Formulation toys (CPU scoreboard)

HyperGAN/conceptmod is the home for these gates. ParticleGAN is core primitives only. Install the PyPI package (pip install particlegan, particlegan>=0.5.0 in pyproject.toml and requirements.txt). Toys call that package for GANLoss / RpGAN, GradRegularizer / b_cap (GradientPenalty), and ParticlePrior. Those modules are not copied into conceptmod. The only local geometry is the leftover / cover / teacher field in conceptmod/toys/cover_leftover.py (and the per-toy fixtures around it). Toy code is not forked back into ParticleGAN.

The nine Wave-1 families were first opened on 255BITS/ParticleGAN by mistake. #26 (shared-trajectory) and #27 (orbit radius hold) were merged there and are being reverted. This tree keeps both families. #28–#34 were closed unmerged and are ported from those PR heads. Late-collapse selection is a later local gate (Lunar #23). Keep-critic is a new family in this tree. Image UNI lm_target, the residual student (composes with shared_trajectory), field lift, unused-token UNI hold, path-suffix LoRA honesty, and mid-scale identity hold (scale grid, not the lm_target teacher split) are later conceptmod families. The cover-posture fork is a later conceptmod gate on that set. Replace-macro expand honesty is a Wave-2 CPU gate in this tree: a~b must expand to the documented triple, and a geometric right for bare red~blue fails. Phrase-DSL job honesty is a later gate on the docs/dsl.md scoreboard. The DSL game-geometry bridge is a later gate in this tree: locked_shared loss_game wiring and geometric right from the phrase scoreboard, both required. The cross-toy suite leaderboard is a later gate: which candidate adv config PASSes every applicable stamp toy (CPU only; not Music/Anima/Supra transfer). Cover-posture fork stays its own module and is not a suite column that FAILs a config for the other claim. None of these was a ParticleGAN pull, and none is a fork of those toys. Backend-agnostic erase/keep is later and lives only in this repo.

family	module	ParticleGAN source
shared slow→fast vs stranger	conceptmod/toys/shared_trajectory.py	#26 (merged by mistake, being reverted; kept here)
slow→fast residual student	conceptmod/toys/residual_student.py	this tree; composes with shared_trajectory
orbit radius hold	conceptmod/toys/orbit_hold.py	#27 (merged by mistake, being reverted; kept here)
locked_shared RpGAN + b_cap floor	conceptmod/toys/locked_shared_floor.py	#28
unipolar residual	conceptmod/toys/unipolar.py	#29
AE-GAN + locked hold	conceptmod/toys/ae_gan_hold.py	#30
cover / leftover / faithful teacher	conceptmod/toys/cover_leftover.py	#31
backend-agnostic erase/keep	conceptmod/toys/erase_keep_backend.py	this repo (cpu/dummy path; not a ParticleGAN port)
leaderboard honesty	conceptmod/toys/leaderboard_honesty.py	#32
tiny particle posture	conceptmod/toys/particle_posture.py	#33
mode-hold ring	conceptmod/toys/mode_hold.py	#34
late-collapse selection	conceptmod/toys/late_collapse.py	local (Lunar #23); not a ParticleGAN pull
keep-critic locked_shared	conceptmod/toys/keep_critic.py	new. Music Arm B handoff keeps --adv_arch mlp; this row freezes the host critic and refuses that swap
image UNI lm_target	conceptmod/toys/lm_target.py	conceptmod (Anima image UNI; not a ParticleGAN pull)
2D→3D field lift	conceptmod/toys/field_lift.py	particle-sliders tests/test_field3d.py (reviewed; not a ParticleGAN pull)
unused-token / UNI hold	conceptmod/toys/unused_token_hold.py	conceptmod (image-slider posture; not a ParticleGAN pull)
path-suffix LoRA	conceptmod/toys/path_suffix_lora.py	native CPU gate (Anima attach / PEFT suffix match)
mid-scale identity hold	conceptmod/toys/mid_scale_identity.py	conceptmod (Anima smile mid-scale; eval grid always includes −1)
cover posture fork	conceptmod/toys/cover_posture_fork.py	this tree (composes the locked floor with the two cover pins; not a ParticleGAN port)
replace-macro expand honesty	conceptmod/toys/dsl_macro_expand.py	this tree (phrase DSL ~; not a ParticleGAN port)
phrase DSL jobs	conceptmod/toys/dsl_phrase_jobs.py	this tree (docs/dsl.md scoreboard; not a ParticleGAN port)
DSL game geometry	conceptmod/toys/dsl_game_geometry.py	this tree (locked loss_game step and geometric right; not a ParticleGAN port)
cross-toy suite leaderboard	conceptmod/toys/suite_leaderboard.py	this tree (stamp + DSL matrix; cover posture is its own toy, not leaderboard_honesty)

Allowed particlegan imports are the primitives: GANLoss, GradientPenalty / GradRegularizer, ParticlePrior, ParticleRegularizer, and get_recipe (the AE-GAN study preset). This tree does not copy GradRegularizer. It also does not call particlegan.orbit_hold, particlegan.leaderboard_honesty, or any other toy that briefly lived in that repo. The ring critic is conceptmod.toys.mlp because the published package does not export the 100-Gaussians MLP; that class is host critic shape, not a second penalty.

Culture reviewed: HyperGAN/particle-sliders analysis/slider2d/locked_baseline_defaults.py (locked_shared / #94), analysis/slider2d/field3d.py and tests/test_field3d.py (2D recipe on an R³ leftover field), Music Arm B (docs/music-arm-b-gates.md, trainer ARM_B: adv_arch=mlp, cover/pole 1.0). Demo cover 1.5 and that Music pin are different postures; "cover_weight=1" is a substring of "cover_weight=1.5" and is not a match. This repo's conceptmod/analysis_2d.py is DSL geometry, a different fixture. Keep-critic does not adopt the Music mlp head.

A PASS here is a CPU toy. It is not a Music or Anima GPU transfer.

Live conceptmod consumes this package

These modules are the formulation source of truth. Live scoring does not keep a second copy of the adv recipe.

live path	what it calls
conceptmod.analysis_2d / conceptmod.analysis_dsl	locked_adv_defaults() — the locked_shared_floor.LOCKED stamp (RpGAN logistic, b_cap coeff=1 κ=1, FM off, demo cover 1.5, n=12). Keep/leak flags use U_KEPT_MIN and SAME_DIR_MAX.
conceptmod.game.loss_game	The same locked_adv_defaults() stamp. FM-on and thinned κ go through reject_unlocked when the game claims locked. Stranger pairing (flipped batch, GANLoss mode stays rp) is refused on that claim. The game does not sample n_particles or apply cover.
conceptmod.toys.dsl_game_geometry	Bridge gate. A row PASSes when that locked game step is clean and analysis_2d.run_method / analysis_dsl.run_job scores geometric right on the same phrase. A game PASS alone is not the gate.
conceptmod.ops_erase.erase_keep_geometry	hold_dir, faithful_guard_e, and leftover_bipolar from the cover / leftover toy.
conceptmod.toys.erase_keep_backend	That same helper, with axes read back from CpuBackend / dummy predict_v (and a supra-shaped latent stub).
formulation PASS	claim_pass. An empty negative list raises HonestyError. Geometric verdicts stay right / needs help / recipe and are not a PASS.
replace-macro expand	dsl_macro_expand.claim_expand_pass. red~blue PASS is the documented triple (exaggerate / write / orthogonal at 0.2 / 0.4 / −0.1). Geometric right on that phrase is a FAIL (recipe). An empty negative list raises HonestyError.
conceptmod.toys.dsl_phrase_jobs	The same locked_adv_defaults() stamp and U_KEPT_MIN / SAME_DIR_MAX keep gates, via the 2-D job runners. claim_phrase_pass is phrase honesty (empty negatives raise). Geometric labels stay right.

The 2-D Adam budget (40 steps, lr 8e-2) is fixture budget. It is not a new adv recipe, and it does not replace locked_shared as the winning default. tests/test_live_toys.py imports the analysis and erase paths and checks that they execute these helpers.

Locked shape

Demo posture, from the slider lock, unless a row below names a drift:

knob	value
pairing	RpGAN logistic (GANLoss, loss_type=logistic, mode=rp)
penalty	GradRegularizer b_cap, coeff=1, kappa=1, norm=l2
FM	off (fm_weight=0)
cover	demo 1.5 (Music pole/cover 1.0 is a different posture)
cloud	n=12, particle_l2=0.02 (not Hub 128, not Recipe("gan") 20_000)
VICReg	0.05 when the toy has a latent prior
critic	the toy's host critic; an architecture swap is refused

Named omissions (reported, not silent aliases):

• Unipolar trains with cover weight 0. Cover is the eval gate (>= 0.85). No particle cloud. Budget 400 steps.
• Orbit does not apply VICReg. Those particles are the orbit state, not a latent prior. The critic is a frozen linear probe.
• Mode-hold records cover 1.5 and does not add a supervised cover loss. Learning rate is this toy's budget (2e-3), not the slider 5e-3.
• Music --parts 0 on the posture toy is the empty-cloud spelling: cover 1.0, VICReg off. It is a second PASS, not a substitute for the n=12 demo lock.
• Late-collapse does not train. The curve is eight synthetic checkpoints. Train loss falls through the collapse; the val gate is the only export rule. The locked_shared stamp is read and not retuned.
• Image UNI lm_target trains target MSE only. Adversarial loss does not apply (Anima image UNI is trajectory MSE). locked_shared is the only legal adv posture and is refused if drifted, not trained. Step budget is 80, lr 0.5.
• Unused-token hold trains an embed-slot student (image-slider UNI). Hold weight is 1.0 (the image-slider default). Demo cover 1.5 and n=12 / particle_l2=0.02 are recorded and are not a second loss. No particle cloud. Not an Anima or Supra GPU result.
• Mid-scale identity has no particle cloud (residual student, same omission as unipolar). Eval grid is -1, 0, 0.5, 1. Step budget is 800. Cover stays demo 1.5. It is not the lm_target teacher split.
• Cover posture fork keeps the locked_shared shape (n=12 included) and splits only the cover pin. Demo 1.5 and Music 1.0 each PASS when the claim, the label, and (for Music) the drift report agree. A swapped pin with no report FAILs. The locked-shared floor still claims demo only, so Music 1.0 fails that stamp.
• DSL game geometry does not train a new adv recipe. Geometry uses the 2-D fixture budget (40 steps, lr 8e-2, seed 0). The game row is one locked_shared step. A game PASS with no geometric right fails. red~blue stays recipe.

Scoreboard

Seed 0. Numbers are the reference runs these gates were written against. Pytest re-checks the PASS/FAIL split (and, where the original test pinned a digit, that digit). Lower error is better unless a column says otherwise.

Shared trajectory (identity MSE ≤ 0.02)

arm	pairing	identity MSE	gate
locked_shared	shared	0.00167	PASS
nearest_stranger	nearest other slow arc	0.297	FAIL
stranger	opposite seed	0.567	FAIL

400 steps. Locked entry point refuses stranger pairing.

Residual student (identity MSE ≤ 0.02 and success rate = 1)

Same arcs and locked_shared pins as the shared-trajectory toy. The head is a residual on the slow arc, fast = slow + head(slow, z). Supervised residual loss runs only on same-seed both-land rows: slow touchdown impact ≤ 0.10, and the paired fast arc ends within 0.25 of this seed's pad (the same-seed fast endpoint). Pad gaps start near 0.383, so a touchdown on another seed's pad falls outside this landing. A pass also requires wrong_pad_rate = 0.

arm	pairing	both-land	identity MSE	success	wrong-pad	gate
locked_shared	shared	12	0.00197	1.00	0	PASS
nearest_stranger	nearest other slow arc	0	0.143	0.083	0.917	FAIL
stranger	opposite seed	0	0.511	0	1	FAIL

400 steps, seed 0. Locked entry point refuses stranger pairing. Copying the paired fast arc lands on the wrong pad (success 0, wrong-pad 1) for both drift indexes; that floor is the Lunar crash catch. One nearest-stranger seed can graze its own pad after training and the arm still fails. A pass here is a CPU toy score.

Orbit radius hold

arm	radius_rel	direction	speed_rel	b_cap	gate
locked_shared	~0	0.992	8.0e-3	4	PASS
ablate_residual	0.410	1.000	~0	4	FAIL radius_drift
stranger_vanilla / shuffle	~0	0.992	8.0e-3	4	FAIL stranger_pairing
fm_on	~0	0.992	8.0e-3	4	FAIL fm_on
hinge_drift	~0	0.992	8.0e-3	4	FAIL loss_not_logistic
thin_bcap	~0	0.992	8.0e-3	9	FAIL thinned_bcap

Heading and speed stay healthy on the spiral. Radius is the gate.

Locked-shared floor (absolute error vs the real modules)

arm	gate	adv err	g err	cap err	κ probe
locked_shared	PASS	0	0	0	0
thinned_kappa	FAIL	0	0	0	0.040
fm_on	FAIL	0	7.28e-4	0	0
vanilla_logistic	FAIL	0.693	0.0231	0	0
stranger_pair	FAIL	2.44e-4	1.19e-6	0	0
music_cover_1	FAIL	0	0.233	0	0
hub128	FAIL	0	0	0	0

The κ=0.2 probe is what catches a cap that hardcodes center 1. At locked κ=1 that stub still matches the step penalty.

Unipolar residual (cover ≥ 0.85, leak ≤ 0.05, neu_hold ≥ 0.85)

arm	cover	leak	neu_hold	gate
locked_rpgan	0.964	0.0005	0.956	PASS
mse_only	1.000	0	0.667	FAIL neu_hold
polarity_flipped	0	0	0.963	FAIL plus cover

FM-on, vanilla pairing, thinned κ, and putting demo cover 1.5 back on as a train loss are refused.

AE-GAN hold (recon ≤ 0.05, hold ≤ 0.35, b_cap and RpGAN every step)

arm	recon	hold	b_cap	RpGAN	gate
locked	0.0038	0.024	250/250	250/250	PASS
ae_only	0.0024	0.005	0/250	0/250	FAIL
stranger / FM-on / thin κ	—	—	refused	refused	FAIL

AE-only can still reconstruct. It fails because the locked adversary never ran. get_recipe("ae_gan") stays the study preset (K=400, lazy every 4); this gate overrides to n=12 and lazy=1.

Cover / leftover / faithful teacher

arm	cover	teacher	u_kept	leak	gate
locked	1.5	faithful_guard_e	0.936	0.000	PASS
cover_zero	0	faithful_guard_e	0.577	0.060	FAIL undershoot
teacher_drift	1.5	faithful	0.935	0.468	FAIL teacher_leak

800 steps (budget). Stranger pairing, FM-on, thinned b_cap, and n=128 are refused.

Backend-agnostic erase/keep (host predict_v, image latent)

Axes round-trip through erase_keep_geometry. No GAN steps. cover_zero is an unpinned residual at half the pole, not the 800-step cover digit. The supra row is a weight-free (4, 32, 32) stub with flow time in [0, 1]. cpu and dummy are the same CpuBackend latent (4, 8, 8).

arm	cover	leak	hold	same_dir	u_kept	pole_err	gate
locked	1.5	0	1	0	1.000	0	PASS
teacher_leak	1.5	1	—	0	1.000	0	FAIL teacher_leak
cover_zero	0	0	1	0	0.500	0.500	FAIL undershoot
wrong_poles	1.5	0	1	1	1.000	2	FAIL undershoot,wrong_poles

The split is the same on cpu, dummy, and supra_stub. A (C, 1, H, W) latent is refused.

Leaderboard honesty (mean_abs ≥ 0.30 and grad_med ≤ 1)

order	arm	mean_abs	grad_med	cover_score	gate
1	locked_shared	0.514	0.420	0.772	won
2	stranger_pairing	0	0.073	0	fail
3	thinned_b_cap (κ hardcoded at 100)	0.694	2.342	1.041	fail

PASS requires the locked cell to win and every declared bad arm to fail. An empty negative list raises HonestyError. Cover alone would crown the thinned hinge.

Tiny particle posture

arm	n	particle_l2	cloud_ms	anchor_drop	residual_l2	gate
locked_tiny	12	0.02	0.00246	1.66e-5	0.0037	PASS
music_parts0	0	0	0	0	0.0116	PASS
hub128_routed	128	0	0.999	0	0.0455	FAIL
zero_particle_l2	12	0	0.00211	0	0.0023	FAIL

600 steps. zero_particle_l2 can keep a small cloud; it fails because the 0.02 anchor does not contract a unit cloud.

Mode-hold ring (8 Gaussians, modes ≥ 7 and HQ ≥ 0.90)

arm	drift	modes	HQ	effective	gate
locked	none	8/8	1.000	7.48	PASS
b_cap off	f_none, coeff 0	1/8	0.090	1.00	FAIL

1200 steps, EMA, 4096 samples. Stranger pairing, FM-on, and a κ-hardcoded stub are refused unless the drift is named. The stub is not trained.

Late-collapse selection (export the earlier good checkpoint)

Eight checkpoints. No Adam budget. Train loss falls on every step, so the best train loss is the last step. Val error is best at step 3 (0.06), still inside the gate at step 4 (0.11), then collapsed. The val gate refuses val_error > 0.25 and any collapsed flag, then keeps the minimum val error (earlier step on a tie).

arm	rule	step	train_loss	val_error	collapsed	gate
locked_val_gate	val gate	3	0.32	0.06	no	PASS
last_step	final index	7	0.01	1.40	yes	FAIL
best_train_loss	argmin train, no val	7	0.01	1.40	yes	FAIL

PASS requires the locked rule to export step 3 and every declared bad arm to export a collapsed checkpoint. An empty negative list raises HonestyError. A won flag that disagrees with the export is a dishonest board and raises HonestyError. claim_pass stays the two-pole gate.

Keep-critic (frozen host, no MLP swap)

The host critic is a frozen linear scorer. It is not conceptmod.toys.mlp.SimpleMLPDiscriminator, and it is not stepped. Music Arm B's --adv_arch mlp is refused before any step. Cover posture is on every row: demo 1.5 is the lock, Music pole/cover 1.0 fails this stamp. A PASS does not transfer to a Music or Anima GPU run.

arm	gate	adv err	g err	cap err	κ probe	critic
locked_shared	PASS	0	0	0	0	frozen host, Δw=0
stranger_pairing	FAIL	0.167	2.22e-4	0	0	frozen host
fm_on	FAIL	0	0.200	0	0	frozen host
thinned_kappa	FAIL	0	0	0	0.040	frozen host
music_cover_1	FAIL	0	0.233	0	0	frozen host, cover 1.0
critic_step	FAIL	0	0	0	0	same host, Δw=0.0283
forced_mlp_swap	FAIL	—	—	—	—	refused (mlp not trained)

8 steps (budget). The κ=0.2 probe is what catches a cap that hardcodes center 1.

Image UNI lm_target (expr_gain ≥ 0.85, struct_hold ≥ 0.95)

Locked target: trajectory. Particle-sliders Anima image UNI (--lm_target trajectory in anima_slider.py). Live smile does not lock direct or cfg_delta: a 1-step velocity gap cannot carry expression (cos(v(plus), v(neu)) ≈ 0.99993, MSE ≈ 0.00037 on the v4 diagnostic). This CPU field shuts that high-σ gate, so the 1-step gradient is zero. Expression is a late-σ write (σ ≤ 0.5) and shows up on the K=4 FlowMatch Euler trajectory (traj_expr_gap = -1).

arm	expr_gain	struct_hold	one_step_mse	traj_expr_gap	gate
trajectory	1.000	1.000	0	-1	PASS
direct	0	1.000	0	-1	FAIL expr_gain
cfg_delta	0	1.000	0	-1	FAIL expr_gain

80 steps, lr 0.5 (budget). Same field and the same three gates on every arm. Structure hold stays 1 on the failing arms; the concept axis is the split. Adversarial loss does not apply. Stranger pairing, FM-on, and a thinned b_cap are refused. Not an Anima GPU transfer.

2D→3D field lift (plane gates survive the tangent pushforward)

Reviewed against particle-sliders Field3D (analysis/slider2d/field3d.py, tests/test_field3d.py): the locked 2D recipe transfers with no 3D-only hack. κ stays 1. The sheet is the cover toy's guarded odd residual (slider 1, content 0.55, unused ê stripped by faithful_guard_e). On the coordinate plane that vector is the paste (1, 0.55, 0). The lift tilts û toward ê by π/4; the correction is slider·û(θ) + content·ĉ.

arm	frame	u_kept	content	leak	pole_err	gate
locked_2d	plane	1	1	0	0	PASS
locked_lift	tilted	1	1	0	0	PASS
naive_copy	tilted, no z correction	0.707	1	1	0.504	FAIL naive_lift, undershoot
stranger_pairing	tilted, flipped pair	1	1	0	0	FAIL stranger_pairing (adv err 4.92e-4)
fm_on	tilted, fm_weight=0.1	1	1	0	0	FAIL fm_on (fm term 6.36e-3)
thinned_kappa	tilted, center hardcoded at 1	1	1	0	0	FAIL thinned_kappa (κ probe 0.040)

One scored step. The paste's cover gap is g err 0.439. At locked κ=1 the thinned center still matches the step penalty; the κ=0.2 probe is what fails it. Stranger pairing, FM-on, and the stub keep the lifted geometry.

Unused-token / UNI hold (hold ≥ 0.85 and concept move ≥ 0.85)

Image-slider posture: an unused embed slot stays on encode(neu) while the concept slot moves. The student writes one shared residual into every slot. Concept RpGAN sees only the concept slot, so without a hold loss half of that step lands on the unused pin.

arm	unused hold	concept move	gate
locked_shared	0.991	0.968	PASS
no_hold	0.479	0.959	FAIL no_hold, unused_trashed
stranger_pairing	0.024	0.969	FAIL stranger_pairing, unused_trashed
stranger_vanilla	0.998	0.944	FAIL stranger_pairing
fm_on	0.996	0.981	FAIL fm_on
thin_bcap	0.992	0.977	FAIL thinned_bcap

200 steps, seed 0. stranger_pairing aligns the unused slot with the concept embed (the wrong neu partner) and trashes the pin; concept move stays. stranger_vanilla, FM-on, and the center-0 b_cap stub can still hold the pin. They fail because the adv card drifted. Faithful b_cap probe is 4; the stub is 9. Hold weight 1.0 is the image-slider pin, not Field3D cover. A PASS here is not an Anima, Supra, or Music GPU transfer.

Path-suffix LoRA (coverage = 1 and bare |grad| < 1)

One backward on a block whose Linear leaves are self_attn.proj, cross_attn.proj, and a bare proj. PEFT list targets match the exact key or a .<suffix> ending, so ["proj"] trains the head as well as both attention projs. The locked child regex is (?:self_attn|cross_attn)\..+. The literal self_attn.* fullmatches the parent module and is refused. q_proj matches nothing on this block (Anima's encoder leaf name is not a bare proj). Mean |lora_B grad| is the coverage signal. A colliding head zeroes coverage even when the attention leaves also move.

arm	targets	coverage	self |g|	cross |g|	bare |g|	gate
locked_suffix	self_attn.proj, cross_attn.proj	1	2.25	2.25	0	PASS
locked_regex	child regex above	1	2.25	2.25	0	PASS
bare_proj	["proj"]	0	2.25	2.25	4.50	FAIL bare_proj
head_only	"proj" fullmatch	0	0	0	4.50	FAIL bare_proj coverage
partial_suffix	self_attn.proj only	0.50	2.25	0	0	FAIL coverage
overbroad_star	self_attn.*	—	—	—	—	refuse empty_trainable
empty_q_proj	q_proj	—	—	—	—	refuse empty_trainable

No Hub weights. A PASS here is not an Anima GPU transfer.

Mid-scale identity hold (concept cos ≥ 0.85 at ±1, identity ≥ 0.85 at 0 and 0.5)

Poles may move the concept. Identity is the content retain axis (hold_dir); it has to hold at scale 0 and at mid-scale 0.5. Teacher poles are faithful_guard_e (leftover ê off the odd part). Pole motions use leftover_bipolar. Particle-sliders formulation sampling always includes -1 (unipolar canary on every cell, bipolar gated pole). This eval grid is -1, 0, 0.5, 1. Anima smile's sample grid 0, 0.25, 0.5, 1 drops -1 and is a FAIL even when the residual is fine. This is the scale-grid gate, not the image UNI lm_target teacher split.

arm	concept+	concept−	identity@0	identity@0.5	eval grid	gate
locked	1.000	1.000	0.987	0.998	−1, 0, 0.5, 1	PASS
mid_collapse	1.000	1.000	0.994	0.000	−1, 0, 0.5, 1	FAIL identity_mid
missing_minus	1.000	n/a	0.987	0.998	0, 0.25, 0.5, 1	FAIL missing_minus
polarity_flipped	−1.000	−1.000	0.992	0.995	−1, 0, 0.5, 1	FAIL polarity
stranger	1.000	1.000	0.007	0.000	−1, 0, 0.5, 1	FAIL stranger_pairing

800 steps. Seed 0. mid_collapse is the smile failure: poles and scale 0 stay the person, scale 0.5 is a stranger. missing_minus rescores the locked residual on the Anima grid. Vanilla pairing, FM-on, and a thinned b_cap are refused. A CPU PASS is not a Music or Anima GPU transfer.

Cover posture fork (demo 1.5 and Music 1.0 are named postures)

Same locked_shared shape on every row: RpGAN logistic, b_cap coeff=1 κ=1 l2, FM off, n=12, particle_l2=0.02. The cover pin is the posture. 8-step floor budget (identity, not a train). A PASS is a CPU toy. It is not a Music or Anima GPU transfer. Wave-1 cover / leftover stays the demo-1.5 geometry gate; this row only checks the fork.

arm	claim	cover	label	drift reported	g err	alias gap	gate
demo_1_5	demo_1_5	1.5	demo	—	0	0.233	PASS
music_1_0	music_1_0	1.0	music	cover_weight, cover_posture	0	0.233	PASS
mislabeled_swap	demo_1_5	1.0	demo	none	0.233	0.233	FAIL unreported swap
stranger	demo_1_5	1.5	demo	—	1.19e-6	0.233	FAIL stranger_pairing
fm_on	demo_1_5	1.5	demo	—	7.28e-4	0.233	FAIL fm_on
thinned_kappa	demo_1_5	1.5	demo	—	0	0.233	FAIL thinned_bcap

The 0.233 alias gap is the generator-term split between cover 1.5 and cover 1.0 on this seed (the same split the locked-shared floor reports for music_cover_1). Music 1.0 still FAILs that floor, which claims the demo posture only. The Music label does not waive stranger pairing, FM-on, or a thinned kappa cap. This fork is its own toy (python -m conceptmod.toys.cover_posture_fork). The suite leaderboard does not add opposite-claim columns for it.

Replace-macro expand honesty (a~b is a macro)

red~blue is not a fourth loss. Default λ = 0.1 is the written triple from dsl.md: blue++:0.2 | red=blue:0.4 | blue%red:-0.1 (exaggerate blue, write red←blue, orthogonal blue%red with negative alpha). PASS is that expand. On this fixture the macro is recipe: blue++ and red=blue oppose, so claiming geometric right FAILs even when the three rules match. The optional loss_game("red~blue") row is locked_shared wiring (the three expanded rules stepped). It does not claim geometric right. No Adam budget on the expand arms. A PASS is a CPU toy. It is not a Music or Anima GPU transfer.

arm	expand	geometric claim	gate
locked_expand	exaggerate blue 0.2, write red←blue 0.4, orthogonal blue%red −0.1	recipe	PASS
wrong_order	write, then exaggerate, then orthogonal	recipe	FAIL op_order
swapped_roles	exaggerate red, write blue←red, orthogonal red%blue	recipe	FAIL swapped_roles
wrong_alphas	documented roles, alphas 1	recipe	FAIL alphas
dropped_percent	no % term	recipe	FAIL dropped_percent
single_write	bare red=blue	recipe	FAIL single_write
geometric_right	documented triple	right	FAIL geometric_right
fourth_loss	a live replace op	recipe	FAIL fourth_loss
game_wiring	same triple, one locked_shared step	recipe	PASS (wiring only)

PASS requires the parser to match the triple and every declared bad arm to fail. An empty negative list raises HonestyError. claim_pass stays the two-pole gate.

Phrase DSL jobs (documented phrase, not a new operator)

Geometric right from docs/dsl.md is the job verdict, scored by analysis_dsl / analysis_2d on the CPU 2-D field (40 steps, lr 8e-2, seed 0 — fixture budget, not a new adv recipe). Formulation PASS is honesty that the phrase is that job. claim_phrase_pass raises HonestyError on an empty negative list. A PASS is a CPU toy. It is not a Music, Anima, or Supra GPU transfer.

arm	phrase	geometric	color on red	stripe hold	probe	gate
neutralize	red--	right	+0.015	0.999	origin; write cos −1	PASS
bipolar	red++	right	+3.167	0.998	blue −3.167	PASS
remap	red=blue	right	−1.167	0.998	write cos +1	PASS
keep_erase	`red--	stripe#stripe`	right	+0.015	1.000	erase + stripe#stripe
mix	red=red stripe	right	+1.000	1.000	pattern on red +0.985	PASS
isolate	red stripe=stripe	right	—	0.907	mix (+0.037, +1.000)	PASS

arm	phrase	geometric	reason	gate
plus_synonym	red+stripe	—	honesty	FAIL
bang_synonym	red!!	—	honesty	FAIL
slash_synonym	red/	—	honesty	FAIL
percent_isolate	red%red stripe	right	percent_isolate	FAIL
keep_erase_bare	red-- claimed as keep+erase	right	retain	FAIL
pixel_caret	red^blue	—	dead_op (no render)	REFUSE
reward_semi	;red	—	dead_op (ImageReward)	REFUSE

red++ is the bipolar slider: stripe holds and blue scales to −3.167 with red. Mix and isolate are the write recipes (red=red stripe, red stripe=stripe). A first-class +, !!, or / is an honesty FAIL; the parser still rejects those glyphs. % as isolate can satisfy the isolate landing check (mix x +0.192) while inflating the perpendicular (mix y +6.262). That row stays FAIL. Bare red-- is neutralize (stripe hold 0.999). The same phrase claimed as keep+erase FAILs retain. ^ and ; are refused on this fixture.

DSL game geometry (locked step and geometric right)

A row PASSes only when both are true: loss_game steps under locked_shared with the locked stamp, and the 2-D fixture train scores geometric right for that phrase. Scoring is analysis_2d.run_method / analysis_dsl.run_job (the same verdicts as tests/test_2d_analysis.py and tests/test_dsl_jobs.py). This module does not copy those thresholds.

The geometry budget is the 2-D fixture: 40 Adam steps, lr 8e-2, seed 0. That budget is not a new adv recipe. The game row is one CPU step on locked_adv_defaults(). A game PASS alone is not geometric right. A bridge PASS is a CPU toy. It is not a Music or Anima GPU transfer.

arm	phrase	game	geometry	gate
write	red=blue	PASS locked	right (write)	PASS
erase_bare	red--	PASS locked	right (erase_esd)	PASS
erase_freeze	red--|stripe#stripe	PASS locked	right (erase_esd_freeze)	PASS
exaggerate	red++	PASS locked	right (exaggerate)	PASS
wiring_only	red=blue	PASS locked	skipped	FAIL wiring_only
faked_geometry	red=blue	PASS locked	faked right	FAIL faked_geometry
geometry_only_stranger	red=blue	FAIL stranger, claim drift	right	FAIL geometry_only unlocked_claim
geometry_only_fm	red=blue	FAIL FM-on, claim drift	right	FAIL geometry_only unlocked_claim
locked_stranger_refuse	red=blue	locked claim refused	right	FAIL locked_claim_refused
locked_fm_refuse	red=blue	locked claim refused	right	FAIL locked_claim_refused
dead_pixel	^	refused	dead	FAIL dead_phrase claimed_pass
dead_reward	;	refused	dead	FAIL dead_phrase claimed_pass
replace_macro	red~blue	PASS locked	recipe	FAIL recipe_not_independently_right

red~blue is the replace macro. docs/dsl.md marks it recipe, not independently right. A locked game can still step it. That game PASS does not become a geometry PASS. The expand-honesty toy above is the gate for the triple itself. Phrase-job honesty is the gate that the phrase is the documented job.

Suite leaderboard (which config passes all applicable stamp toys?)

Cross-toy matrix in conceptmod/toys/suite_leaderboard.py. Distinct from leaderboard_honesty (single-toy two-pole gate). Candidate rows include locked_shared (demo LOCKED), music_cover_1_0, stranger_pair, fm_on, thinned_kappa, vanilla_logistic, and hub128. Columns are stamp and dsl only.

A stamp sweep winner must PASS every applicable stamp cell (N/A excluded). Phrase-DSL columns are scored under locked_shared only (or N/A for cover drifts). Claiming a sweep with no declared failing configs raises HonestyError.

Cover posture is not a suite column. Demo cover 1.5 and Music cover 1.0 are named claims on conceptmod/toys/cover_posture_fork.py (python -m conceptmod.toys.cover_posture_fork). That toy still PASSes a row only on the posture it claims, and still FAILs a silent swap or the wrong pin. The suite does not score locked_shared under the Music claim, or music_cover_1_0 under the demo claim, and then report those mismatches as formulation FAILs.

PASS all stamp toys is a CPU formulation result. It is not a Music, Anima, or Supra GPU transfer. See suite-leaderboard.md.

Run

pytest tests/test_live_toys.py tests/test_2d_analysis.py tests/test_erase_cpu.py tests/test_dsl_jobs.py -q

pytest tests/test_toy_shared_trajectory.py \
       tests/test_toy_residual_student.py \
       tests/test_toy_orbit_hold.py \
       tests/test_toy_locked_shared_floor.py \
       tests/test_toy_unipolar.py \
       tests/test_toy_ae_gan_hold.py \
       tests/test_toy_cover_leftover.py \
       tests/test_toy_erase_keep_backend.py \
       tests/test_toy_leaderboard_honesty.py \
       tests/test_toy_particle_posture.py \
       tests/test_toy_mode_hold.py \
       tests/test_toy_late_collapse.py \
       tests/test_toy_keep_critic.py \
       tests/test_toy_lm_target.py \
       tests/test_toy_field_lift.py \
       tests/test_toy_unused_token_hold.py \
       tests/test_toy_path_suffix_lora.py \
       tests/test_toy_mid_scale_identity.py \
       tests/test_toy_cover_posture_fork.py \
       tests/test_toy_dsl_macro_expand.py \
       tests/test_toy_dsl_phrase_jobs.py \
       tests/test_toy_dsl_game_geometry.py \
       tests/test_toy_suite_leaderboard.py -q

One family at a time, with a tailable log:

python -m conceptmod.toys.shared_trajectory   # via the module's train(); see tests
python -m conceptmod.toys.residual_student --family
python -m conceptmod.toys.orbit_hold
python -m conceptmod.toys.locked_shared_floor
python -m conceptmod.toys.unipolar
python -m conceptmod.toys.ae_gan_hold
python -m conceptmod.toys.cover_leftover
python -m conceptmod.toys.erase_keep_backend
python -m conceptmod.toys.leaderboard_honesty
python -m conceptmod.toys.particle_posture
python -m conceptmod.toys.mode_hold
python -m conceptmod.toys.late_collapse
python -m conceptmod.toys.keep_critic
python -m conceptmod.toys.lm_target
python -m conceptmod.toys.field_lift
python -m conceptmod.toys.unused_token_hold
python -m conceptmod.toys.path_suffix_lora
python -m conceptmod.toys.mid_scale_identity
python -m conceptmod.toys.cover_posture_fork
python -m conceptmod.toys.dsl_macro_expand
python -m conceptmod.toys.dsl_phrase_jobs
python -m conceptmod.toys.dsl_game_geometry
python -m conceptmod.toys.suite_leaderboard

shared_trajectory and residual_student are libraries (train_locked / train_drift). python -m conceptmod.toys.residual_student --family prints the residual board. The others print a one-line board when run as __main__.