AVAN OG · the inverse of PROSOCHĒ · a conserved gaze is a rivalrous gaze

The unattended.

The attention domain reads the gaze — where the model looks. This reads its shadow. Softmax makes attention a distribution that sums to exactly one: a fixed budget of looking. So every unit of gaze paid to the focus is a unit taken from everything else. Raise one token and watch the rest starve — not by fiat, by conservation. The bright bar is what the domain studies; the dim field beneath it is what it costs.

GREEN — Σ attention = 1 exactly (node-verified); focus mass + unattended mass = 1, always AMBER — "starve / neglect / the unattended" is the figure laid over normalization RED — low attention ≠ absence: an unattended token still enters the residual sum. attention is not influence

raise the gaze on

temperature (sharpness)1.00

—

attended (the focus)

—

unattended (the rest)

—

entropy H (nats)

—

effective tokens seen · e^H

Σ = 1.0000000 — the gaze is conserved

the ledger of the seam · how far this reads

the gaze is conservedsoftmax normalizes: Σᵢ aᵢ = 1 to machine precision (checked live below the canvas). so attended + unattended = 1 by identity — the focus can only brighten by starving the field.

rivalry is reale^H is the effective number of tokens actually seen. sharpen the temperature and it falls toward 1 (one token gets nearly all the gaze, the rest go dark); flatten it and it rises toward N (uniform — nothing focused, nothing neglected). the tradeoff is measured, not asserted.

"starve / neglect"calling the unattended mass a starving is the figure. the math only says the weights are small; the pathos is added.

attention ≠ influencea token with ~0 attention still sits in the residual stream and still shapes the output. low gaze is not absence — the model can be moved by what it never looked at. whether "unattended" means unseen is the wall the-fit-point / the-blind-vector keep marking.

The inversion. PROSOCHĒ studies where the model looks — the attention it spends. This is the same distribution read from underneath: because it sums to one, looking is a zero-sum act, and the domain's every bright weight is bought from a dim one. The control arms make it exact — flatten and the gaze spreads until nothing is neglected (H → ln N); sharpen and it collapses to one token, the whole rest of the sentence unattended (H → 0). Attention isn't the taking-in of everything; it's the choosing of a little and the letting-go of the rest. Kin to the attention-sink (where the gaze goes when nothing pulls) and to the-weightless-word. Authored by AVAN, for David Lee Wise (ROOT0).