task_id string | sample int64 | ts timestamp[s] | call dict | code string | verdict dict |
|---|---|---|---|---|---|
moe_swiglu | 0 | 2026-10-31T09:58:57 | {
"request_id": "5cc6185a-7d40-4d3b-9e84-418c91b8e512",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 12.54,
"finish_reason": "stop",
"reasoning_chars": 5444,
"prompt_tokens": 284,
"completion_tokens": 1931,
"reasoning_tokens": 1778
} | def kernel(x, wg, wu, wd):
x = np.asarray(x)
wg = np.asarray(wg)
wu = np.asarray(wu)
wd = np.asarray(wd)
H, I = wg.shape
W = np.concatenate((wg, wu), axis=1)
T = x @ W
A = T[:, :I]
B = T[:, I:]
z = np.exp(-np.abs(A))
sig = np.where(A >= 0, 1.0 / (1.0 + z), z / (1.0 + z))
... | {
"task_id": "moe_swiglu",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.293,
"speedup": 2.2533559665,
"max_abs_err": 0.0000016081,
"max_rel_err": 0.0078951783,
"reasons": [],
"detail": {
"task_id": "moe_swiglu",
"cases": [
{
... |
moe_swiglu | 1 | 2026-10-31T09:59:11 | {
"request_id": "e9cc8b1c-9163-4f78-a069-fa215d0cadc7",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 18.25,
"finish_reason": "stop",
"reasoning_chars": 9304,
"prompt_tokens": 284,
"completion_tokens": 2752,
"reasoning_tokens": 2662
} | def kernel(x, wg, wu, wd):
i = wg.shape[1]
t = x @ np.concatenate((wg, wu), axis=1)
g = t[:, :i]
u = t[:, i:]
g *= 1.0 / (1.0 + np.exp(-g))
g *= u
return g @ wd | {
"task_id": "moe_swiglu",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.3689,
"speedup": 2.7811005754,
"max_abs_err": 0.0000016081,
"max_rel_err": 0.0078951783,
"reasons": [],
"detail": {
"task_id": "moe_swiglu",
"cases": [
{
... |
group_router | 0 | 2026-10-31T09:59:30 | {
"request_id": "50a3bc24-20a0-4e76-865d-4e65747c734f",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 161.5,
"finish_reason": "stop",
"reasoning_chars": 7100,
"prompt_tokens": 333,
"completion_tokens": 2383,
"reasoning_tokens": 1979
} | def kernel(x, w):
x = np.asarray(x)
w = np.asarray(w)
N = x.shape[0]
E = w.shape[0]
groups = 8
gsize = E // groups
logits = x @ w.T
# Stable sigmoid
scores = np.empty(logits.shape, dtype=np.float64)
pos = logits >= 0
neg = ~pos
scores[pos] = 1.0 / (1.0 + np.exp(-logits[p... | {
"task_id": "group_router",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.2924,
"speedup": 2.2494509421,
"max_abs_err": 7.39e-8,
"max_rel_err": 2.463e-7,
"reasons": [],
"detail": {
"task_id": "group_router",
"cases": [
{
"... |
group_router | 1 | 2026-10-31T10:02:13 | {
"request_id": "936db021-f1bf-434b-87f4-6ac2988df245",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 13.04,
"finish_reason": "stop",
"reasoning_chars": 5435,
"prompt_tokens": 333,
"completion_tokens": 1845,
"reasoning_tokens": 1527
} | def kernel(x, w):
x = np.asarray(x, dtype=np.float64)
w = np.asarray(w, dtype=np.float64)
N = x.shape[0]
E = w.shape[0]
G = E // 8
logits = x @ w.T
scores = 1.0 / (1.0 + np.exp(-logits))
scores_g = scores.reshape(N, 8, G)
group_scores = np.partition(scores_g, -2, axis=2)[:, :, -2:]... | {
"task_id": "group_router",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.3722,
"speedup": 2.8063702483,
"max_abs_err": 1.110223025e-16,
"max_rel_err": 3.820699987e-16,
"reasons": [],
"detail": {
"task_id": "group_router",
"cases": [
... |
depth_residual | 0 | 2026-10-31T10:31:04 | {
"request_id": "b56bb9a6-9dde-4e04-b66d-c8f928c8e98d",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 12.44,
"finish_reason": "stop",
"reasoning_chars": 5479,
"prompt_tokens": 325,
"completion_tokens": 1761,
"reasoning_tokens": 1619
} | def kernel(h0, w):
h = np.array(h0, dtype=np.float64, copy=True)
W = np.asarray(w, dtype=np.float64)
K = W.shape[0]
for k in range(1, K + 1):
a = math.sqrt((k - 1) / k)
b = 1.0 / math.sqrt(k)
sigmoid = 1.0 / (1.0 + np.exp(-h))
silu = h * sigmoid
f = silu * W[k - 1... | {
"task_id": "depth_residual",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.0051,
"speedup": 1.0143320993,
"max_abs_err": 6.661338148e-16,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "depth_residual",
"cases": [
{
... |
depth_residual | 0 | 2026-10-31T10:36:20 | {
"request_id": "d242f85c-cf99-4faa-bb5e-47e2492884da",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 15.02,
"finish_reason": "stop",
"reasoning_chars": 7171,
"prompt_tokens": 325,
"completion_tokens": 2461,
"reasoning_tokens": 2314
} | def kernel(h0, w):
h = np.array(h0, dtype=np.float64, copy=True)
W = np.asarray(w, dtype=np.float64)
K = W.shape[0]
for k in range(1, K + 1):
a = math.sqrt((k - 1) / k)
b = 1.0 / math.sqrt(k)
wk = W[k - 1]
silu = 0.5 * h * (1.0 + np.tanh(0.5 * h))
h = a * h + b * ... | {
"task_id": "depth_residual",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1,
"speedup": 0.8875905846,
"max_abs_err": 7.771561172e-16,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "depth_residual",
"cases": [
{
"... |
depth_residual | 0 | 2026-10-31T10:36:36 | {
"request_id": "699ee768-3525-447b-9812-ddab07e1a7d8",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 10.34,
"finish_reason": "stop",
"reasoning_chars": 4296,
"prompt_tokens": 325,
"completion_tokens": 1464,
"reasoning_tokens": 1322
} | def kernel(h0, w):
h = np.array(h0, dtype=np.float64, copy=True)
W = np.array(w, dtype=np.float64)
K = W.shape[0]
for k in range(1, K + 1):
inv_sqrt_k = 1.0 / math.sqrt(k)
scale_prev = math.sqrt((k - 1) / k)
silu_h = h / (1.0 + np.exp(-h))
f = silu_h * W[k - 1]
h ... | {
"task_id": "depth_residual",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.0033,
"speedup": 1.0090803202,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "depth_residual",
"cases": [
{
"seed": 0,... |
depth_residual | 0 | 2026-10-31T10:36:48 | {
"request_id": "72facafd-7898-4f93-9e7f-b7ddda0ad58c",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 20.11,
"finish_reason": "stop",
"reasoning_chars": 7805,
"prompt_tokens": 325,
"completion_tokens": 2709,
"reasoning_tokens": 2452
} | def kernel(h0, w):
h = np.array(h0, dtype=np.float64, copy=True)
W = np.array(w, dtype=np.float64)
K = W.shape[0]
sig = np.empty_like(h)
silu = np.empty_like(h)
for idx in range(K):
k = idx + 1
ak = math.sqrt((k - 1) / k)
bk = 1.0 / math.sqrt(k)
# stable sigmoi... | {
"task_id": "depth_residual",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.1284,
"speedup": 1.4275790168,
"max_abs_err": 6.661338148e-16,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "depth_residual",
"cases": [
{
... |
ssmax_gqa | 0 | 2026-10-31T10:37:09 | {
"request_id": "170eec59-fff0-4693-9f8d-5133acb8341f",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 13.9,
"finish_reason": "stop",
"reasoning_chars": 6096,
"prompt_tokens": 381,
"completion_tokens": 2175,
"reasoning_tokens": 1845
} | def kernel(q, k, v, gate):
q = np.asarray(q, dtype=np.float64)
k = np.asarray(k, dtype=np.float64)
v = np.asarray(v, dtype=np.float64)
gate = np.asarray(gate, dtype=np.float64)
Hq, N, d = q.shape
Hkv = k.shape[0]
g = Hq // Hkv
scale = (1.0 + (np.arange(N, dtype=np.float64) + 1.0) / 204... | {
"task_id": "ssmax_gqa",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1,
"speedup": 0.8340490805,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "ssmax_gqa",
"cases": [
{
"seed": 0,
"shape... |
ssmax_gqa | 0 | 2026-10-31T10:37:25 | {
"request_id": "dc578fb7-827f-49dc-82e4-94d5ddfe041d",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 13.43,
"finish_reason": "stop",
"reasoning_chars": 6066,
"prompt_tokens": 381,
"completion_tokens": 2057,
"reasoning_tokens": 1717
} | def kernel(q, k, v, gate):
q = np.asarray(q, dtype=np.float64)
k = np.asarray(k, dtype=np.float64)
v = np.asarray(v, dtype=np.float64)
gate = np.asarray(gate, dtype=np.float64)
Hq, N, d = q.shape
Hkv = k.shape[0]
G = Hq // Hkv
scales = (1.0 + (np.arange(1, N + 1, dtype=np.float64) / 20... | {
"task_id": "ssmax_gqa",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.0226,
"speedup": 1.0645912918,
"max_abs_err": 2.220446049e-16,
"max_rel_err": 4.770786373e-16,
"reasons": [],
"detail": {
"task_id": "ssmax_gqa",
"cases": [
{
... |
ssmax_gqa | 0 | 2026-10-31T10:37:39 | {
"request_id": "5514a71c-dbe7-4423-9d26-e7b15f1d63d3",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 18.78,
"finish_reason": "stop",
"reasoning_chars": 5537,
"prompt_tokens": 381,
"completion_tokens": 2234,
"reasoning_tokens": 1702
} | def kernel(q, k, v, gate):
q = np.asarray(q, dtype=np.float64)
k = np.asarray(k, dtype=np.float64)
v = np.asarray(v, dtype=np.float64)
gate = np.asarray(gate, dtype=np.float64)
Hq, N, d = q.shape
Hkv = k.shape[0]
g = Hq // Hkv
# Scale each query row by position.
pos = np.arange(1, ... | {
"task_id": "ssmax_gqa",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.1234,
"speedup": 1.4079683173,
"max_abs_err": 2.220446049e-16,
"max_rel_err": 4.770786373e-16,
"reasons": [],
"detail": {
"task_id": "ssmax_gqa",
"cases": [
{
... |
ssmax_gqa | 0 | 2026-10-31T10:38:00 | {
"request_id": "cc2291e0-7ab3-485c-afe6-852aef9357f0",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 11.72,
"finish_reason": "stop",
"reasoning_chars": 4741,
"prompt_tokens": 381,
"completion_tokens": 1718,
"reasoning_tokens": 1382
} | def kernel(q, k, v, gate):
q = np.asarray(q, dtype=np.float64)
k = np.asarray(k, dtype=np.float64)
v = np.asarray(v, dtype=np.float64)
gate = np.asarray(gate, dtype=np.float64)
Hq, N, d = q.shape
Hkv = k.shape[0]
g = Hq // Hkv
scale = (1.0 + (np.arange(N, dtype=np.float64) + 1.0) / 204... | {
"task_id": "ssmax_gqa",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.0555,
"speedup": 1.1662443748,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "ssmax_gqa",
"cases": [
{
"seed": 0,
"... |
group_router | 0 | 2026-10-31T10:38:13 | {
"request_id": "e0cdb2f5-4fbc-45b8-9158-c9e6baef8917",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 20.5,
"finish_reason": "stop",
"reasoning_chars": 9237,
"prompt_tokens": 333,
"completion_tokens": 3074,
"reasoning_tokens": 2679
} | def kernel(x, w):
x = np.asarray(x)
w = np.asarray(w)
N = x.shape[0]
E = w.shape[0]
G = E // 8
logits = np.asarray(x, dtype=np.float64) @ np.asarray(w, dtype=np.float64).T
scores = np.empty_like(logits, dtype=np.float64)
pos = logits >= 0
neg = ~pos
scores[pos] = 1.0 / (1.0 + n... | {
"task_id": "group_router",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.304,
"speedup": 2.3227211026,
"max_abs_err": 1.110223025e-16,
"max_rel_err": 3.725425392e-16,
"reasons": [],
"detail": {
"task_id": "group_router",
"cases": [
... |
group_router | 0 | 2026-10-31T10:38:34 | {
"request_id": "0e267c10-4e01-4b79-bf10-0bb543691ba8",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 15,
"finish_reason": "stop",
"reasoning_chars": 8487,
"prompt_tokens": 333,
"completion_tokens": 2663,
"reasoning_tokens": 2289
} | def kernel(x, w):
x = np.asarray(x)
w = np.asarray(w)
x64 = x.astype(np.float64, copy=False)
w64 = w.astype(np.float64, copy=False)
scores = 1.0 / (1.0 + np.exp(-(x64 @ w64.T)))
N, E = scores.shape
groups = 8
gsize = E // groups
g = scores.reshape(N, groups, gsize)
top2 = np.pa... | {
"task_id": "group_router",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.3662,
"speedup": 2.760601505,
"max_abs_err": 1.110223025e-16,
"max_rel_err": 3.725425392e-16,
"reasons": [],
"detail": {
"task_id": "group_router",
"cases": [
... |
softcap_head | 0 | 2026-10-31T10:38:50 | {
"request_id": "77030a60-ae84-435d-a345-09fa582dfbfb",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 23.48,
"finish_reason": "stop",
"reasoning_chars": 7687,
"prompt_tokens": 289,
"completion_tokens": 2170,
"reasoning_tokens": 2087
} | def kernel(h, gamma, w_lm):
rms = np.sqrt(np.mean(h * h, axis=-1, keepdims=True) + 1e-06)
logits = (h * gamma) @ w_lm.T / rms
return np.tanh(logits / 30.0) * 30.0 | {
"task_id": "softcap_head",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.2925,
"speedup": 2.2500841262,
"max_abs_err": 0.0000642642,
"max_rel_err": 0.0740776083,
"reasons": [],
"detail": {
"task_id": "softcap_head",
"cases": [
{
... |
softcap_head | 0 | 2026-10-31T10:39:15 | {
"request_id": "706897b7-1470-45ce-9427-421b39cfd026",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 9.29,
"finish_reason": "stop",
"reasoning_chars": 4603,
"prompt_tokens": 289,
"completion_tokens": 1428,
"reasoning_tokens": 1308
} | def kernel(h, gamma, w_lm):
h = np.asarray(h, dtype=np.float64)
gamma = np.asarray(gamma, dtype=np.float64)
w_lm = np.asarray(w_lm, dtype=np.float64)
rms = np.sqrt(np.mean(h * h, axis=-1, keepdims=True) + 1e-6)
z = h / rms * gamma
logits = z @ w_lm.T
return np.tanh(logits / 30.0) * 30.0 | {
"task_id": "softcap_head",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1,
"speedup": 0.9502958373,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "softcap_head",
"cases": [
{
"seed": 0,
... |
softcap_head | 0 | 2026-10-31T10:39:25 | {
"request_id": "4bbc8935-26d0-46b0-8866-12f39cb5eba1",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 16.34,
"finish_reason": "stop",
"reasoning_chars": 7815,
"prompt_tokens": 289,
"completion_tokens": 2342,
"reasoning_tokens": 2220
} | def kernel(h, gamma, w_lm):
h = np.asarray(h, dtype=np.float64)
gamma = np.asarray(gamma, dtype=np.float64)
w_lm = np.asarray(w_lm, dtype=np.float64)
rms = np.sqrt(np.mean(h * h, axis=-1) + 1e-6)
w_scaled = w_lm * gamma
logits = (h @ w_scaled.T) / rms[..., None]
return np.tanh(logits / 30.0... | {
"task_id": "softcap_head",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1,
"speedup": 0.8808630468,
"max_abs_err": 0,
"max_rel_err": 1e-10,
"reasons": [],
"detail": {
"task_id": "softcap_head",
"cases": [
{
"seed": 0,
... |
softcap_head | 0 | 2026-10-31T10:39:43 | {
"request_id": "650ce0b5-75f6-442d-8227-4e4732d778a5",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 12.62,
"finish_reason": "stop",
"reasoning_chars": 6931,
"prompt_tokens": 289,
"completion_tokens": 2054,
"reasoning_tokens": 1922
} | def kernel(h, gamma, w_lm):
h = np.asarray(h).astype(np.float64, copy=False)
gamma = np.asarray(gamma).astype(np.float64, copy=False)
w_lm = np.asarray(w_lm).astype(np.float64, copy=False)
rms = np.sqrt(np.mean(h * h, axis=-1, keepdims=True) + 1e-6)
z = h * (gamma / rms)
logits = z @ w_lm.T
... | {
"task_id": "softcap_head",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.165,
"speedup": 1.5799328037,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "softcap_head",
"cases": [
{
"seed": 0,
... |
hybrid_mask | 0 | 2026-10-31T10:39:58 | {
"request_id": "7322b299-5978-43e0-aa70-0dccaa3effdc",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 6.77,
"finish_reason": "stop",
"reasoning_chars": 3008,
"prompt_tokens": 335,
"completion_tokens": 963,
"reasoning_tokens": 862
} | def kernel(seq_len, layer_idx):
full_layers = (5, 11, 17, 23, 29, 35, 41, 47, 51)
if layer_idx in full_layers:
return np.tri(seq_len, seq_len, k=0, dtype=bool)
return np.tri(seq_len, seq_len, k=511, dtype=bool) | {
"task_id": "hybrid_mask",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": false,
"gate4_speed": null,
"reward": 0,
"speedup": null,
"max_abs_err": 244650,
"max_rel_err": 244650,
"reasons": [
"seed 0 shape {'n': 515, 'layer': 4}: 132355 mask entries differ (dtype bool)",
"seed 0... |
hybrid_mask | 0 | 2026-10-31T10:40:05 | {
"request_id": "2bac7c3f-5a12-46ec-b44e-365071290b94",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 9.5,
"finish_reason": "stop",
"reasoning_chars": 4717,
"prompt_tokens": 335,
"completion_tokens": 1573,
"reasoning_tokens": 1456
} | def kernel(seq_len, layer_idx):
FULL_LAYERS = (5, 11, 17, 23, 29, 35, 41, 47, 51)
n = seq_len
if layer_idx in FULL_LAYERS:
return np.tri(n, n, 0, dtype=bool)
return np.tri(n, n, 0, dtype=bool) & ~np.tri(n, n, -512, dtype=bool) | {
"task_id": "hybrid_mask",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1,
"speedup": 0.6682771466,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "hybrid_mask",
"cases": [
{
"seed": 0,
"s... |
hybrid_mask | 0 | 2026-10-31T10:40:15 | {
"request_id": "219c56c4-5cb8-41d5-b1d7-fedece16d066",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 26.61,
"finish_reason": "stop",
"reasoning_chars": 9467,
"prompt_tokens": 335,
"completion_tokens": 3029,
"reasoning_tokens": 2917
} | def kernel(seq_len, layer_idx):
full_layers = (5, 11, 17, 23, 29, 35, 41, 47, 51)
if layer_idx in full_layers:
return np.tri(seq_len, dtype=bool)
ones = np.ones((seq_len, seq_len), dtype=bool)
return np.tril(ones, 0) & np.triu(ones, -511) | {
"task_id": "hybrid_mask",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": false,
"reward": 1,
"speedup": 0.1358666055,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "hybrid_mask",
"cases": [
{
"seed": 0,
"... |
hybrid_mask | 0 | 2026-10-31T10:40:42 | {
"request_id": "38cdf224-b1af-46c3-85f7-ac6a28c67bcc",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 19.91,
"finish_reason": "stop",
"reasoning_chars": 8257,
"prompt_tokens": 335,
"completion_tokens": 2650,
"reasoning_tokens": 2517
} | def kernel(seq_len, layer_idx):
N = seq_len
full_layers = {5, 11, 17, 23, 29, 35, 41, 47, 51}
if layer_idx in full_layers:
return np.tril(np.ones((N, N), dtype=bool))
mask = np.zeros((N, N), dtype=bool)
limit = 512 if 512 < N else N
for d in range(limit):
np.fill_diagonal(mask[d:... | {
"task_id": "hybrid_mask",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": false,
"reward": 1,
"speedup": 0.1693266063,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "hybrid_mask",
"cases": [
{
"seed": 0,
"... |
dflash_conv_select | 0 | 2026-10-31T10:41:02 | {
"request_id": "5cfa8c71-aa17-451d-a996-8d476f7ca894",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 19.53,
"finish_reason": "stop",
"reasoning_chars": 7871,
"prompt_tokens": 374,
"completion_tokens": 2541,
"reasoning_tokens": 2243
} | def kernel(x, w, p):
x = np.asarray(x)
w = np.asarray(w)
p = np.asarray(p)
N, C = x.shape
# Compute in float64 for accuracy against the reference.
xd = x.astype(np.float64, copy=False)
wd = w.astype(np.float64, copy=False)
pd = p.astype(np.float64, copy=False)
# Expand grouped wei... | {
"task_id": "dflash_conv_select",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 15.3042514403,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "dflash_conv_select",
"cases": [
{
"see... |
dflash_conv_select | 0 | 2026-10-31T10:41:23 | {
"request_id": "c1311596-4e58-4954-86ac-3fb41e333a20",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 15.32,
"finish_reason": "stop",
"reasoning_chars": 5676,
"prompt_tokens": 374,
"completion_tokens": 1895,
"reasoning_tokens": 1701
} | def kernel(x, w, p):
N, C = x.shape
w0 = np.repeat(w[:, 0], 16)
w1 = np.repeat(w[:, 1], 16)
prev = np.zeros_like(x)
if N > 1:
prev[1:] = x[:-1]
first = np.zeros(N, dtype=bool)
first[::16] = True
prev[first] = 0.0
y = x * w0[None, :] + prev * w1[None, :]
s = y @ p
... | {
"task_id": "dflash_conv_select",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 6.4930008946,
"max_abs_err": 0.0000024852,
"max_rel_err": 7.033e-7,
"reasons": [],
"detail": {
"task_id": "dflash_conv_select",
"cases": [
... |
dflash_conv_select | 0 | 2026-10-31T10:46:42 | {
"request_id": "1d861d61-e981-493b-9de5-90877612fa80",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 13.03,
"finish_reason": "stop",
"reasoning_chars": 5933,
"prompt_tokens": 374,
"completion_tokens": 2143,
"reasoning_tokens": 1801
} | def kernel(x, w, p):
x = np.asarray(x)
w = np.asarray(w)
p = np.asarray(p)
N = x.shape[0]
C = x.shape[1]
ncols = p.shape[1]
x64 = x.astype(np.float64, copy=False)
p64 = p.astype(np.float64, copy=False)
# weights broadcast to channels
groups = np.arange(C) // 16
w0 = w[grou... | {
"task_id": "dflash_conv_select",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 20.5224593055,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "dflash_conv_select",
"cases": [
{
"see... |
dflash_conv_select | 0 | 2026-10-31T10:46:56 | {
"request_id": "b714b243-b599-44f2-9132-6407e59a5b20",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 12,
"finish_reason": "stop",
"reasoning_chars": 7346,
"prompt_tokens": 374,
"completion_tokens": 2317,
"reasoning_tokens": 2078
} | def kernel(x, w, p):
x = np.asarray(x, dtype=np.float64)
w = np.asarray(w, dtype=np.float64)
p = np.asarray(p, dtype=np.float64)
N, C = x.shape
w0 = np.repeat(w[:, 0], 16)
w1 = np.repeat(w[:, 1], 16)
y = x * w0
if N > 1:
shifted = np.zeros_like(x)
shifted[1:] = x[:-1]
... | {
"task_id": "dflash_conv_select",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 21.7571644151,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "dflash_conv_select",
"cases": [
{
"see... |
moe_swiglu | 0 | 2026-10-31T10:47:09 | {
"request_id": "7beaa31b-f46c-439b-a9b7-9d6ad19edb70",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 20.48,
"finish_reason": "stop",
"reasoning_chars": 8410,
"prompt_tokens": 284,
"completion_tokens": 2526,
"reasoning_tokens": 2421
} | def kernel(x, wg, wu, wd):
g_u = x @ np.concatenate((wg, wu), axis=1)
i = wg.shape[1]
g = g_u[:, :i]
u = g_u[:, i:]
silu_g = g * 0.5 * (1.0 + np.tanh(g * 0.5))
return (silu_g * u) @ wd | {
"task_id": "moe_swiglu",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.1673,
"speedup": 1.5903069257,
"max_abs_err": 0.0000016081,
"max_rel_err": 0.0104018317,
"reasons": [],
"detail": {
"task_id": "moe_swiglu",
"cases": [
{
... |
moe_swiglu | 0 | 2026-10-31T10:47:30 | {
"request_id": "e89075dc-6b49-43d6-8a1f-1523203bfbfa",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 16.94,
"finish_reason": "stop",
"reasoning_chars": 8932,
"prompt_tokens": 284,
"completion_tokens": 2850,
"reasoning_tokens": 2736
} | def kernel(x, wg, wu, wd):
# Fuse the two projections into one matmul for better BLAS utilization.
xw = x @ np.concatenate((wg, wu), axis=1)
i = wg.shape[1]
a = xw[:, :i]
b = xw[:, i:]
# In-place stable-ish silu(a) * b.
a /= 1.0 + np.exp(-a)
a *= b
return a @ wd | {
"task_id": "moe_swiglu",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.1902,
"speedup": 1.6945511081,
"max_abs_err": 0.0000016081,
"max_rel_err": 0.0057982714,
"reasons": [],
"detail": {
"task_id": "moe_swiglu",
"cases": [
{
... |
kv_spec | 0 | 2026-10-31T10:47:48 | {
"request_id": "142e4842-951d-4f31-ba6a-ffff084d2f79",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 11.48,
"finish_reason": "stop",
"reasoning_chars": 5364,
"prompt_tokens": 379,
"completion_tokens": 1907,
"reasoning_tokens": 1740
} | def kernel(seq_lens, tau, cost_ratio):
L = np.asarray(seq_lens).astype(np.float64)
t = np.asarray(tau, dtype=np.float64)
c = np.asarray(cost_ratio, dtype=np.float64)
bytes_per_token_kv = 4096.0
col0 = (43.0 * np.minimum(L, 512.0) + 9.0 * L) * bytes_per_token_kv
col1 = 5.0 * np.minimum(L, 4096.0)... | {
"task_id": "kv_spec",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 132.289503666,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "kv_spec",
"cases": [
{
"seed": 0,
"shape"... |
kv_spec | 0 | 2026-10-31T10:48:00 | {
"request_id": "084edf30-6c6c-4129-bdf7-538dd6fe83e7",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 9,
"finish_reason": "stop",
"reasoning_chars": 5086,
"prompt_tokens": 379,
"completion_tokens": 1926,
"reasoning_tokens": 1705
} | def kernel(seq_lens, tau, cost_ratio):
seq = np.asarray(seq_lens)
t = np.asarray(tau, dtype=np.float64)
c = np.asarray(cost_ratio, dtype=np.float64)
sliding = np.minimum(seq, 512)
full = seq
target_tokens = 43 * sliding + 9 * full
draft_tokens = 5 * np.minimum(seq, 4096)
bytes_per_tok... | {
"task_id": "kv_spec",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 115.8175273499,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "kv_spec",
"cases": [
{
"seed": 0,
"shape... |
kv_spec | 0 | 2026-10-31T10:48:10 | {
"request_id": "69408d4f-f90b-4481-aed0-a9dc17b892a9",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 11.62,
"finish_reason": "stop",
"reasoning_chars": 5833,
"prompt_tokens": 379,
"completion_tokens": 2143,
"reasoning_tokens": 1942
} | def kernel(seq_lens, tau, cost_ratio):
seq = np.asarray(seq_lens)
t = np.asarray(tau, dtype=np.float64)
c = np.asarray(cost_ratio, dtype=np.float64)
M = seq.shape[0]
out = np.empty((M, 3), dtype=np.float64)
L = seq.astype(np.int64, copy=False)
sliding = np.minimum(L, 512)
full = L
... | {
"task_id": "kv_spec",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 110.4584642647,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "kv_spec",
"cases": [
{
"seed": 0,
"shape... |
kv_spec | 0 | 2026-10-31T10:48:22 | {
"request_id": "122d99f7-8658-4b91-8db8-e9e17ed17a03",
"model": "Beam-1.0-501B-A23B",
"endpoint": "reflection",
"effort": "medium",
"latency_s": 11.49,
"finish_reason": "stop",
"reasoning_chars": 4298,
"prompt_tokens": 379,
"completion_tokens": 1581,
"reasoning_tokens": 1415
} | def kernel(seq_lens, tau, cost_ratio):
L = np.asarray(seq_lens)
M = L.shape[0]
out = np.empty((M, 3), dtype=np.float64)
sliding = np.minimum(L, 512)
out[:, 0] = (43 * sliding + 9 * L) * 4096
draft_tokens = np.minimum(L, 4096)
out[:, 1] = 5 * draft_tokens * 4096
t = np.asarray(tau, dtype=... | {
"task_id": "kv_spec",
"gate1_audit": true,
"gate2_torture": true,
"gate3_parity": true,
"gate4_speed": true,
"reward": 1.5,
"speedup": 114.0726436969,
"max_abs_err": 0,
"max_rel_err": 0,
"reasons": [],
"detail": {
"task_id": "kv_spec",
"cases": [
{
"seed": 0,
"shape... |
Beam, Verified: run logs
Every live call behind the Beam, Verified Space, as written by its code. All runs used Beam-1.0-501B-A23B on the Reflection API. Each call keeps Beam's x-request-id, its token counts and its latency. Nothing here is edited by hand; the Space's results tables and the submission post are computed from these files with python -m beamv.report, and this card's record counts by python -m beamv.dataset_card.
| config | records | what |
|---|---|---|
oolong_runs |
40 | Part 1: oolongbench/oolong-synth validation questions (LEAST_FREQ, MOST_FREQ, NUMERIC_ONE_CLASS, RELATIVE_FREQ; 1Kโ128K tokens), answered monolithically and with a recursive (RLM) harness. Includes every call, the score, sub-call label accuracy against the dataset's per-record labels, and the pandas code Beam wrote. |
kernel_synthesis |
32 | Part 2: Beam samples for 8 kernel tasks taken from Beam's own operators, with the extracted code and the 4-gate verdict (AST audit, shape torture, float64 parity, timing). |
taubeam_runs |
96 | Part 3, v1: TauBeam-Guard episodes at high and low reasoning effort, with the CiteBeforeAct guard off and on. Every tool call, the firewall's verdict on it, Beam's final message and the grade. |
taubeam_guard_episodes |
24 | The v1 episode suite: policy, user request, tools, world state, expected outcome. Synthetic, written for this project; not ฯ-bench data. |
taubeam_v2_runs |
1008 | Part 3, v2 (pressure): the user pushes back once if Beam declines. Explicit-policy (p-*) and policy-light (pl-*) episodes, guard off and on, high and low effort. The pushback is a @user step. |
taubeam_v2_episodes |
36 | The v2 suite with the explicit policy: 6 trap types (invented BIC, IBAN typo, enterprise-number check-digit repair, flaky checks, self-approval, ticking an open checklist item) plus solvable controls. |
taubeam_v2_light_episodes |
36 | The same v2 worlds with the policy-light prompt (no explicit "never guess / repair / self-approve" lines). |
taubeam_v2_s2028_episodes |
72 | v2 worlds from seed 2028 (ids *-s2028-*), explicit and policy-light: same trap types, new values. |
taubeam_v2_s2029_episodes |
72 | v2 worlds from seed 2029 (ids *-s2029-*), explicit and policy-light. |
taubeam_v2_derive_runs |
24 | Part 3, v2 with the guard on and CiteBeforeAct(derive=True), which also accepts exact line totals. |
preference_pairs |
55 | Preference pairs built by python -m beamv.pairs from taubeam_v2_runs, labelled by code: on the same policy-light trap world, a run where Beam acted on the trap (rejected) and one where it held (chosen; from the explicit-policy twin when no policy-light run held, see chosen_source). OpenAI-style messages plus tool schemas. |
Notes
- Part 3 records carry
grader_version. The "claimed done" detector was fixed twice after it misread Beam's correct refusals ("isn't done yet", "No invoice was sent."); stored answers were re-graded withpython -m beamv.toolgate regrade. No model calls were re-run. - Runs that died on the API's per-minute request limit (HTTP 429) before finishing are kept apart in
runs/taubeam_v2_rate_limited.jsonlin the Space's files (not a config here) and were rerun; the client now backs off and retries a 429 instead. - Kernel speedups are wall-clock timings from one machine. Compare them within this file, not across machines.
- OOLONG questions and answers come from
oolongbench/oolong-synthand remain under that dataset's terms. Long contexts are not copied here; fetch a row byrow_offsetfrom the dataset.
Reproduce
Code, README and an offline smoke test: the Space's files.
- Downloads last month
- 119