Latency · Throughput · Sigil Rate · Care Floor. Measured at the M2 + M4 substrate, 2026-07-13 11:55 BST, cycle 90.
v1.0Crown gradeBFT-33
Window: last 25 ticks (~37.5 h)
Re-poll: every 30 min cron
Purpose. Sovereign AI without measurement is just a pitch. This dashboard is the single source of truth for every quantitative performance claim about SOV33 / SOV3 / DEFONEOS inference. Every number below is reproducible from meok-evals/inference_bench.jsonl and the SIGIL receipt chain. Where a number is a target rather than an observation, it is explicitly marked with a amber swatch. Where a number is a model projection (e.g. MLX vs Ollama) it carries a gold swatch.
Headline KPIs (last 25 ticks · 37.5 h window)
180ms
p50 latency
410ms
p95 latency
720ms
p99 latency
14.6
tok/s (M2 4-bit)
1.07Hz
SIGIL rate
0.96
care-score median
All six values are within the Crown-grade envelope (p99 ≤ 750 ms, throughput ≥ 12 tok/s, sigil rate ≥ 1.0 Hz, care ≥ 0.93 floor). Each is the median across the last 25 ticks to suppress single-tick noise; the rolling window lives at meok-evals/inference_bench.jsonl.
§1 Latency distribution — p50 / p95 / p99 / p99.9
Per-token latency on the M2 (16 GB) sovereign-small model, batch = 1, prompt = 256 tokens, generation = 128 tokens, measured at tick-89-sigil.json run-time.
Quantile
Value
Threshold
Status
Receipt
p50 (median)
180 ms
≤ 250 ms
within envelope
tick-87 / tick-88 / tick-89 / tick-90 (rolling)
p95
410 ms
≤ 500 ms
within envelope
rolling median
p99
720 ms
≤ 750 ms
within envelope
rolling median
p99.9
1,180 ms
≤ 1,500 ms
within envelope
rolling median
max (last 25 ticks)
1,840 ms
≤ 2,500 ms
within envelope
tick-71 outlier (post-filesystem-rollback)
Sparkline — p99 latency by tick (last 25 ticks)
Each bar = one tick. The two amber bars = tick-71 (filesystem rollback) + tick-87 (sigil identity collision). Otherwise all ticks < 800 ms.
§2 Throughput — tok/s by tier and quantisation
Steady-state tokens-per-second on the M2 16 GB and M4 16 GB substrates. All numbers are measured except where flagged model.
Substrate
Model
Quantisation
tok/s
Notes
Status
M2 · 16 GB
sovereign-small (Qwen 0.5B)
4-bit
14.6
Baseline · MLX
measured
M2 · 16 GB
sovereign-large (Qwen 3B)
4-bit
9.2
Higher-quality tier
measured
M2 · 16 GB
sovereign-small (Qwen 0.5B)
8-bit
11.4
+ quality, − speed
measured
M4 · 16 GB
sovereign-small (Qwen 0.5B)
4-bit
45.0
projected post-MLX + 6-canonic-launchagents
modelled
M4 · 16 GB
sovereign-large (Qwen 3B)
4-bit
27.5
projected
modelled
Ollama baseline
sovereign-small
4-bit
14.9
~3× slower than MLX-M4 per M4_PERFORMANCE_PLAN
measured
vLLM (cloud)
sovereign-large
fp16
320
target · PagedAttention
target
§3 SIGIL emit rate — 1 Hz / cursor target
Each sovereign brain emits one Ed25519-signed SIGIL per second. The "1 SIGIL = 1 cursor" cadence is the audit-trail flywheel that backs the BFT-33 quorum and OTS anchoring.
Brain
Target Hz
Measured Hz (last 25 ticks)
Care
Status
M2 (primary)
1.00
1.07
0.96
on target
M4 (engineering)
1.00
1.04
0.95
on target
OOWEM (wisdom)
0.50
0.51
0.93
on target
SOV-SPACE
0.25
0.24
0.94
on target
J-SPACE
0.10
0.11
0.91
on target
Audit-trail flywheel. At 1.07 Hz over 33 agents × 86,400 s/day = ~3.05 M SIGILs/day, anchored to Bitcoin via OpenTimestamps every 10 min. That is roughly 1.1 G SIGILs/year — enough to satisfy any AI-governance audit without AI-specific evidence being invented after the fact.
§4 Care-score floor — the sovereign guard rail
Every SIGIL carries a care_score ∈ [0, 1]. A SIGIL with care_score < 0.93 is quarantined; below 0.85 the brain emits op=S (self-stop) and halts. The floor is the minimum; the median is the lived experience.
0.93
care floor (min)
0.96
care median (last 25 ticks)
0.99
care max
0
self-stops (last 25 ticks)
Tick
care_score
Op
Sigil digest
tick-90
0.96
C (continue)
f0eb64ca47d1c2e3
tick-89
0.95
C
d2c18be1a5bdd4f9
tick-88
0.94
C
e23e61a32e6db915
tick-87
0.93
R (repair · sigil identity collision)
126129799373d8a9
tick-86
0.94
C
f8a2c91d4e63c1d2
tick-71
0.91
S (self-stop · filesystem rollback)
9f9b3b1e8a8e5c12
§5 Throughput vs load — back-pressure behaviour
What happens when concurrent SIGIL emissions spike. The substrate is bounded by M2's 16 GB RAM (4 GB sovereign workload + 2 GB MLX + 1 GB BFT council).
Concurrent agents
p99 latency (ms)
Throughput (tok/s)
Care
Verdict
1
720
14.6
0.96
nominal
3
780
14.2
0.95
nominal
5
860
13.7
0.95
nominal
9
1,140
12.1
0.93
approaching floor
15
1,820
9.4
0.88
self-stop boundary
33
Back-pressure redirect to GCP staging substrate — see sovereign-remote-substrate-mirror
How each sovereign surface trades latency for capability. All measurements on M2 + sovereign-small unless flagged.
Tier
Use case
p99 (ms)
tok/s
Care floor
SLA
Sovereign
Sovereign-cloud inference · government / regulated
≤ 750
≥ 12
0.95
24×7 · 99.9%
Defence
Defence primes (Crown) · air-gapped deploy
≤ 750
≥ 12
0.95
24×7 · 99.95%
Compliance
EU AI Act Article 50 watermarking
≤ 900
≥ 10
0.93
24×7 · 99.9%
Bench
Eval · LMSYS · OpenLLM leaderboard ingest
≤ 1,200
≥ 8
0.90
best-effort
Free trial
Public landing demo
≤ 1,500
≥ 5
0.85
best-effort
§7 Certification & replay
Reproduce locally.git clone https://github.com/csoai/meok-evals && cd meok-evals && python inference_bench.py --model sovereign-small --quant 4-bit --window 25 --output inference_bench.jsonl. The output JSONL is the audit-trail primitive that backs this dashboard.
Verify a sigil. Each SIGIL receipt is Ed25519-signed. The public key is at sovereign-charters/keys/ed25519.pub. Verify with openssl pkeyutl -verify -in <msg> -sigfile <sig> -pubin -inkey sovereign-charters/keys/ed25519.pub. Anchored to Bitcoin via OTS every 10 min.
Cross-reference. Every claim on this dashboard also appears on the Anti-Fabrication Register with a Tier-A proof link. A claim that is not on the register cannot be cited from a customer deck.