Governed Agent Harness — benchmark report

Read this first. These are synthetic microbenchmarks of local harness operations, run on shared CI runners with no warm-up. They quantify runtime overhead in the recorded environment only. They do not measure developer productivity, code quality or statistical significance.

5 repetition(s) of harness benchmark run; values are the median across repetitions of each run's median (p95 likewise).

Latency per operation

Median and p95 latency per operation (log scale, milliseconds)
medianp95
0.001 ms0.01 ms0.1 ms1 ms10 ms100 msState transitionState transition — median: 0.00125 ms0.00125State transition — p95: 0.00128 ms0.00128Gate evaluationGate evaluation — median: 0.002 ms0.002Gate evaluation — p95: 0.00204 ms0.00204Canonical hash (digest)Canonical hash (digest) — median: 0.0112 ms0.0112Canonical hash (digest) — p95: 0.0124 ms0.0124Artifact put (deduplicated)Artifact put (deduplicated) — median: 0.0729 ms0.0729Artifact put (deduplicated) — p95: 0.118 ms0.118Event append (hash chain)Event append (hash chain) — median: 0.257 ms0.257Event append (hash chain) — p95: 0.519 ms0.519Process launch, directProcess launch, direct — median: 12.394 ms12.394Process launch, direct — p95: 13.416 ms13.416Process launch, governedProcess launch, governed — median: 13.122 ms13.122Process launch, governed — p95: 13.889 ms13.889
Latency per operation across repetitions
OperationIterations/runMedian msp95 msOps/sMin–max of mediansStd. dev.
State transition10000.001250.00128777,704.20.00124 – 0.001333.68e-05
Gate evaluation10000.0020.00204489,138.20.00197 – 0.002157.31e-05
Canonical hash (digest)10000.01120.012485,834.40.0107 – 0.01160.000321
Artifact put (deduplicated)10000.07290.11812,320.90.0726 – 0.0770.00187
Event append (hash chain)10000.2570.5193,195.40.248 – 0.2680.00803
Process launch, direct2012.39413.41680.50712.117 – 12.8090.27
Process launch, governed2013.12213.88975.94013.033 – 13.4450.199

Direct versus governed process launch

Process launch median: governed overhead 0.639 ms (5.16 %)
median
10 ms100 msProcess launch, directProcess launch, direct — median: 12.394 ms12.394Process launch, governedProcess launch, governed — median: 13.122 ms13.122
Governance overhead per repetition. Local process-launch overhead only; excludes validators, agents and human wait.
RepetitionAbsolute median overhead (ms)Relative median overhead (%)
10.6144.79
21.32810.96
30.6395.16
40.867.04
50.6244.99

Throughput

Operations per second (log scale; higher is faster)
ops/s
10 ops/s100 ops/s1,000 ops/s10,000 ops/s100,000 ops/s1,000,000 ops/sState transitionState transition — ops/s: 777,704.2 ops/s777,704.2Gate evaluationGate evaluation — ops/s: 489,138.2 ops/s489,138.2Canonical hash (digest)Canonical hash (digest) — ops/s: 85,834.4 ops/s85,834.4Artifact put (deduplicated)Artifact put (deduplicated) — ops/s: 12,320.9 ops/s12,320.9Event append (hash chain)Event append (hash chain) — ops/s: 3,195.4 ops/s3,195.4Process launch, directProcess launch, direct — ops/s: 80.507 ops/s80.507Process launch, governedProcess launch, governed — ops/s: 75.940 ops/s75.940

Dispersion across repetitions

Median of each repetition (one dot per repetition, log scale)
0.001 ms0.01 ms0.1 ms1 ms10 ms100 msState transitionState transition: 0.00133 msState transition: 0.00125 msState transition: 0.00127 msState transition: 0.00124 msState transition: 0.00125 msGate evaluationGate evaluation: 0.00215 msGate evaluation: 0.00197 msGate evaluation: 0.00202 msGate evaluation: 0.002 msGate evaluation: 0.00198 msCanonical hash (digest)Canonical hash (digest): 0.0116 msCanonical hash (digest): 0.0107 msCanonical hash (digest): 0.0112 msCanonical hash (digest): 0.011 msCanonical hash (digest): 0.0114 msArtifact put (deduplicated)Artifact put (deduplicated): 0.077 msArtifact put (deduplicated): 0.0726 msArtifact put (deduplicated): 0.0729 msArtifact put (deduplicated): 0.0726 msArtifact put (deduplicated): 0.0741 msEvent append (hash chain)Event append (hash chain): 0.255 msEvent append (hash chain): 0.248 msEvent append (hash chain): 0.268 msEvent append (hash chain): 0.266 msEvent append (hash chain): 0.257 msProcess launch, directProcess launch, direct: 12.809 msProcess launch, direct: 12.117 msProcess launch, direct: 12.394 msProcess launch, direct: 12.213 msProcess launch, direct: 12.499 msProcess launch, governedProcess launch, governed: 13.423 msProcess launch, governed: 13.445 msProcess launch, governed: 13.033 msProcess launch, governed: 13.072 msProcess launch, governed: 13.122 ms
Per-repetition medians
OperationRun 1 median msRun 2 median msRun 3 median msRun 4 median msRun 5 median ms
State transition0.001330.001250.001270.001240.00125
Gate evaluation0.002150.001970.002020.0020.00198
Canonical hash (digest)0.01160.01070.01120.0110.0114
Artifact put (deduplicated)0.0770.07260.07290.07260.0741
Event append (hash chain)0.2550.2480.2680.2660.257
Process launch, direct12.80912.11712.39412.21312.499
Process launch, governed13.42313.44513.03313.07213.122

Thesis reference point

Transcribed from the thesis, chapter 6 (implementation and demonstration), for the v0.8.0 cut: 1,000 iterations per operation, no warm-up, 20 samples for process launches. The raw evidence file is not part of this repository.

Reference values recorded in the thesis; environments differ, compare orders of magnitude only.
OperationThesis median ms (v0.8.0 cut)This report median ms
State transition0.000960.00125
Gate evaluation0.001530.002
Canonical hash (digest)0.02280.0112
Event append (hash chain)0.2810.257
Artifact put (deduplicated)0.05490.0729

Thesis governed-process overhead: 1.21 ms, 10.09 %, 11.23 %.

End-to-end scenarios

Direct patch-and-test path versus the complete governed path (programmatic approval, no human wait)
ProfileStatusIterationsDirect median msGoverned median msOverhead %Events (last run)Chain valid
nodePASSED3219.41,541.4602.7129True
pythonPASSED3325.34,175.51183.5140True
Scenario methodology
KeyValue
directPathApply the same deterministic patch and execute the mandatory test command.
governedPathInitialize, persist task, execute all normative phases, validate, review, evaluate the gate, apply a programmatic benchmark approval, close, and project status.
humanWaitExcluded: the benchmark approval is immediate and programmatic.
networkNo network access required.
scopeSynthetic local fixtures; not a productivity or quality claim.

Environment and methodology

Environment reported by the harness
KeyValue
platformLinux-7.0.0-1017-azure-x86_64-with-glibc2.39
processorx86_64
python3.12.16
Methodology reported by the harness
KeyValue
clocktime.perf_counter_ns
scopeMicrobenchmarks on this execution environment; not general performance claims.
warmupNo explicit warmup; every sample is reported in the aggregate.

Generated from runs between 2026-11-12T16:07:09.881327+00:00 and 2026-11-12T16:07:16.820834+00:00 (schemaVersion 1.0).