Performance / measured behavior

The whole
working session.

Measurements with their methods, observations and limits. Each report stays attached to the application that was measured. Same-host comparisons show the cost of a complete tool stack and web-app responsiveness. Fixed scales, declared targets and every observation keep the numbers in context.

NUS / local observationsComparisons + homepage evidence

One environment. Six complete stacks.

Installed application bytes · lower is smaller. NUS compared with VS Code + a browser + a terminal, all on this Mac.

Reported statistic┆ NUS bundle: 391.24 MiB
NUSLocal optimized build
391.24MiB
VS Code + Zen + GhosttyThree application bundles
1,514.22MiB
VS Code + Zen + KittyThree application bundles
1,706.38MiB
VS Code + Zen + WezTermThree application bundles
1,720.57MiB
VS Code + Arc + GhosttyThree application bundles
1,792.84MiB
VS Code + Arc + KittyThree application bundles
1,985MiB
VS Code + Arc + WezTermThree application bundles
1,999.19MiB
05121,0241,5362,0482,560

Shared 0–2.5 GiB scale for installed application file bytes.

Regular-file bytes, each inode counted once; symlinks and user profiles excluded. Packages are measured as distributed, including universal binaries where supplied. This is storage, not RAM. The homepage’s macOS package figure uses allocated APFS storage for published preview.7; this comparison uses logical bytes of the local runtime build.

Values, scope and evidence

NUS

One application bundle. Runtime build identity is recorded in the source data.

Download source observations

Deterministic size measurement; no sampling interval.

VS Code + Zen + Ghostty

VS Code 1.138.0; Zen 1.22.2b; Ghostty 1.2.3

Download source observations

Deterministic size measurement; no sampling interval.

VS Code + Zen + Kitty

VS Code 1.138.0; Zen 1.22.2b; Kitty 0.49.0

Download source observations

Deterministic size measurement; no sampling interval.

VS Code + Zen + WezTerm

VS Code 1.138.0; Zen 1.22.2b; WezTerm 20240203-110809-5046fc22

Download source observations

Deterministic size measurement; no sampling interval.

VS Code + Arc + Ghostty

VS Code 1.138.0; Arc 1.165.1; Ghostty 1.2.3

Download source observations

Deterministic size measurement; no sampling interval.

VS Code + Arc + Kitty

VS Code 1.138.0; Arc 1.165.1; Kitty 0.49.0

Download source observations

Deterministic size measurement; no sampling interval.

VS Code + Arc + WezTerm

VS Code 1.138.0; Arc 1.165.1; WezTerm 20240203-110809-5046fc22

Download source observations

Deterministic size measurement; no sampling interval.

Earlier workspace baseline: three surfaces open.

Process-tree RSS · lower is smaller. A 100 KiB file, a local web page and an idle terminal command; editor foreground.

Reported statistic● Individual observations▨ Follow-up · retained Arc session┆ NUS median: 611.69 MiB
NUS (earlier build)Median · n=5
611.69MiB
VS Code + Zen + GhosttyMedian · n=5
4,483.3MiB
VS Code + Zen + KittyMedian · n=5
4,460.5MiB
VS Code + Zen + WezTermMedian · n=5
4,416.7MiB
VS Code + Arc + GhosttyMedian · 5 trials / Arc session retained
3,598.7MiB
VS Code + Arc + KittyMedian · 5 trials / Arc session retained
3,543.97MiB
VS Code + Arc + WezTermMedian · 5 trials / Arc session retained
3,552.52MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Earlier complete workspace baseline retained. Current NUS workspace rerun was stopped at user request; all displayed workspace figures belong to the earlier build. Recorded 2026-09-23. NUS and Zen stacks: five fresh-profile trials with rotating run order. Each dot summarizes five RSS snapshots. Summed RSS double-counts shared pages; it is not physical footprint. NUS has one window; the stack has three. Hatched Arc rows are a later batch: fresh VS Code/terminal launches share one retained Arc test-account session after the browser benchmark. Profile and prior workload differ; no matched causal speedup or independent Arc-memory confidence claim.

Values, scope and evidence

NUS (earlier build)

Earlier complete workspace baseline retained. Current NUS workspace rerun was stopped at user request; all displayed workspace figures belong to the earlier build. Median of five process-tree RSS snapshots, one second apart, after all three readiness oracles and 3 seconds settling Sum of the unique descendant PID union, RSS still double-counts shared pages across processes. Same tasks, NUS one window versus three app windows. VS Code includes a small readiness-only development extension; helper Python process present in every configuration. Fresh profiles/configs; OS caches and background activity uncontrolled. Observed range 611.27–613.05. Five independent trials are too few for a finite distribution-free 95% median interval.

Download source observations

Median · n=5 · MiB
ObservationValue
1611.6875
2611.328125
3613.046875
4612.03125
5611.265625

VS Code + Zen + Ghostty

Earlier complete workspace baseline retained. Current NUS workspace rerun was stopped at user request; all displayed workspace figures belong to the earlier build. Median of five process-tree RSS snapshots, one second apart, after all three readiness oracles and 3 seconds settling Sum of the unique descendant PID union, RSS still double-counts shared pages across processes. Same tasks, NUS one window versus three app windows. VS Code includes a small readiness-only development extension; helper Python process present in every configuration. Fresh profiles/configs; OS caches and background activity uncontrolled. Observed range 4,412.27–4,537.11. Five independent trials are too few for a finite distribution-free 95% median interval.

Download source observations

Median · n=5 · MiB
ObservationValue
14456.1875
24524.59375
34412.265625
44483.296875
54537.109375

VS Code + Zen + Kitty

Earlier complete workspace baseline retained. Current NUS workspace rerun was stopped at user request; all displayed workspace figures belong to the earlier build. Median of five process-tree RSS snapshots, one second apart, after all three readiness oracles and 3 seconds settling Sum of the unique descendant PID union, RSS still double-counts shared pages across processes. Same tasks, NUS one window versus three app windows. VS Code includes a small readiness-only development extension; helper Python process present in every configuration. Fresh profiles/configs; OS caches and background activity uncontrolled. Observed range 4,451.94–4,486.3. Five independent trials are too few for a finite distribution-free 95% median interval.

Download source observations

Median · n=5 · MiB
ObservationValue
14486.296875
24486
34456.375
44460.5
54451.9375

VS Code + Zen + WezTerm

Earlier complete workspace baseline retained. Current NUS workspace rerun was stopped at user request; all displayed workspace figures belong to the earlier build. Median of five process-tree RSS snapshots, one second apart, after all three readiness oracles and 3 seconds settling Sum of the unique descendant PID union, RSS still double-counts shared pages across processes. Same tasks, NUS one window versus three app windows. VS Code includes a small readiness-only development extension; helper Python process present in every configuration. Fresh profiles/configs; OS caches and background activity uncontrolled. Observed range 4,359.23–4,467.66. Five independent trials are too few for a finite distribution-free 95% median interval.

Download source observations

Median · n=5 · MiB
ObservationValue
14379.9375
24467.65625
34416.703125
44432.609375
54359.234375

VS Code + Arc + Ghostty

Median of five process-tree RSS snapshots, one second apart, after all three readiness oracles and 3 seconds settling Arc uses the user-prepared empty account in the existing macOS session; process, browser storage and caches retained, no app restarts. Account change is not filesystem profile isolation. VS Code and terminals use fresh profiles/configs. Follow-up batch, not paired with the earlier NUS/Zen runs. RSS sums unique descendant PIDs, double-counts shared pages; not physical footprint. Editor foreground, three windows. Small readiness extension and common Python helper included. Background activity uncontrolled. Observed range 3,514.98–3,860.94. Shared Arc session; no independence assumption or confidence interval.

Download source observations

Median · 5 trials / Arc session retained · MiB
ObservationValue
13860.9375
23579.984375
33609.046875
43598.703125
53514.984375

VS Code + Arc + Kitty

Median of five process-tree RSS snapshots, one second apart, after all three readiness oracles and 3 seconds settling Arc uses the user-prepared empty account in the existing macOS session; process, browser storage and caches retained, no app restarts. Account change is not filesystem profile isolation. VS Code and terminals use fresh profiles/configs. Follow-up batch, not paired with the earlier NUS/Zen runs. RSS sums unique descendant PIDs, double-counts shared pages; not physical footprint. Editor foreground, three windows. Small readiness extension and common Python helper included. Background activity uncontrolled. Observed range 3,521.11–3,561.33. Shared Arc session; no independence assumption or confidence interval.

Download source observations

Median · 5 trials / Arc session retained · MiB
ObservationValue
13524.609375
23521.109375
33543.96875
43550.140625
53561.328125

VS Code + Arc + WezTerm

Median of five process-tree RSS snapshots, one second apart, after all three readiness oracles and 3 seconds settling Arc uses the user-prepared empty account in the existing macOS session; process, browser storage and caches retained, no app restarts. Account change is not filesystem profile isolation. VS Code and terminals use fresh profiles/configs. Follow-up batch, not paired with the earlier NUS/Zen runs. RSS sums unique descendant PIDs, double-counts shared pages; not physical footprint. Editor foreground, three windows. Small readiness extension and common Python helper included. Background activity uncontrolled. Observed range 3,496.63–3,572.5. Shared Arc session; no independence assumption or confidence interval.

Download source observations

Median · 5 trials / Arc session retained · MiB
ObservationValue
13552.515625
23563.4375
33572.5
43530.109375
53496.625

Web applications, on the same workload.

Speedometer 3.1 · higher is better. NUS and Zen: five process launches each. Arc: five repetitions in one test-account session.

Reported statistic● Individual observations▨ Follow-up · retained Arc session┆ NUS median: 44.34 score
NUSMedian · n=5
44.34score
ZenMedian · n=5
40.12score
ArcMedian · 5 repetitions / one session
36.08score
0255075100

Fixed 0–100 display range; higher is better. Scores are not percentages and can exceed 100.

NUS/Zen: fresh profiles, alternating application order. All use unchanged pinned Speedometer workloads served locally. Ten internal iterations produce each dot; they are not ten independent app launches. This measures web-app responsiveness, not human productivity. Hatched Arc row: later follow-up batch with retained profile/cache state in the existing macOS account. Repetitions share one process; no confidence interval or claim of a profile-matched speedup.

Values, scope and evidence

NUS

Speedometer 3.1 default suites; ten internal iterations per independent run. Fresh profiles and unchanged workloads. Document visibility checked at readiness and completion; continuous focus/occlusion was not instrumented. Zen trial 4 stalled and appeared as a macOS thumbnail; it completed before timeout after read-only UI inspection, which may have changed focus. Its score is retained. Background activity uncontrolled; no tightly controlled causal competitor speedup claim. Observed range 43.36–45.14. Five independent trials are too few for a finite distribution-free 95% median interval.

Download source observations

Median · n=5 · score
ObservationValue
144.341557010989064
243.93417010189581
343.35802154661351
445.13739982787466
544.34290308614549

Zen

Speedometer 3.1 default suites; ten internal iterations per independent run. Fresh profiles and unchanged workloads. Document visibility checked at readiness and completion; continuous focus/occlusion was not instrumented. Zen trial 4 stalled and appeared as a macOS thumbnail; it completed before timeout after read-only UI inspection, which may have changed focus. Its score is retained. Background activity uncontrolled; no tightly controlled causal competitor speedup claim. Observed range 33.57–40.93. Five independent trials are too few for a finite distribution-free 95% median interval.

Download source observations

Median · n=5 · score
ObservationValue
140.08335744338647
240.9250872362475
340.64409637671488
433.57356098064117
540.11976363199968

Arc

User-prepared empty Arc account, existing macOS account and app process. Account changes do not establish a new filesystem profile. Browser storage and caches retained; no app restart. Repetitions share session state and are not independent process trials. Follow-up batch, not paired with the earlier NUS/Zen runs; background activity, thermal and power state uncontrolled. No causal speedup or confidence interval claim. Observed range 34.76–36.57. Shared Arc session; no independence assumption or confidence interval.

Download source observations

Median · 5 repetitions / one session · score
ObservationValue
135.93426400337875
236.51041026214705
336.573223953214324
436.07586209257458
534.76157245191473

Browser workload: Speedometer source 1386415be8fe. NUS Chromium 152.0.7977.83 · Zen 1.22.2b. Arc 1.165.1 · Native viewport and user-agent values are retained in each run.

The package figures on the homepage.

Published v0.0.1-preview.7, authenticated archives. The comparison above uses the local runtime build; these are the shipping package measurements.

Download and installed size are distinct measurements
PlatformDownloadOn disk / unpacked
macOS ARM64154.08 MiB379.3 MiB
Windows x64190.75 MiB458.27 MiB
Linux x64358.34 MiB1,526.42 MiB (1.49 GiB)

macOS: allocated APFS storage after extraction. Windows/Linux: archive file payload bytes, not native filesystem allocation. Archive hashes and exact sizes ↓

The file figures on the homepage.

Request to loaded-content submission · lower is better. Both file sizes share the same 0–200 ms axis.

Reported statistic● Individual observations┆ 10 MiB target only: 100 ms
10 MiB filep95 · 20 independent opens
16.1ms
100 MiB fileMaximum · 5 independent opens
65.49ms
050100150200

Shared file-open scale; 100 ms is the existing NUS 10 MiB target.

The homepage keeps its original statistics: p95 for 20 opens of 10 MiB, maximum for five opens of 100 MiB. One excluded warmup per experiment. The 100 ms target applies only to the 10 MiB workload. These are NUS measurements against its target, not an editor comparison or a photon-latency measurement.

Values, scope and evidence

10 MiB file

File open request to first loaded-content submission; independent launches, one excluded warmup; OS caches uncontrolled; not display scanout.

Download source observations

p95 · 20 independent opens · ms
ObservationValue
114.854792
215.423916
315.876999999999999
415.367917
514.663458
615.700124999999998
715.843124999999999
814.873958
914.525583
1015.723458999999998
1116.103959
1214.631165999999999
1314.797459
1414.869959
1515.288292
1614.221875
1715.417
1816.638333000000003
1914.795041999999999
2014.633999999999999

100 MiB file

File open request to first loaded-content submission; independent launches, one excluded warmup; OS caches uncontrolled; not display scanout.

Download source observations

Maximum · 5 independent opens · ms
ObservationValue
164.999
264.76404199999999
364.55291700000001
465.491209
564.522458

What an additional browser tab costs.

Additional process-tree RSS, tabs 2–8; first-tab initialization excluded. Lower is smaller.

Reported statistic● Individual observations┆ Existing NUS budget: 143.05 MiB
NUS additional tabMedian · 35 increments / five profiles
107.17MiB
064128192256

Fixed tab-increment scale; the 150 MB target is 143.05 MiB.

Observed range 96.39–111.56 MiB. Homepage reports the median and range; all increments remain inspectable. Five fresh profiles, tabs 2–8 after first-tab initialization. No confidence interval treats these dependent increments as independent runs. This target is a NUS resource budget, not a competitor result.

Values, scope and evidence

NUS additional tab

Additional process-tree RSS, tabs 2–8; first-tab initialization excluded. Range 96.39–111.56 MiB. Increments within a profile are dependent.

Download source observations

Median · 35 increments / five profiles · MiB
ObservationValue
1:tabs-2109.078125
1:tabs-3107.421875
1:tabs-4107.078125
1:tabs-5108.015625
1:tabs-6106.84375
1:tabs-7107.234375
1:tabs-896.390625
2:tabs-2110.125
2:tabs-396.453125
2:tabs-4106.859375
2:tabs-5106.984375
2:tabs-6107.171875
2:tabs-7108.78125
2:tabs-8106.640625
3:tabs-2109.828125
3:tabs-396.625
3:tabs-4107.96875
3:tabs-5106.953125
3:tabs-6105.90625
3:tabs-7107.109375
3:tabs-8107.265625
4:tabs-2110.859375
4:tabs-397.1875
4:tabs-4111.5625
4:tabs-5107.1875
4:tabs-6107.875
4:tabs-7106.59375
4:tabs-8106.9375
5:tabs-2105.390625
5:tabs-3100.859375
5:tabs-4107.40625
5:tabs-5107.984375
5:tabs-6106.828125
5:tabs-7107.265625
5:tabs-8107.1875

What an additional empty window costs.

Observed parent-process RSS change, windows 2–4; not total GPU memory. Negative deltas can reflect concurrent page reclamation, not negative window allocation. Lower is smaller.

Reported statistic● Individual observations┆ Existing NUS budget: 23.84 MiB
NUS additional windowMedian · 15 increments / five profiles
2.08MiB
08162432

Fixed window-increment scale; the 25 MB target is 23.84 MiB.

Outside the fixed plot domain: NUS additional window, observation 4:windows-3: -19.98 MiB. Edge arrows preserve these observations; exact values remain below.

Observed range -19.98–29.58 MiB. Homepage reports the range; all increments remain inspectable. Five fresh profiles, windows 2–4. No confidence interval treats these dependent increments as independent runs. This target is a NUS resource budget, not a competitor result.

Values, scope and evidence

NUS additional window

Observed parent-process RSS change, windows 2–4; not total GPU memory. Negative deltas can reflect concurrent page reclamation, not negative window allocation. Range -19.98–29.58 MiB. Increments within a profile are dependent.

Download source observations

Median · 15 increments / five profiles · MiB
ObservationValue
1:windows-23.859375
1:windows-32.015625
1:windows-42.109375
2:windows-23.203125
2:windows-31.921875
2:windows-42.09375
3:windows-23.21875
3:windows-31.515625
3:windows-41.921875
4:windows-229.578125
4:windows-3-19.984375
4:windows-41.71875
5:windows-22.71875
5:windows-32.078125
5:windows-41.734375
Machine, versions and reproducibility
Hardware
Apple M4 Pro · 48 GiB RAM · macOS-15.6-arm64-arm-64bit-Mach-O. One host; OS caches and background activity uncontrolled.
Runtime measured
2026-09-23 · local optimized build, including uncommitted source changes. Binary SHA-256: ef3c08021ff3b29bc7168b05a28bd0c14432772d18fca94301868ee6d54f1422
Applications
NUS 0.0.1 · VS Code 1.138.0 · Arc 1.165.1 · Zen 1.22.2b · Ghostty 1.2.3 · Kitty 0.49.0 · WezTerm 20240203-110809-5046fc22
Published packages
v0.0.1-preview.7. Download bytes, archive hashes, unpacked bytes and macOS allocation are recorded separately in homepage evidence.
Reproduce
Comparison protocol and commands. Download current collector sources / Exact baseline collectors (requires a matching NUS build and external apps). Raw JSON retains every successful observation and the exact harness identities. Failed calibration evidence: browser v1 / workspace v1.
Native workflow timings and startup evidence

Earlier NUS-only pilots. Different workflows share a 0–5 second axis; these timings are not competitor comparisons.

NUS / startup

2026-09-22 · complete

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1 (4740af3)

10/10 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Main entry to first presentation call

Median · lower is better

Reported statistic● Individual observations
NUSMedian
276.77ms
05001,0001,5002,000

Fixed 0–2 second startup scale; not a pass/fail budget.

Every observation is retained. Range 263.38–1,616.91 ms · n=10 · missing=0. Band: median interval 270.95–290.24, at least 95% coverage under IID assumptions. Rust main entry → application present-call return; not display scanout

Values, scope and evidence

NUS

Rust main entry → application present-call return; not display scanout

Download source observations

Median · ms
ObservationValue
11616.914708
2276.21012499999995
3280.328959
4270.945916
5290.24025
6273.343708
7277.334708
8263.379959
9271.58408299999996
10284.26291599999996

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
389.58MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 384.56–394.03 MiB · n=10 · missing=0. Band: median interval 386.3–391.42, at least 95% coverage under IID assumptions. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1389.78125
2388.609375
3394.03125
4391.03125
5389.375
6386.296875
7391.09375
8391.421875
9387.5
10384.5625
Method & evidence
Execution
Application-owned scenario; not human task timing
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
8a115cd9ae5dbaaf340207968943e31fd39874a42317407ca56eaa12b7109586
Source-report SHA-256
be7d6746858c1b818069ac5871b92af9acfc196768889c59005259a435111478
Public observations
Sanitized run records; source logs and paths are excluded.
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / edit-verify

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1 (4740af3)

10/10 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Edit, save and verify in browser

Median · lower is better

Reported statistic● Individual observations
NUSMedian
220.3ms
01,0002,0003,0004,0005,000

Shared 0–5 second scale for all scripted workflow durations.

Every observation is retained. Range 199.93–262.36 ms · n=10 · missing=0. Band: median interval 214.63–233.19, at least 95% coverage under IID assumptions. Prepared server → page open, file edit and save → browser DOM confirms saved bytes

Values, scope and evidence

NUS

Prepared server → page open, file edit and save → browser DOM confirms saved bytes

Download source observations

Median · ms
ObservationValue
1214.634084
2233.187084
3222.918583
4216.35479099999998
5217.270875
6199.928166
7262.355375
8224.896375
9217.67666699999998
10224.688458

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
728.5MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 724.88–733.28 MiB · n=10 · missing=0. Band: median interval 725.52–732.08, at least 95% coverage under IID assumptions. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1726.359375
2730.5625
3725.515625
4732.078125
5729.234375
6728.609375
7724.875
8727.15625
9733.28125
10728.390625
Method & evidence
Execution
Local scripted pilot; application-owned actions include event-loop scheduling and oracle overhead. Not human task timing or input-to-photon latency.
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
0b3c7fac62c9a3cafb94da2cace76c76f5e9b71dc69ebdf779ec4b7d8bb2d1a6
Source-report SHA-256
ee9a6bde64c31f1235b9d5abb82c44b0b6c7948d9a9cce0553dce3fc74e2a176
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 7ccc69d84e4f9a9a7cc7efb416925b27115ece2c6a374fefc1722871e29e8cfb
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / port-conflict

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1 (4740af3)

10/10 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Recover an occupied local port

Median · lower is better

Reported statistic● Individual observations
NUSMedian
2,149.52ms
01,0002,0003,0004,0005,000

Shared 0–5 second scale for all scripted workflow durations.

Every observation is retained. Range 2,117.1–2,260.3 ms · n=10 · missing=0. Band: median interval 2,132.64–2,182.77, at least 95% coverage under IID assumptions. Contender launch → real conflict, owner jump, stop and restart → both services confirmed in browser

Values, scope and evidence

NUS

Contender launch → real conflict, owner jump, stop and restart → both services confirmed in browser

Download source observations

Median · ms
ObservationValue
12138.364583
22156.822458
32117.103292
42148.872792
52182.767541
62141.6165419999998
72132.635792
82164.921166
92260.301458
102150.160375

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
761.72MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 759.55–767.17 MiB · n=10 · missing=0. Band: median interval 760.14–764.14, at least 95% coverage under IID assumptions. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1759.546875
2760.140625
3760.984375
4760.453125
5760.3125
6764.140625
7762.453125
8767.171875
9763.890625
10763.34375
Method & evidence
Execution
Local scripted pilot; application-owned actions include event-loop scheduling and oracle overhead. Not human task timing or input-to-photon latency.
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
0b3c7fac62c9a3cafb94da2cace76c76f5e9b71dc69ebdf779ec4b7d8bb2d1a6
Source-report SHA-256
3cc7d43fc5f819d8e75291e85308c438b33e1a0cd37caf9d98c8e6eaf0a81149
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 7ccc69d84e4f9a9a7cc7efb416925b27115ece2c6a374fefc1722871e29e8cfb
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / project-switch

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1 (4740af3)

10/10 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Switch project buffers with PTY output

Median · lower is better

Reported statistic● Individual observations
NUSMedian
200.84ms
01,0002,0003,0004,0005,000

Shared 0–5 second scale for all scripted workflow durations.

Every observation is retained. Range 192.93–208.33 ms · n=10 · missing=0. Band: median interval 192.99–201.47, at least 95% coverage under IID assumptions. Dirty project A → edit/save project B → return/save A → browser confirms A; output counter advances during interval

Values, scope and evidence

NUS

Dirty project A → edit/save project B → return/save A → browser confirms A; output counter advances during interval

Download source observations

Median · ms
ObservationValue
1201.376417
2192.932833
3201.47458300000002
4208.330333
5201.390416
6201.463291
7200.296542
8192.985542
9198.36812500000002
10195.257917

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
630.53MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 628.48–634.47 MiB · n=10 · missing=0. Band: median interval 628.75–632.27, at least 95% coverage under IID assumptions. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1632.046875
2630.6875
3629.328125
4631.890625
5628.75
6629.578125
7632.265625
8634.46875
9628.484375
10630.375
Method & evidence
Execution
Local scripted pilot; application-owned actions include event-loop scheduling and oracle overhead. Not human task timing or input-to-photon latency.
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
0b3c7fac62c9a3cafb94da2cace76c76f5e9b71dc69ebdf779ec4b7d8bb2d1a6
Source-report SHA-256
6e878972ce450fceeaa54d3c0a1a3bae9f4cea03f33d133747d6c4c524348ec2
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 7ccc69d84e4f9a9a7cc7efb416925b27115ece2c6a374fefc1722871e29e8cfb
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / startup

2026-09-23 · complete

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1-dev (4740af3)

10/10 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Main entry to first presentation call

Median · lower is better

Reported statistic● Individual observations
NUSMedian
270.98ms
05001,0001,5002,000

Fixed 0–2 second startup scale; not a pass/fail budget.

Every observation is retained. Range 255.36–280.35 ms · n=10 · missing=0. Band: median interval 258.27–277.46, at least 95% coverage under IID assumptions. Rust main entry → application present-call return; not display scanout

Values, scope and evidence

NUS

Rust main entry → application present-call return; not display scanout

Download source observations

Median · ms
ObservationValue
1277.45599999999996
2268.16341700000004
3273.797125
4275.173417
5280.346
6266.846
7258.26979100000005
8255.364667
9274.08441700000003
10263.41845800000004

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
124.36MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 122.05–125.75 MiB · n=10 · missing=0. Band: median interval 124.02–125.52, at least 95% coverage under IID assumptions. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1125.265625
2124.0625
3124.015625
4122.046875
5124.625
6125.75
7124.390625
8125.515625
9124.125
10124.328125
Method & evidence
Execution
Application-owned scenario; not human task timing
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
ef3c08021ff3b29bc7168b05a28bd0c14432772d18fca94301868ee6d54f1422
Source-report SHA-256
a5c1cb77ce96f57dac93b62b4b9ee8096934cfdf4579a2c1b34c711d6abe7603
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 9587973728ccb8976b805ad1e0e567caea544fe29b22a485bffcb9a07a8116bd
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / editor-open

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1-dev (4740af3)

20/20 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Open 10 MiB text file

Median · lower is better

Reported statistic● Individual observations
NUSMedian
15.08ms
050100150200

Shared file-open scale; 100 ms is the existing NUS 10 MiB target.

Every observation is retained. Range 14.22–16.64 ms · n=20 · missing=0. Band: median interval 14.8–15.7, at least 95% coverage under IID assumptions. File open request → loaded content frame submission; not display scanout

Values, scope and evidence

NUS

File open request → loaded content frame submission; not display scanout

Download source observations

Median · ms
ObservationValue
114.854792
215.423916
315.876999999999999
415.367917
514.663458
615.700124999999998
715.843124999999999
814.873958
914.525583
1015.723458999999998
1116.103959
1214.631165999999999
1314.797459
1414.869959
1515.288292
1614.221875
1715.417
1816.638333000000003
1914.795041999999999
2014.633999999999999

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
139.9MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 137.39–142.56 MiB · n=20 · missing=0. Band: median interval 138.97–140.23, at least 95% coverage under IID assumptions. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1140.875
2140.421875
3138.6875
4140.234375
5142.5625
6141.234375
7140.1875
8138.296875
9139.328125
10139.203125
11138.8125
12140.03125
13140.203125
14137.390625
15138.96875
16139.5625
17138.015625
18141.28125
19140.21875
20139.765625
Method & evidence
Execution
Application-owned scenario; not human task timing
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
ef3c08021ff3b29bc7168b05a28bd0c14432772d18fca94301868ee6d54f1422
Source-report SHA-256
da4ac061b939f6198fffb7ab4d9cb31ae0576ffbca1836adae7bdd9c48567fb0
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 9587973728ccb8976b805ad1e0e567caea544fe29b22a485bffcb9a07a8116bd
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / editor-open

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1-dev (4740af3)

5/5 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Open 100 MiB text file

Median · lower is better

Reported statistic● Individual observations
NUSMedian
64.76ms
050100150200

Shared file-open scale; 100 ms is the existing NUS 10 MiB target.

Every observation is retained. Range 64.52–65.49 ms · n=5 · missing=0. Median interval unavailable for this sample size. File open request → loaded content frame submission; not display scanout

Values, scope and evidence

NUS

File open request → loaded content frame submission; not display scanout

Download source observations

Median · ms
ObservationValue
164.999
264.76404199999999
364.55291700000001
465.491209
564.522458

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
241.13MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 238.67–242.16 MiB · n=5 · missing=0. Median interval unavailable for this sample size. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1242.15625
2240.546875
3238.671875
4241.859375
5241.125
Method & evidence
Execution
Application-owned scenario; not human task timing
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
ef3c08021ff3b29bc7168b05a28bd0c14432772d18fca94301868ee6d54f1422
Source-report SHA-256
be773c210eec05a3418b400f8f238d813d7d0776f84feed94d7fbe2c643adf8d
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 9587973728ccb8976b805ad1e0e567caea544fe29b22a485bffcb9a07a8116bd
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / edit-verify

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1-dev (4740af3)

5/5 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Edit, save and verify in browser

Median · lower is better

Reported statistic● Individual observations
NUSMedian
308.52ms
01,0002,0003,0004,0005,000

Shared 0–5 second scale for all scripted workflow durations.

Every observation is retained. Range 300.07–499.92 ms · n=5 · missing=0. Median interval unavailable for this sample size. Prepared server → page open, file edit and save → browser DOM confirms saved bytes

Values, scope and evidence

NUS

Prepared server → page open, file edit and save → browser DOM confirms saved bytes

Download source observations

Median · ms
ObservationValue
1375.85025
2499.92141699999996
3300.070917
4308.519792
5308.367208

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
741.83MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 738.67–750.91 MiB · n=5 · missing=0. Median interval unavailable for this sample size. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1750.90625
2750.40625
3741.828125
4740.90625
5738.671875
Method & evidence
Execution
Local scripted pilot; application-owned actions include event-loop scheduling and oracle overhead. Not human task timing or input-to-photon latency.
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
ef3c08021ff3b29bc7168b05a28bd0c14432772d18fca94301868ee6d54f1422
Source-report SHA-256
4f0324aa555d188928ad29aa60a97967815aac010b11e1dbc8ed61d024861a5b
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 9587973728ccb8976b805ad1e0e567caea544fe29b22a485bffcb9a07a8116bd
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / port-conflict

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1-dev (4740af3)

5/5 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Recover an occupied local port

Median · lower is better

Reported statistic● Individual observations
NUSMedian
2,084.3ms
01,0002,0003,0004,0005,000

Shared 0–5 second scale for all scripted workflow durations.

Every observation is retained. Range 1,688.32–2,182.12 ms · n=5 · missing=0. Median interval unavailable for this sample size. Contender launch → real conflict, owner jump, stop and restart → both services confirmed in browser

Values, scope and evidence

NUS

Contender launch → real conflict, owner jump, stop and restart → both services confirmed in browser

Download source observations

Median · ms
ObservationValue
11688.318708
22106.011416
32062.467459
42182.12175
52084.297833

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
781.22MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 756.17–1,191.13 MiB · n=5 · missing=0. Median interval unavailable for this sample size. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1781.21875
2756.171875
31190.484375
4777.21875
51191.125
Method & evidence
Execution
Local scripted pilot; application-owned actions include event-loop scheduling and oracle overhead. Not human task timing or input-to-photon latency.
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
ef3c08021ff3b29bc7168b05a28bd0c14432772d18fca94301868ee6d54f1422
Source-report SHA-256
7d37032143c0d63179738bea915d4b97c2edd7a9bc8c834fc49ed99b49462315
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 9587973728ccb8976b805ad1e0e567caea544fe29b22a485bffcb9a07a8116bd
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

NUS / project-switch

2026-09-23 · complete

Scripted pilot · uncontrolled background activity

macOS-15.6-arm64-arm-64bit-Mach-O · nus 0.0.1-dev (4740af3)

5/5 measured runs · 1 warmup · 0 failed attempts · 0 unattempted

Switch project buffers with PTY output

Median · lower is better

Reported statistic● Individual observations
NUSMedian
295.12ms
01,0002,0003,0004,0005,000

Shared 0–5 second scale for all scripted workflow durations.

Every observation is retained. Range 262.25–348.08 ms · n=5 · missing=0. Median interval unavailable for this sample size. Dirty project A → edit/save project B → return/save A → browser confirms A; output counter advances during interval

Values, scope and evidence

NUS

Dirty project A → edit/save project B → return/save A → browser confirms A; output counter advances during interval

Download source observations

Median · ms
ObservationValue
1322.28562500000004
2295.120625
3262.247
4281.244791
5348.0845

Whole process tree RSS

Median · lower is better

Reported statistic● Individual observations
NUSMedian
638.67MiB
02,0484,0966,1448,192

Shared 0–8 GiB scale for process-tree RSS; not a resource budget.

Every observation is retained. Range 634.22–649.27 MiB · n=5 · missing=0. Median interval unavailable for this sample size. Point-in-time summed resident pages, including shared-page double counting; not private footprint

Values, scope and evidence

NUS

Point-in-time summed resident pages, including shared-page double counting; not private footprint

Download source observations

Median · MiB
ObservationValue
1649.265625
2638.671875
3635.34375
4634.21875
5639.171875
Method & evidence
Execution
Local scripted pilot; application-owned actions include event-loop scheduling and oracle overhead. Not human task timing or input-to-photon latency.
Cache state
Fresh onboarded profiles; OS/filesystem caches not evicted
Hardware limits
Hardware model, power and thermal state not recorded by this source
Binary SHA-256
ef3c08021ff3b29bc7168b05a28bd0c14432772d18fca94301868ee6d54f1422
Source-report SHA-256
0ef5084ba4e83d1f0d24b69ae739f491b0e4715b94f64b19364cc76ba1568d45
Public observations
Sanitized run records; source logs and paths are excluded.
Harness identity
benchmark_fixture.py: 3c7507010f64762a2f0c87ba65800ca682dc965f0e3ec1ccef654bba3c973c83
benchmark_workflows.py: 5106f07cb3eb9a9f0129f5f77fd69661a8fb65fe277e9915464143b007ef7614
perf-native.py: 9587973728ccb8976b805ad1e0e567caea544fe29b22a485bffcb9a07a8116bd
Statistics
Median and observed range across independent process summaries. Intervals use binomial order statistics for the population median, with at least 95% coverage assuming independent, identically distributed runs. They do not establish a tail-latency bound.

What remains unmeasured

Broader workflow evidence

Local pilots cover scripted actions. Cross-machine replication, sustained mixed workloads, and human task studies remain open.

Remaining comparisons

Profile-matched Arc replication remains open. Equivalent cross-app workflow timings, terminal throughput and editor-open comparisons remain open; storage and browser scores cannot stand in for them.

Physical latency

Submission counters do not measure keyboard-to-screen latency. External display measurements remain separate.

Correctness has its own evidence.

Formal verification applies to named invariants under explicit assumptions. A proof of retained output does not prove the whole application correct or establish its speed.

Verified: bounded output retention. 3 production methods verified; the deliberately broken input-retention mutation was rejected. The retained bytes equal the newest suffix up to the configured capacity.

Proof scope & assumptions
Source identity
7124d59e9963b47073440a3042adc63277bf2758d54e311fe41abb8bf457908d
Verifier
Verus 0.2026.09.20.aef82ed · 2026-09-23 · --no-cheating
Assumptions
Verus, SMT solver and Rust compiler correctness; Pinned vstd contracts for Vec and slice operations; Successful allocation; no claim about allocator failure or physical footprint; Private ring fields; state created by new and mutated only by push.
Excluded
PTY and socket delivery; Replay screen reconstruction; Whole-application correctness; Timing and performance.
Relationship to measurements
This is a source-level proof. It is not an attestation of the measured executable or of the earlier startup build.

Proof result · JSON / Verified source / Reproduction instructions

Benchmark and verification protocol ↓ / Run the native workflows ↓