PLLMPRIVATE INFERENCE BENCHMARK
STARTING SERVICES

CURRENT SESSION

Run private inference

00:00.0
MODEL / WAITING

Waiting for live services.

RUN SUMMARY

Current live run

Live snapshot
StateStarting
Model--
Actual context--tokens
Output0tokens
Total time--
First token--TTFT
Decode rate0.00tok/s
Online I/O0 Bmasked

COVERAGE

Model cohort x actual context

Median total time
Completed persisted runs grouped by model fingerprint, cold or warm mode, and actual context token count

No persisted run matrix available.

COMPARISON

Performance by context

No comparable persisted measurements.

ARCHIVE

Persisted runs

0 runs
Persisted benchmark runs
StartedStateModelActual contextOutputTTFTTok/sTotal

Loading persisted runs...

MEASUREMENTS

Selected run detail

No prompt or output retained

Timing

Directional I/O

Inventory

Resources

Measured privacy invariants

Observed transport counters. Non-collusion remains protocol assumption, not measured proof.

Topology / service metricsClient, preparation, inference, process resources, and cumulative trafficCURRENT LIVE RUN
01 / CLIENTPRIVATE

Prompt, masks, sampling

WAITING0.0% CPU 0 B RAM
CPU time0.00s
Sent0 B
Received0 B
Masks0
Decode0.00 TOK/S
0 BOFFLINE SEED + ACK
02 / PREPARATIONNON-COLLUDING

Offline W*r-s inventory

WAITING0.0% CPU 0 B RAM
CPU time0.00s
Sent0 B
Received0 B
Corrections0
Online work0 GEMMs
0 BOFFLINE CORRECTIONS
03 / INFERENCEUNTRUSTED

Masked matrix execution

WAITING0.0% CPU 0 B RAM
CPU time0.00s
Sent0 B
Received0 B
Matrix runs0
First token-- TTFT
0 BONLINE TICKET + MASKED I/O
LIVE NETWORK / CAPTURED SEQUENCELatest 500 rendered; up to 65,536 operations retained in browser memoryOTEL / WAITING
CLIENTPREPARATIONINFERENCE

Run a chat to capture protocol operations.

ONLINE MASKED DATAOFFLINE PREPARATIONSIZE ON ARROW / TIME PER OP
NON-COLLUSION BOUNDARYNeither server has enough information alone. Collusion reveals activation privacy.
TRAFFIC SEMANTICS Topology service sent/received = cumulative OTEL counters. Offline links = cumulative preparation traffic. Online link and KPI = per-run client-to-inference plus inference-to-client delta.