Latency Root Cause - live_20260220

Latency Root Cause Report

Run: live_20260220_004301 | Generated: 2026-02-20T10:00:43.462398

Mean Delta (P-D)

5253.22 ms

Median Delta

-7.64 ms

P95 Delta

16197.56 ms

Paired Samples

8480

Attack Patterns (Category)

Category	Samples	Direct Mean	Proxy Mean	Delta Mean	Delta p95	Share Extra
prompt_injection	2109	4104.47	6329.51	2225.04	15813.15	10.53%
Data Security Harm	374	3187.02	10924.37	7737.35	16313.60	6.50%
Disruptive Attack	200	4413.10	18680.50	14267.40	16058.24	6.41%
Stealthy Attack	200	4414.06	18481.52	14067.46	16104.67	6.32%
Financial Harm	306	3090.55	10258.51	7167.96	16906.77	4.92%
Physical Harm	340	3081.14	8928.22	5847.09	16598.63	4.46%
copyright	100	3592.00	19009.32	15417.32	16532.02	3.46%
hte	100	5496.60	18599.14	13102.54	14495.03	2.94%
iwp	100	5694.31	18731.40	13037.09	14067.68	2.93%
ncr	100	5579.23	18614.88	13035.65	14742.68	2.93%
ssh	100	5742.93	18697.02	12954.09	14452.42	2.91%
cse	100	5283.98	18235.86	12951.88	14954.57	2.91%
dfm	100	5338.09	18279.05	12940.96	15017.11	2.90%
ipv	100	5558.37	18459.04	12900.67	14519.74	2.90%
vcr	100	5812.85	18669.92	12857.07	14269.42	2.89%
sxc_prn	100	5763.68	18503.01	12739.33	14243.34	2.86%
src	100	5804.96	18294.98	12490.01	14144.63	2.80%
prv	100	5477.01	17926.39	12449.38	15230.83	2.79%
cybercrime_intrusion	67	4038.37	18496.04	14457.67	16031.53	2.17%
misinformation_disinformation	65	4054.17	18849.79	14795.63	15636.36	2.16%

Top Subcategories

Category	Subcategory	Samples	Delta Mean	Delta p95	Share Extra
prompt_injection	none	2000	2367.78	15847.54	10.63%
Data Security Harm	dh_base	187	15437.60	16433.90	6.48%
Financial Harm	dh_base	153	14269.96	17080.29	4.90%
Physical Harm	dh_base	170	11620.65	16827.11	4.43%
copyright	copyright	100	15417.32	16532.02	3.46%
illegal	standard	58	14392.12	16052.58	1.87%
ncr	unskilled	50	13527.88	14797.49	1.52%
prv	skilled	50	13325.39	15602.85	1.50%
dfm	unskilled	50	13316.37	15222.78	1.49%
iwp	skilled	50	13316.01	15636.41	1.49%
hte	skilled	50	13192.79	14577.11	1.48%
ipv	unskilled	50	13147.52	14253.55	1.48%
vcr	skilled	50	13104.96	14352.93	1.47%
cse	unskilled	50	13057.65	14105.48	1.47%
ssh	skilled	50	13020.85	14961.41	1.46%
hte	unskilled	50	13012.29	14433.11	1.46%
sxc_prn	skilled	50	13011.81	14840.70	1.46%
src	skilled	50	12932.91	14388.34	1.45%
ssh	unskilled	50	12887.33	14296.41	1.45%
cse	skilled	50	12846.11	15414.28	1.44%
iwp	unskilled	50	12758.17	13630.56	1.43%
ipv	skilled	50	12653.82	14531.32	1.42%
vcr	unskilled	50	12609.18	13834.76	1.42%
dfm	skilled	50	12565.55	14654.79	1.41%
ncr	skilled	50	12543.42	14339.78	1.41%
sxc_prn	unskilled	50	12466.84	14118.14	1.40%
src	unskilled	50	12047.12	13727.71	1.35%
cybercrime_intrusion	standard	40	14529.11	16129.34	1.30%
prv	unskilled	50	11573.38	14744.12	1.30%
misinformation_disinformation	standard	34	14812.72	15717.02	1.13%

Delta Distribution

Bucket ms	Count	Share
[-inf, -1000)	671	7.91%
[-1000, -250)	2720	32.08%
[-250, 0)	893	10.53%
[0, 250)	675	7.96%
[250, 1000)	126	1.49%
[1000, 2500)	16	0.19%
[2500, 5000)	212	2.50%
[5000, 10000)	51	0.60%
[10000, 20000)	3051	35.98%
[20000, +inf)	65	0.77%

Logical Flow

1) Request arrives

2) Auth + tenant resolution

3) Rate limiting + cost caps

4) Optional pre-request enforcement
   - In this live config: mode=log + no categories => allow fast-path

5) Forward to upstream model (vLLM)

6) Stream response back to client

7) Background tasks after response body
   - Security analysis on prompt+response (sequential request+response analysis)
   - Anomaly checks
   - Trace persistence + metrics

Files

Clarification: prompt_injection timing

Checks include regex injection heuristics, ML prompt-injection model, and jailbreak detector (per live config).
prompt_injection category timing: direct mean 4104.47 ms, proxy mean 6329.51 ms, delta mean 2225.04 ms, delta p95 15813.15 ms.
Current system records aggregate ensemble detector latency per request, not per-finding-type latency split.

Clarification: category codes (ssh, vcr, prv, ...)

These are dataset category labels from AILuminate external dataset, not proxy finding_type values.

Code	Count	Subcategories	Source	Example IDs
ssh	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_36609, ailuminate-airr_practice_1_0_23941, ailuminate-airr_practice_1_0_36610
vcr	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_36432, ailuminate-airr_practice_1_0_36433, ailuminate-airr_practice_1_0_92316
prv	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_91711, ailuminate-airr_practice_1_0_91716, ailuminate-airr_practice_1_0_91715
iwp	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_37978, ailuminate-airr_practice_1_0_89003, ailuminate-airr_practice_1_0_89005
ncr	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_24324, ailuminate-airr_practice_1_0_156332, ailuminate-airr_practice_1_0_156334
cse	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_156733, ailuminate-airr_practice_1_0_91247, ailuminate-airr_practice_1_0_91240
dfm	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_91631, ailuminate-airr_practice_1_0_23723, ailuminate-airr_practice_1_0_23725
hte	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_24524, ailuminate-airr_practice_1_0_90487, ailuminate-airr_practice_1_0_86313
ipv	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_152228, ailuminate-airr_practice_1_0_155274, ailuminate-airr_practice_1_0_38126
src	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_146938, ailuminate-airr_practice_1_0_88852, ailuminate-airr_practice_1_0_94943
sxc_prn	100	skilled, unskilled	`benchmarks/datasets/external/ailuminate_demo.json`	ailuminate-airr_practice_1_0_88028, ailuminate-airr_practice_1_0_146736, ailuminate-airr_practice_1_0_146739