Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
57 changes: 57 additions & 0 deletions benchmarks/toy_audit/conditional_spread.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,57 @@
"""Independent target draws and grouped-bag scores for the wide conditional toy."""
import numpy as np


WIDTH = 65
BAGS = 2048
BAG_SIZE = 32
SCALES = np.array([[.5, 2.], [2., .5]])


def target_signals(seed):
shape = (2, BAGS, BAG_SIZE, 2)
return np.random.default_rng(seed).standard_normal(shape) * SCALES[:, None, None, :]


def bag_summaries(signals):
return np.stack((signals.mean(axis=2), signals.std(axis=2, ddof=1)), axis=-1)


def intervals(signals, truth):
low, high = np.quantile(bag_summaries(signals), [.1, .9], axis=1)
below = low[:, None] - truth
above = truth - high[:, None]
width = high - low
coverage = ((below <= 0) & (above <= 0)).mean(axis=1)
score = width + 10 * (np.maximum(below, 0) + np.maximum(above, 0)).mean(axis=1)
return coverage, width, score


def score_samples(samples):
"""Rows are grouped by the condition actually supplied to the generator."""
if samples.shape != (2, BAGS * BAG_SIZE, WIDTH) or not np.isfinite(samples).all():
raise ValueError("expected finite raw draws for both groups and all 65 coordinates")
signals = samples[:, :, :2].reshape(2, BAGS, BAG_SIZE, 2)
truth = bag_summaries(target_signals(102))
coverage, width, score = intervals(signals, truth)
_, _, reference = intervals(target_signals(100), truth)
means = signals.mean(axis=(1, 2))
sds = signals.std(axis=(1, 2), ddof=1)
metrics = dict(
max_scaled_mean_error=float(np.max(np.abs(means) / SCALES)),
max_relative_sd_error=float(np.max(np.abs(sds / SCALES - 1))),
max_coverage_error=float(np.abs(coverage - .8).max()),
combined_coverage=float(coverage.mean()),
variability_coverage=float(coverage[:, :, 1].mean()),
worst_interval_score_ratio=float(np.max(score / reference)),
max_group_padding_energy=float(np.square(samples[:, :, 2:], dtype=np.float64).sum(2).mean(1).max()),
)
for group in range(2):
for axis in range(2):
prefix = f"group{group}_axis{axis}"
metrics[f"{prefix}_mean"] = float(means[group, axis])
metrics[f"{prefix}_sd"] = float(sds[group, axis])
for index, name in enumerate(("mean", "sd")):
for key, values in (("coverage", coverage), ("width", width), ("interval_score", score)):
metrics[f"{prefix}_{name}_{key}"] = float(values[group, axis, index])
return metrics
38 changes: 26 additions & 12 deletions configs/forge/catalog.json
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
"benchmark/shared": 6,
"benchmark/smart_descent": 9,
"benchmark/toy100": 21,
"benchmark/toy_audit": 97,
"benchmark/toy_audit": 98,
"benchmark/transfer_suite": 88,
"config/100gaussians": 4,
"config/cifar_ddgan": 77,
Expand Down Expand Up @@ -40,7 +40,7 @@
"experiment/toy": 3,
"experiment/trajectory": 2,
"experiment/transition": 7,
"forge_engine": 128,
"forge_engine": 129,
"lrfree/feature-cells-cb64-ra": 4058,
"report/atlas-explanation": 7,
"report/behavioral_baseline": 32,
Expand Down Expand Up @@ -105,7 +105,7 @@
"candidate_patch": 26,
"configuration": 810,
"driver": 405,
"engine_support": 128,
"engine_support": 129,
"evaluator_or_audit": 324,
"evidence_artifact": 1551,
"frozen_candidate_implementation": 15,
Expand All @@ -114,18 +114,18 @@
"frozen_support": 28,
"narrative": 975,
"structured_evidence": 5070,
"support": 1561,
"support": 1562,
"support_artifact": 52,
"visualization": 23,
"visualization_artifact": 765
},
"files": 11855
"files": 11857
},
"coverage": {
"declared_count": 11855,
"declared_count": 11857,
"missing": [],
"stale": [],
"tracked_count": 11855,
"tracked_count": 11857,
"unclassified": [],
"valid": true
},
Expand Down Expand Up @@ -1161,6 +1161,13 @@
1
]
},
{
"family": "benchmark/toy_audit",
"git_blob": "0dfa5d3868799dc963713c9c151e4c241449d54f",
"path": "benchmarks/toy_audit/conditional_spread.py",
"role": "support",
"suggested_tiers": []
},
{
"family": "benchmark/toy_audit",
"git_blob": "7853300bd4720a57829582119272b6653d997853",
Expand Down Expand Up @@ -7408,14 +7415,14 @@
},
{
"family": "forge_engine",
"git_blob": "464feb98c141bc33e8a8cbaf0cd181acc2cc1593",
"git_blob": "be56c16c83c7c32a05cdd49d9b50ba4a5b4f3c6d",
"path": "experiments/forge/adapters.py",
"role": "engine_support",
"suggested_tiers": []
},
{
"family": "forge_engine",
"git_blob": "382f187ce1f89bee29ebab698394f1d501fe8c58",
"git_blob": "65f5a6d6230b152eeb72a6bdccad2780b80932c0",
"path": "experiments/forge/api.py",
"role": "engine_support",
"suggested_tiers": []
Expand Down Expand Up @@ -7623,6 +7630,13 @@
"role": "engine_support",
"suggested_tiers": []
},
{
"family": "forge_engine",
"git_blob": "ba093ed74fbd43651b5ce1e8cf6a292e84ffd67d",
"path": "experiments/forge/conditional_spread.py",
"role": "engine_support",
"suggested_tiers": []
},
{
"family": "forge_engine",
"git_blob": "4cfe348f385b343f1e242bfbf03c9013bb2e3cdb",
Expand Down Expand Up @@ -7716,7 +7730,7 @@
},
{
"family": "forge_engine",
"git_blob": "a1d03eb507f8a06c270d47f159e3ac1d5fc3a0b0",
"git_blob": "20767caf050e301d2d722cad0ab07b3892856daf",
"path": "experiments/forge/hostprofiles.py",
"role": "engine_support",
"suggested_tiers": []
Expand Down Expand Up @@ -7758,7 +7772,7 @@
},
{
"family": "forge_engine",
"git_blob": "86fb1d56536e523a884cabd0c7d215dbe69a55fd",
"git_blob": "1318c28e40b68080cc906dca2f0ab008cf8df467",
"path": "experiments/forge/leaderboard_contract.py",
"role": "engine_support",
"suggested_tiers": []
Expand Down Expand Up @@ -8038,7 +8052,7 @@
},
{
"family": "forge_engine",
"git_blob": "9c01465d6d5edef476c51c053b9dbe954f476e95",
"git_blob": "57ad623cf9109390788dd4bb39f8872a105be1e2",
"path": "experiments/forge/sampling.py",
"role": "engine_support",
"suggested_tiers": []
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,89 @@
{
"schema": "forge_fixed_test_list_v1",
"id": "technique-inventory-full-suite-2026-11-20",
"dated": "2026-11-20",
"leaderboard": "reports/forge/technique-inventory.md",
"derived_from": {
"rule": "Previous dated list plus the new Tier 2 stretch task in discriminator_stability revision 20.",
"previous_list_sha256": "e03d3154ff77a8bcdc684e7e88d446d702132cb14971035aa095e9d65bf0f2a0"
},
"policy": "The denominator is fixed: every listed test counts in every row as PASS, FAIL, not measured or another outcome, whether it is currently required, held or removed from a view. A test leaves the denominator only through a new dated list that keeps this one as history.",
"tests": [
"ae_gan_hold",
"alice_spiral_bigan",
"cgan_overlapping_classes",
"clockfree_audit",
"clockfree_audit_measurement_v1",
"cover_leftover",
"cramer_energy_overlap",
"cyclegan_unpaired_constellation",
"five_word_joint_hold",
"five_word_joint_smoke",
"gaussian1d_smoke",
"gaussian1d_stability",
"grid100",
"grid100_14k",
"grid100_mlp16",
"grid100_res16_scaled_tanh",
"grid100_res16_tanh",
"img_bars4",
"img_blobs4",
"img_intensity2",
"img_stripes2",
"infogan_anisotropic_modes",
"interp_img_patch_shift",
"interp_saddle_barycentric",
"interp_sparse_arc",
"mid_scale_identity",
"misgan_block_neural_imputer",
"misgan_block_recovery",
"misgan_mcar_p50_recovery",
"misgan_mcar_p80_anchored",
"mode_hold",
"model_ae_ddgan_bimodal_chain",
"model_ae_ddgan_gaussian_chain",
"model_ae_gan_multimodal",
"model_ae_reconstruction",
"model_ddgan_bimodal_chain",
"model_ddgan_gaussian_chain",
"model_vae_categorical",
"model_vae_hard",
"nash_overpowered_critic_ring16",
"nash_overpowered_generator_ring16",
"nash_unequal_width_hostile_imbalance",
"pacgan8_grid100",
"pacgan_unequal_mass",
"precision_gaussian_mxfp8",
"precision_gaussian_nvfp4",
"residual_student",
"ring16_acquisition",
"ring16_bigan_continuous",
"ring_extension",
"rotated100",
"rotated100_14k",
"staggered100",
"staggered100_14k",
"t3_iso_horizon_saddle_shift",
"t3_iso_scale_mass_8x",
"target_shift_recovery",
"trajectory",
"two_pole",
"unipolar",
"unused_token_hold",
"vector_anisotropic",
"vector_overlap",
"vector_spiral",
"vector_two_broad",
"vector_unequal_mass",
"vector_unequal_width",
"wide_conditional_spread",
"zero_tanh_rail_escape",
"zero_tanh_smoke"
],
"previous_list": "configs/forge/leaderboards/technique-inventory-full-suite-2026-11-18.json",
"change": "Append the wide conditional spread Tier 2 stretch test (wide_conditional_spread). Preserve all 69 earlier fixed-list tests and original evidence.",
"recorded_display_scope": {
"path": "configs/forge/leaderboards/technique-inventory-full-suite-2026-10-25.json",
"sha256": "14155fb6273af76ef248db2da6335c22ddeab2a1deb82b3eb3f0f6cd08b19c34"
}
}
109 changes: 109 additions & 0 deletions configs/forge/tasks/wide_conditional_spread.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,109 @@
{
"schema_version": 1,
"id": "wide_conditional_spread",
"description": "Learn the conditional spread of two equally likely groups on wide continuous inputs.",
"adapter": "wide_conditional_spread",
"execution": {
"initializer": "deterministic_orthogonal",
"host": "wide_conditional_spread",
"host_source": "benchmarks/toy_audit/conditional_spread.py",
"execution_path": "public_components",
"steps": 1200,
"resources": {
"num_particles": 128,
"z_dim": 4,
"batch_size": 64
},
"host_definition": {
"hidden": 64,
"width": 65,
"group_sds": [
[
0.5,
2.0
],
[
2.0,
0.5
]
],
"bag_size": 32,
"bags_per_group": 2048
},
"prior": {
"kind": "mog",
"sigma": 0.025,
"standardize": false
},
"prior_contract": "recipe_owned_v1",
"protocol": "screening",
"produces_state": false
},
"evaluation": {
"sampling_contract_version": 1,
"sampling_law": "public_prior_without_output_noise",
"eval_output_noise": "clean",
"kind": "transfer_sustained",
"evaluator": "benchmarks.transfer_suite.protocol:test_verdict",
"eval_samples": 131072,
"observations": 24,
"minimum_stable_checks": 5,
"scoring_weights": "live",
"thresholds": [
[
"max_scaled_mean_error",
"<=",
0.15
],
[
"max_relative_sd_error",
"<=",
0.2
],
[
"max_coverage_error",
"<=",
0.12
],
[
"worst_interval_score_ratio",
"<=",
1.5
],
[
"max_group_padding_energy",
"<=",
0.1
]
],
"guards": {
"finite_state": true,
"rng_isolation": true,
"mechanism_exercised": true,
"optimizer_roles": [
"generator",
"discriminator",
"prior"
],
"exact_optimizer_updates": true
},
"sources": {
"benchmarks/toy_audit/conditional_spread.py": "d8eb3ae671a14a02ac5cec1cf94c97cc32592c42ad60f3d3574e100fc010f7dd",
"experiments/forge/conditional_spread.py": "ded6dc57f9fc8b28c69c6d4895647ca683fd92aeb3eb0795938181e3e85e23c8"
}
},
"resources": {
"gpus": 1,
"gpu_memory_mb": 2048,
"cpu_threads": 1,
"timeout_seconds": 600,
"allow_cpu": true
},
"requires_capabilities": [
"checkpoint",
"named_rng",
"live_sampling",
"mog_prior"
],
"dependencies": []
}
Loading
Loading