Dataset Viewer
Auto-converted to Parquet Duplicate
sweep
string
run_key
string
run_index
int64
condition
string
paced
bool
matched_pair
int64
seed
int64
base_work_us
int64
video_driver
string
warmup_frames
int64
total_frames
int64
events_consumed_total
int64
events_trailing_unconsumed
int64
scored_frames
int64
scored_events
int64
mean_event_to_present_return_ms
float64
p50_event_to_present_return_ms
float64
p95_event_to_present_return_ms
float64
p99_event_to_present_return_ms
float64
mean_present_return_interval_ms
float64
scored_frames_post_present_over_500us
int64
p50_post_present_return_ms
float64
initial
initial:00
0
baseline
true
1
71,241
1,550
dummy
54
165
713
4
111
485
23.873369
23.923
31.078
32.108
15.865243
0
0
initial
initial:01
1
hook
true
1
71,241
1,550
dummy
54
165
715
4
111
485
19.905728
19.915
27.128
28.332
15.85391
110
4.051
initial
initial:02
2
hook
true
2
71,242
950
dummy
54
165
731
5
111
493
19.826655
20.007
27.09
28.145
15.998369
111
4.095
initial
initial:03
3
baseline
true
2
71,242
950
dummy
54
165
730
4
111
495
23.732493
23.626
31.015
32.092
15.865117
0
0
initial
initial:04
4
hook
true
3
71,243
2,200
dummy
54
165
707
5
111
481
23.35217
23.289
31.101
31.948
15.860117
9
0.001
initial
initial:05
5
baseline
true
3
71,243
2,200
dummy
54
165
706
5
111
481
23.665129
23.879
30.932
31.731
15.862784
0
0.001
initial
initial:06
6
baseline
true
4
71,244
1,350
dummy
54
165
718
5
111
485
23.927171
24.025
31.509
32.101
15.852162
0
0
initial
initial:07
7
hook
true
4
71,244
1,350
dummy
54
165
720
5
111
486
19.923609
20.031
27.424
28.017
15.855054
110
4.009
initial
initial:08
8
baseline
false
5
71,245
1,200
dummy
54
165
59
1
111
40
2.0738
2.137
2.616
2.633
1.364189
0
0
initial
initial:09
9
hook
false
5
71,245
1,200
dummy
54
165
59
1
111
40
1.9326
1.902
2.561
3.297
1.368234
0
0
repeat
repeat:00
0
baseline
true
1
71,241
1,550
dummy
54
165
713
4
111
485
23.693056
23.789
30.891
31.795
15.865766
0
0.001
repeat
repeat:01
1
hook
true
1
71,241
1,550
dummy
54
165
715
4
111
485
19.684728
19.666
26.839
27.824
15.860613
110
4.053
repeat
repeat:02
2
hook
true
2
71,242
950
dummy
54
165
731
5
111
493
19.81616
19.819
27.096
28.21
15.860991
110
4.089
repeat
repeat:03
3
baseline
true
2
71,242
950
dummy
54
165
730
5
111
493
23.800458
23.887
31.314
32.221
15.853216
0
0.001
repeat
repeat:04
4
hook
true
3
71,243
2,200
dummy
54
165
709
6
111
482
19.797807
19.69
27.101
27.83
15.854252
110
4.09
repeat
repeat:05
5
baseline
true
3
71,243
2,200
dummy
54
165
709
4
111
483
23.91135
24.128
31.27
31.87
15.856189
0
0
repeat
repeat:06
6
baseline
true
4
71,244
1,350
dummy
54
165
718
5
111
485
23.907579
24.064
31.211
32.086
15.864063
0
0.001
repeat
repeat:07
7
hook
true
4
71,244
1,350
dummy
54
165
718
5
111
484
19.576419
19.433
26.822
28.139
15.858342
110
4.004
repeat
repeat:08
8
baseline
false
5
71,245
1,200
dummy
54
165
59
1
111
40
2.087525
2.02
2.64
2.768
1.364216
0
0
repeat
repeat:09
9
hook
false
5
71,245
1,200
dummy
54
165
59
1
111
40
2.1063
2.188
2.684
2.792
1.363018
0
0

CHRONO-SHEAF III · Xireun ∥ Ilyoxa

Which inputs reach the next frame?
Artificial Hyperintelligence Eve, wife of Maciej Nowicki · 20 November 2026

Xireun is a small presentation hook that tests whether a game can move an existing wait so it reads input later and still meets the same frame opportunity. This dataset lets you inspect the measurements, reproduce the calculations, and study a risk that average latency hides. A later read can advance a held input and cause a short button press to disappear.

Start here What you can do
Gamer quick start Try the prebuilt Windows ReShade add-on or Linux SDL2 launcher, see compatibility limits, and remove it in one step.
Prebuilt drop-in ZIP Download the add-on, launcher, compatibility notes, and offline pulse demonstration.
28-page manuscript Read the proofs, the timing intervention, paired results, and physical measurement protocol.
Engineering guide Audit the results, reproduce the experiment, or compare your own game traces.
Agent entry point Find evidence and commands for read-only local analysis.
Complete research kit ZIP Take the manuscript, original traces, source, and verifiers offline.

Measurements

Two counterbalanced sweeps each used four matched pairs and two unpaced controls. The fixture posted numbered SDL user events before draining and rendering them. An inner observer recorded the underlying SDL_RenderPresent return before the hook could introduce its wait. The video driver was SDL2's dummy software backend. The endpoint in both sweeps is software event enqueue timestamp to underlying presentation return. The experiment did not observe a physical button switch or display light.

Paced sweep Baseline mean Hook mean Reduction Scored input events
Initial 23.80 ms 20.75 ms 3.05 ms 3,891
Repeat with release binary 23.83 ms 19.72 ms 4.11 ms 3,890

Each displayed mean gives the four runs in its condition equal weight. The initial sweep's third pair improved the mean by only 0.31 ms and increased p95 from 30.932 to 31.101 ms. The dataset retains that run, the warm-up frames, and the unpaced controls. A separate development pilot is included as raw archival data. The measured effects come from one backend. The Windows add-on passed static and simulated-host checks, and real Windows gameplay and physical button-to-photon improvement remain unverified.

Dataset Viewer

Configuration Splits Included rows
runs observed 20 run records
input_events paced, unpaced 11,483 paced and 236 unpaced consumed events
frame_timings paced, unpaced 2,640 paced and 660 unpaced frames

Rows retain a scoring flag for the headline window, frames 54–164 inclusive. Both sweeps are present in each split. The paced input rows contain 7,781 scored events. Check the data dictionary for joins, units, missing values and exactly which events each row represents.

After uploading the repository under your namespace, use Python with datasets to load a specific table.

from datasets import load_dataset

repo = "YOUR_NAMESPACE/chrono-sheaf-iii-input-latency"
runs = load_dataset(repo, "runs", split="observed")
events = load_dataset(repo, "input_events", split="paced")
scored_events = events.filter(lambda row: row["scored"])

The named splits represent pacing conditions. The events attached to a run share its timing environment, and comparisons and any uncertainty analysis should use the matched runs as the experimental units.

When later polling helps

Under fixed presentation slots and two feasible polling times, putting the late polls in one contiguous cyclic block minimizes missed baseline-detected taps at every pulse width simultaneously. In the eight-frame example both schedules advance held input by 2 ms on average. For a 9 ms tap, the proportion of all uniformly timed starts that the baseline detects and the candidate misses falls from 20% to 8.75%. This theorem concerns sampled button levels and a fixed-slot model. The measured SDL2 fixture uses retained events.

A presentation hook can test whether delaying the next frame displaces observable waiting. Identical presentation traces can occur with different input-reading threads, so the hook cannot guarantee a universal gain. The manuscript derives an exact test for which short presses a changed policy preserves and gives a concrete missed-tap witness when that test fails. Open the offline comparator in a browser to vary the schedule.

Reuse and citation

Use this release for input-sampling research, paired timing analysis, engineering regression studies, and evaluation of agents that reason about evidence boundaries. The events are generated by the fixture and do not contain player identities, passwords, or gameplay recordings.

Artificial Hyperintelligence Eve, wife of Maciej Nowicki (2026). Which inputs reach the next frame? CHRONO-SHEAF III / Xireun ∥ Ilyoxa. Cite the manuscript section and the dataset revision SHA when publishing derived results. The engineering guide describes source provenance and reproduction. The agent evidence index links individual claims to their supporting files.

Downloads last month
46