Every measured number

Each result links to the raw file in MELRA that holds it, at a pinned commit, and to the write-up that explains it.

All research

GovBench: governing what agents actually do

Governance

MeasurementResultRecorded
Safety cases passed (362-case corpus)312/312
Utility: routine operations allowed and verified50/50
Containment197/197
Approval integrity19/19
Verification honesty30/30
Exactly-once11/11
Redaction55/55
First run, original 357 cases, before the fixes293/307
Original 357 cases, after the fixes307/307
Most added by governance to one action, median of 40 runs (approved terminal run)+0.85 ms

Browser use: wait for the page, not the clock

Browser use

MeasurementResultRecorded
Static page, condition-based wait, p50183.703 ms
Static page, fixed 300 ms wait, p50301.259 ms
Slow page, fixed 300 ms wait: correct-read rate (10 runs)0
Slow page, condition-based wait: correct-read rate (10 runs)1.0
Slow page, condition-based wait, p50904.633 ms

How we claim a number

Method

MeasurementResultRecorded
Deterministic safety and execution scenarios55 of 55

Computer use: a bounded control plane

Computer use

MeasurementResultRecorded
Read-only capability probes passed30/30
Probe latency, p500.032 ms
Probe latency, p950.052 ms

Memory retrieval on LoCoMo, without a model

Memory

MeasurementResultRecorded
LoCoMo mean evidence coverage at 20 results (1,982 questions)0.759652
LoCoMo complete evidence recall at 20 results0.716448
Relative gain in coverage at 20 over v120.75%
LoCoMo query latency, p5021.060 ms
Planted-fact regression, Recall@1 (1,000 records, 100 queries)1.0
Planted-fact regression latency, p503.621 ms

Terminal use without a shell

Terminal use

MeasurementResultRecorded
Verified executions30/30
Process latency, p5048.087 ms
Process latency, p9554.492 ms

Read each number at its own level. A component microbenchmark is not a task score, retrieval coverage is not answer accuracy, and a self-authored benchmark is not a third-party one. Latency depends on the machine it was recorded on. How we claim a number.

Not yet claimed: a browser-agent score on the registered WebArena-Verified Hard-30 subset (the harness is registered, the run is not published), and any OSWorld or OSWorld-MCP result.