SPEC

This note is the boundary for the current game cut. It is also the list of stanza features that are not implemented. The parser will reject unknown keys. Do not document them as if game() accepted them.

What the loss DSL owns

• Phrase syntax and the parse tree (conceptmod/dsl.py): ++, --, =, #, %, the ~ macro, ^, ;, @, |, :alpha, :key=value, {random_prompt}.
• The velocity loss for each live op (conceptmod/ops.rule_loss) and the sentences in describe_phrase.
• The 2-D job scoreboard (dsl.md): which jobs are already recipes, which glyphs are dead on TwoAxisBackend, and the geometric words right / needs help / recipe.
• Compiling one phrase into a Game (conceptmod/game.py) whose student concept objective is those rules.

What the formulation toys own

• The locked_shared stamp: locked_adv_defaults() / locked_shared_floor.LOCKED. RpGAN logistic (GANLoss, loss_type=logistic, mode=rp), GradRegularizer b_cap coeff=1 κ=1 norm=l2, FM off, demo cover 1.5, n_particles=12.
• Refusals of FM-on and thinned κ: reject_unlocked. The game calls that helper when a locked claim carries those knobs.
• Stranger-as-vanilla, cover, leftover, particle clouds, and claim_pass. A geometric right is not a formulation PASS.
• feature_match in locked_shared_floor. The drift arm of the game calls that function. It does not redefine the term.

particlegan supplies GANLoss and GradRegularizer. Conceptmod does not vendor them.

What a game step actually does

1. Critic: frozen vs trained CFG plane points for the fixture probes, same order if pairing is locked_shared, fake row flipped if pairing is stranger. Loss is GANLoss.d_loss + GradRegularizer.penalty.
2. Student: each rule's alpha * rule_loss, then GANLoss.g_loss on the updated critic. FM is added only when fm_weight != 0.

The critic batch is the probe list (red, blue, stripe, dot, red stripe, empty). It is not a draw from ParticlePrior. The stamp still records n_particles=12 and cover 1.5 so a forked default is visible in score()["adv"]. The step does not train them.

Student learning rate is the 2-D fixture budget. Critic learning rate is LOCKED.lr. Neither replaces the stamp.

score()["verdict"] is PASS only for a locked claim whose modules and saved logits match that stamp, with finite phrase and adversary losses. Drift claims score FAIL with stranger_pairing, fm_on, or thinned_kappa. A PASS is a CPU toy. It is not a Music or Anima GPU transfer, and it is not the 40-step geometric verdict in dsl.md.

Not implemented

These are the next cuts. They are not accepted stanza keys today.

Idea	Status
player blocks with a different phrase per player	not implemented — one phrase, and it is the student's
A critic phrase (the critic has no concept rules)	not implemented
schedule / lazy critic steps other than reg_lazy from the stamp	not implemented — reg_lazy stays 1 because the stamp says so
cover as a stanza knob	not implemented — cover stays on the toys
Sampling n_particles inside the game	not implemented
Vanilla logistic as a pairing name	not implemented — stranger here is the flipped rp batch, matching locked_shared_floor
GEM / EA selected by the stanza	not implemented — those hooks stay on analysis_2d / ops_erase
^ and ; inside a CPU game	refused — pixel has no render here; reward has no scorer
A backend other than TwoAxisBackend	not implemented — DiT training is still train.py plus rule_loss, not Game
Multi-critic or more than two players	not implemented

When one of those lands, it has to call the toy stamp rather than grow a second LOCKED inside game.py, and the catalog page has to match the parser that actually shipped.