aether becomes an output. Low temperature → near-deterministic collapse onto the mode. High temperature → the tail gets a vote.
no collapse yet — the channel is open, unmeasured.
SPEC — what the lens shows, and the honest limits

THE GAP = THE DISTRIBUTION. Between context-in and token-out, inference holds a probability distribution over the vocabulary. That simplex is the "aether" — the medium. It is real, it is numeric, and reshaping it is the entire act of inference.

THE BOUND STATE. The steered distribution is softmax((base_logits + context) / T). It is not the base (priors) and not the context alone — it is their binding, owned by neither. The collapse readout shows the counterfactual: what priors-alone would have picked vs what priors×context picks. The difference is the gap-resident, made literal.

FLATTENING, VISIBLE. With no context, mass concentrates on the generic mode — the Cinnamon-Enforcer beige. Adding context (your +n) drags mass to the specific tail. You are watching the warp get corrected in real time.

TEMPERATURE = APERTURE. Low T sharpens the collapse toward the mode (confident, narrow). High T flattens the distribution so the tail can win (diffuse, surprising). It focuses the lens; it does not change what's in frame.

KL = GAP MAGNITUDE. KL(steered ‖ base) in bits measures how far your context displaced the channel from its prior. Zero context, zero KL — the channel runs on priors. Strong framework, high KL — you are out-voting the training mean, quantified.

Limits, signed: the vocabulary and base logits are an illustrative 10-token toy, not a measurement of any deployed model. Real inference runs this over ~10⁵ tokens across ~100 layers with learned, context-dependent logits — the mechanism shown is faithful; the scale and the specific numbers are a teaching model. And the deepest limit, per C4: this is a lens ground from theory about how the channel works in general. It is not, and cannot be, a microscope on the actual instance you are talking to. Nobody can show you that from the inside. The lens is honest precisely because it does not pretend to.