Gauntlet qualification report

run 20260830T183302800258Z-1607abca · 2026-08-30T18:33:20+00:00 · target litert-host · board host:<redacted>:arm64

1Verdict

Limit	Declared	Measured	Status
accuracy.absolute	≥ 0.35	0.59	PASS
accuracy.delta_vs_golden	≤ 0.1	0.01881	PASS
accuracy.answer_flip_rate	≤ 0.25	0.07512	PASS
latency.p99_ms	≤ 60	9.453	PASS
placement.fallback_ops	≤ 0	0	PASS

Unset limits are reported as measured, not passed. NOT_MEASURED rows count as neither pass nor fail.

detect task: a 'flip' is a confident detection (conf >= 0.25) in one column with no same-class IoU >= 0.5 match in the other, either direction

accuracy metric: mAP@0.5 on the eval set (golden 0.6088, target 0.5900)

Redacted for publication: directory paths reduced to filenames and host names removed (model_path, artifact_path, budgets_path, board_id, host.node). Every SHA-256, SDK version, measurement and verdict is unmodified — the hashes, not the paths, are what identify the files this run used.

2Provenance

Item	Value
reference model	yolov8n_fp32.onnx sha256 1607abcabc7b3e3edf9ff56d3845d0167be7113a9722ca9bd6bd878bfc8f1cc9
compiled artifact	yolov8n_int8.tflite sha256 26d097b558ef6b39fd9b07affe3b75b095e03538610170b389b9f888e1363fb0
eval set	sha256 0b448a6934d978d5e278fa7fb59674eaf2e8761f5ff23672c3e9fc0e70d36e5e
calibration set	sha256 —
compiler	none (pre-quantized .tflite) 2.2.0
board	host:<redacted>:arm64
SDKs	numpy 2.5.2, onnxruntime 1.24.3, onnx 1.20.1, ai-edge-litert 2.2.0, ethos-u-vela 5.1.0
host	platform=macOS-26.5.2-arm64-arm-64bit, machine=arm64, python=3.12.7, node=<redacted>
Gauntlet version	0.2.0 git e34ca59e27ec28d99f65fe807795138ff8eaaa91-dirty
budgets file	budgets-yolov8n.yaml sha256 acb9015730e09d6c40ee8a82db66aa392b402d8ef0886db3662abdd067894f5f

3Accuracy deviation — three-column attribution

Column	Runtime	Version	Kernels	Executed	Output digest
golden	onnxruntime	1.24.3	cpu-fp32 (single-threaded, deterministic)	yes	c03d930454042159
scheme	litert (TFLite)	2.2.0	builtin-reference-int8	yes	3ba106022e3dbd29
target	litert-host	2.2.0	litert-host	yes	716cd14f43894f49

Comparison	n	max \|Δ\|	RMSE	SNR dB	cos min	flips	flip rate
Golden vs Scheme (quantization mathematics)	32	0.6448	0.004718	26.229	0.997527	22	0.6875
Scheme vs Target (silicon kernels & compiler)	32	0.4986	0.002516	31.677	0.999209	12	0.375
Golden vs Target (end-to-end)	32	0.6448	0.004686	26.289	0.997605	22	0.6875

Answer-flip rate of the qualified artifact vs Golden: 0.07512 [0.04225, 0.1127] 95% CI. Two models can score identically and still disagree on individual inputs; flips expose what aggregate accuracy hides.

4Operator placement

293operations

293accelerated

0CPU fallback

0unsupported

Every operation was placed on the accelerator. No silent CPU fallback in this build.

Placement reason, all rows: LiteRT host CPU (XNNPACK-optimized kernels)

Full placement map — all 293 operations

#	Op	Layer	Placement	Cycles (est)	Time µs	Precision
0	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
1	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
2	PAD	PAD	ACCELERATED	—	—	int8
3	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
4	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
5	MUL	MUL	ACCELERATED	—	—	int8
6	PAD	PAD	ACCELERATED	—	—	int8
7	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
8	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
9	MUL	MUL	ACCELERATED	—	—	int8
10	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
11	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
12	MUL	MUL	ACCELERATED	—	—	int8
13	SLICE	SLICE	ACCELERATED	—	—	int8
14	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
15	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
16	MUL	MUL	ACCELERATED	—	—	int8
17	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
18	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
19	MUL	MUL	ACCELERATED	—	—	int8
20	ADD	ADD	ACCELERATED	—	—	int8
21	SLICE	SLICE	ACCELERATED	—	—	int8
22	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
23	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
24	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
25	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
26	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
27	MUL	MUL	ACCELERATED	—	—	int8
28	PAD	PAD	ACCELERATED	—	—	int8
29	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
30	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
31	MUL	MUL	ACCELERATED	—	—	int8
32	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
33	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
34	MUL	MUL	ACCELERATED	—	—	int8
35	SLICE	SLICE	ACCELERATED	—	—	int8
36	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
37	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
38	MUL	MUL	ACCELERATED	—	—	int8
39	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
40	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
41	MUL	MUL	ACCELERATED	—	—	int8
42	ADD	ADD	ACCELERATED	—	—	int8
43	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
44	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
45	MUL	MUL	ACCELERATED	—	—	int8
46	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
47	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
48	MUL	MUL	ACCELERATED	—	—	int8
49	ADD	ADD	ACCELERATED	—	—	int8
50	SLICE	SLICE	ACCELERATED	—	—	int8
51	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
52	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
53	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
54	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
55	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
56	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
57	MUL	MUL	ACCELERATED	—	—	int8
58	PAD	PAD	ACCELERATED	—	—	int8
59	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
60	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
61	MUL	MUL	ACCELERATED	—	—	int8
62	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
63	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
64	MUL	MUL	ACCELERATED	—	—	int8
65	SLICE	SLICE	ACCELERATED	—	—	int8
66	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
67	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
68	MUL	MUL	ACCELERATED	—	—	int8
69	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
70	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
71	MUL	MUL	ACCELERATED	—	—	int8
72	ADD	ADD	ACCELERATED	—	—	int8
73	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
74	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
75	MUL	MUL	ACCELERATED	—	—	int8
76	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
77	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
78	MUL	MUL	ACCELERATED	—	—	int8
79	ADD	ADD	ACCELERATED	—	—	int8
80	SLICE	SLICE	ACCELERATED	—	—	int8
81	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
82	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
83	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
84	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
85	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
86	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
87	MUL	MUL	ACCELERATED	—	—	int8
88	PAD	PAD	ACCELERATED	—	—	int8
89	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
90	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
91	MUL	MUL	ACCELERATED	—	—	int8
92	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
93	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
94	MUL	MUL	ACCELERATED	—	—	int8
95	SLICE	SLICE	ACCELERATED	—	—	int8
96	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
97	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
98	MUL	MUL	ACCELERATED	—	—	int8
99	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
100	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
101	MUL	MUL	ACCELERATED	—	—	int8
102	ADD	ADD	ACCELERATED	—	—	int8
103	SLICE	SLICE	ACCELERATED	—	—	int8
104	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
105	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
106	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
107	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
108	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
109	MUL	MUL	ACCELERATED	—	—	int8
110	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
111	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
112	MUL	MUL	ACCELERATED	—	—	int8
113	MAX_POOL_2D	MAX_POOL_2D	ACCELERATED	—	—	int8
114	MAX_POOL_2D	MAX_POOL_2D	ACCELERATED	—	—	int8
115	MAX_POOL_2D	MAX_POOL_2D	ACCELERATED	—	—	int8
116	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
117	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
118	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
119	MUL	MUL	ACCELERATED	—	—	int8
120	RESIZE_NEAREST_NEIGHBOR	RESIZE_NEAREST_NEIGHBOR	ACCELERATED	—	—	int8
121	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
122	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
123	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
124	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
125	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
126	MUL	MUL	ACCELERATED	—	—	int8
127	SLICE	SLICE	ACCELERATED	—	—	int8
128	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
129	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
130	MUL	MUL	ACCELERATED	—	—	int8
131	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
132	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
133	MUL	MUL	ACCELERATED	—	—	int8
134	SLICE	SLICE	ACCELERATED	—	—	int8
135	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
136	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
137	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
138	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
139	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
140	MUL	MUL	ACCELERATED	—	—	int8
141	RESIZE_NEAREST_NEIGHBOR	RESIZE_NEAREST_NEIGHBOR	ACCELERATED	—	—	int8
142	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
143	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
144	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
145	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
146	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
147	MUL	MUL	ACCELERATED	—	—	int8
148	SLICE	SLICE	ACCELERATED	—	—	int8
149	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
150	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
151	MUL	MUL	ACCELERATED	—	—	int8
152	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
153	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
154	MUL	MUL	ACCELERATED	—	—	int8
155	SLICE	SLICE	ACCELERATED	—	—	int8
156	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
157	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
158	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
159	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
160	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
161	MUL	MUL	ACCELERATED	—	—	int8
162	PAD	PAD	ACCELERATED	—	—	int8
163	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
164	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
165	MUL	MUL	ACCELERATED	—	—	int8
166	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
167	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
168	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
169	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
170	MUL	MUL	ACCELERATED	—	—	int8
171	SLICE	SLICE	ACCELERATED	—	—	int8
172	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
173	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
174	MUL	MUL	ACCELERATED	—	—	int8
175	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
176	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
177	MUL	MUL	ACCELERATED	—	—	int8
178	SLICE	SLICE	ACCELERATED	—	—	int8
179	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
180	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
181	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
182	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
183	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
184	MUL	MUL	ACCELERATED	—	—	int8
185	PAD	PAD	ACCELERATED	—	—	int8
186	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
187	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
188	MUL	MUL	ACCELERATED	—	—	int8
189	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
190	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
191	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
192	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
193	MUL	MUL	ACCELERATED	—	—	int8
194	SLICE	SLICE	ACCELERATED	—	—	int8
195	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
196	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
197	MUL	MUL	ACCELERATED	—	—	int8
198	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
199	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
200	MUL	MUL	ACCELERATED	—	—	int8
201	SLICE	SLICE	ACCELERATED	—	—	int8
202	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
203	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
204	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
205	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
206	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
207	MUL	MUL	ACCELERATED	—	—	int8
208	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
209	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
210	MUL	MUL	ACCELERATED	—	—	int8
211	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
212	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
213	MUL	MUL	ACCELERATED	—	—	int8
214	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
215	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
216	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
217	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
218	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
219	MUL	MUL	ACCELERATED	—	—	int8
220	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
221	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
222	MUL	MUL	ACCELERATED	—	—	int8
223	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
224	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
225	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
226	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
227	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
228	MUL	MUL	ACCELERATED	—	—	int8
229	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
230	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
231	MUL	MUL	ACCELERATED	—	—	int8
232	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
233	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
234	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
235	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
236	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
237	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
238	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
239	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
240	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
241	MUL	MUL	ACCELERATED	—	—	int8
242	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
243	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
244	MUL	MUL	ACCELERATED	—	—	int8
245	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
246	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
247	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
248	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
249	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
250	MUL	MUL	ACCELERATED	—	—	int8
251	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
252	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
253	MUL	MUL	ACCELERATED	—	—	int8
254	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
255	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
256	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
257	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
258	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
259	MUL	MUL	ACCELERATED	—	—	int8
260	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
261	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
262	MUL	MUL	ACCELERATED	—	—	int8
263	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
264	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
265	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
266	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
267	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
268	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
269	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
270	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
271	TRANSPOSE	TRANSPOSE	ACCELERATED	—	—	int8
272	SOFTMAX	SOFTMAX	ACCELERATED	—	—	int8
273	CONV_2D	CONV_2D	ACCELERATED	—	—	int8
274	RESHAPE	RESHAPE	ACCELERATED	—	—	int8
275	SLICE	SLICE	ACCELERATED	—	—	int8
276	SLICE	SLICE	ACCELERATED	—	—	int8
277	SUB	SUB	ACCELERATED	—	—	int8
278	ADD	ADD	ACCELERATED	—	—	int8
279	ADD	ADD	ACCELERATED	—	—	int8
280	MUL	MUL	ACCELERATED	—	—	int8
281	SUB	SUB	ACCELERATED	—	—	int8
282	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
283	MUL	MUL	ACCELERATED	—	—	int8
284	LOGISTIC	LOGISTIC	ACCELERATED	—	—	int8
285	MUL	MUL	ACCELERATED	—	—	int8
286	QUANTIZE	QUANTIZE	ACCELERATED	—	—	int8
287	CONCATENATION	CONCATENATION	ACCELERATED	—	—	int8
288	DEQUANTIZE	DEQUANTIZE	ACCELERATED	—	—	int8
289	DELEGATE	DELEGATE	ACCELERATED	—	—	int8
290	DELEGATE	DELEGATE	ACCELERATED	—	—	int8
291	DELEGATE	DELEGATE	ACCELERATED	—	—	int8
292	DELEGATE	DELEGATE	ACCELERATED	—	—	int8

Placement from compiler output. Per-layer drift inside a fused NPU command stream is not observable; deviation attribution is per-segment, not per-layer.

5Latency

Percentile	ms (95% bootstrap CI)
p50	9.264 [9.26, 9.268] 95% CI
p95	9.38 [9.368, 9.394] 95% CI
p99	9.453 [9.421, 9.49] 95% CI
mean	9.172 [9.148, 9.193] 95% CI

960 measured inferences; 3 warm-up passes discarded; 0 retries.

6Memory — capacity

Quantity	Measured
peak RAM	— MB
flash footprint	— MB

7Memory — bandwidth

Not measured — EMC/DDR counters are available on boards exposing them (e.g. Jetson via tegrastats).

8Thermal

Not measured (no sustained-load soak on this run).

9Worst offenders

Input	max \|Δ\|	argmax ref	argmax artifact	Columns
000000000092	0.6208	6189	6238	golden vs target
000000000030	0.5768	6270	6129	golden vs target
000000000078	0.5092	6295	6294	golden vs target
000000000136	0.4834	6189	479	golden vs target
000000000036	0.4505	8370	8320	golden vs target
000000000081	0.4473	999	519	golden vs target
000000000034	0.4348	6276	6292	golden vs target
000000000089	0.4278	6189	6188	golden vs target
000000000074	0.41	6189	6277	golden vs target
000000000144	0.3768	519	399	golden vs target

The specific inputs with the largest deviation, for reproduction.

10Regression diff

baseline run 20260829T235659473087Z-1607abca · 2026-08-29T23:57:17+00:00 · artifact sha256 26d097b558ef6b39…

Measurement	Baseline	Current	Δ	Δ %
accuracy.absolute	0.59	0.59	0	0
accuracy.answer_flip_rate	0.07512	0.07512	0	0
accuracy.delta_vs_golden	0.01881	0.01881	0	0
latency.mean_ms	9.351	9.172	-0.1796	-1.92
latency.p50_ms	9.377	9.264	-0.1128	-1.2
latency.p95_ms	9.462	9.38	-0.0821	-0.868
latency.p99_ms	9.507	9.453	-0.05389	-0.567
placement.fallback_ops	0	0	0	—
placement.total_ops	293	293	0	0
placement.unsupported_ops	0	0	0	—