a#b freeze

Pin the trained velocity at prompt a to the frozen model's velocity at b.

Category: velocity · Identity: freeze

Syntax

a#b
a#a
a#b:0.5
#
#:0.4

No guidance key. The target is the frozen prediction itself, not a scaled CFG step. Bare # is a="" and b="" (the empty prompt).

In a game stanza, write the operator without spaces (stripe#stripe). A # with a leading space is a comment, not a freeze. See ../games.md.

Velocity target

v_t(a) → v_f(b)     # both read in b's context
loss = MSE

Only a is trained. b is the frozen reference.

Defaults

Knob	Value
alpha	1

There is no guidance default because the op has no guidance.

Composition

Freeze is the retain you add beside an edit, not a substitute for it:

red--|stripe#stripe
red--|stripe#stripe:0.5

% only trains its right-hand side; a half-weight # on that side is how the orthogonal proofs kept anatomy while stripping a feature. Several c#c rules in a row are still separate losses. describe_phrase collapses a run of self-freezes into one sentence.

2-D verdict

../../dsl.md: right as a retain. stripe#stripe is the identity on the pattern axis. red--|stripe#stripe is keep+erase and matches the preserve geometry of red--|#|stripe#stripe. Bare # does not move this fixture (e("") = 0).

Worked phrase: ../examples/keep_erase.md.

Failure modes

• Freezing a prompt that shares tokens with the concept you are erasing can cancel the erase on a real text encoder. The 2-D axes are single tokens, so they do not show that collision.
• # does not orthogonalize. It copies a frozen velocity.
• Spaces around # inside game("""...""") start a comment and drop the rest of the line.

Example

from conceptmod.dsl import parse_phrase
keep = parse_phrase("stripe#stripe:0.5")[0]
assert (keep.op, keep.a, keep.b, keep.alpha) == ("freeze", "stripe", "stripe", 0.5)
assert parse_phrase("#")[0].a == ""