Loss DSL

Status. The parser in conceptmod/dsl.py is the live phrase language. The velocity truth table is ../dsl.md — a 2-D job scoreboard, not a promise that every glyph wins on a GPU. This folder is the catalog, the worked phrases, and a GAN-game layer that runs on the same 2-D fixture. A formulation PASS from loss_game is a CPU toy. It is not a Music or Anima GPU transfer.

Install from a checkout (Python 3.11+). particlegan is a dependency; the game calls it for GANLoss and GradRegularizer and does not vendor them.

python -m pip install -e ".[dev]"
pytest tests/test_dsl.py tests/test_dsl_examples.py tests/test_game.py tests/test_dsl_catalog.py -q

Write a phrase, parse it, see the loss, score it

red=blue is one write rule. The loss is ops.rule_loss: trained velocity at red is pulled toward frozen blue. On the CPU fixture that is a remap of the color axis.

from conceptmod.dsl import describe_phrase, parse_phrase
from conceptmod.game import loss_game

phrase = "red=blue"
rules = parse_phrase(phrase)
print(describe_phrase(phrase))

match = loss_game(phrase)     # student vs critic, adv="locked_shared"
print(match)
row = match.score()           # one CPU step on the 2-D fixture
assert row["verdict"] == "PASS"

parse_phrase returns a Rule (op="write", a="red", b="blue", alpha=1). describe_phrase says this is a remap, not a global boost of blue. loss_game builds a Game: the student minimizes that phrase plus the locked RpGAN generator term; the critic minimizes RpGAN logistic d_loss plus GradRegularizer b_cap. The stamp is locked_adv_defaults() — RpGAN logistic, b_cap coeff=1 κ=1, FM off. The step does not sample the toys' 12-particle cloud and does not apply demo cover. Those fields are recorded so the game cannot quietly fork them.

row["verdict"] == "PASS" means that wiring matched. It does not re-score the geometric table in ../dsl.md. Run that table with:

pytest tests/test_dsl_jobs.py tests/test_2d_analysis.py -q
python scripts/analyze_2d.py --jobs --out outputs/2d_analysis

The same phrase as a stanza:

from conceptmod.game import game

match = game("""
phrase red=blue
pairing locked_shared
claim locked
""")

print(match) is the shape table for this cut:

Game  phrase='red=blue'  claim=locked  pairing=locked_shared
fixture  TwoAxisBackend  color=red/blue  pattern=stripe/dot
stamp  loss=logistic mode=rp arm=b_cap coeff=1 kappa=1 fm=0  source=locked_adv_defaults
recorded  n_particles=12 cover=1.5 posture=demo  (toys own these; this step does not sample a cloud)
applied  pairing=locked_shared kappa=1 fm=0
player   role       objective
student  generator  phrase rule_loss + RpGAN g_loss
critic   critic     RpGAN d_loss + GradRegularizer b_cap
index  op           a                 b                 alpha
0      write        red               blue              1

Stranger pairing, fm_weight != 0, and kappa != 1 raise LockedClaimError when the game claims locked. Built with claim="drift" they take one step and score FAIL. Details: games.md. What the next stanza keys are not allowed to pretend: SPEC.md.

Where to read next

Velocity truth table	../dsl.md
Operator catalog	operators/index.md
Worked phrases	examples/bipolar.md, keep+erase, remap, replace, locked game
GAN game	games.md
Add an operator	ADDING_OPERATORS.md
What this layer owns	SPEC.md
2-D pictures	../2d-analysis.md
Formulation stamp	../formulation-toys.md

Glyphs the parser accepts: c++, c--, a=b, a#b, a%b, a~b, a^b, ;c, plus :alpha / :key=value, |, and {random_prompt}. @ is stripped and ignored. Do not add a synonym for a job the scoreboard already calls a recipe.