tiny-geometry-f1-factorized-333e149-e57646b7-32a80ab5
completed —
Hypothesis
Under the Gate E fixed-topology setup, the exact factorized role-uniform control provides a reproducible baseline for comparing within-path corruption correlation.
Why run it
Establishes matched learning curves, changed-token recovery, valid paired trajectories, checkpoint continuation, and artifact identity before interpreting the correlated arm.
Measured behaviour
Table view
| optimizer step | train accuracy |
|---|---|
| 1 | 0.0031 |
| 10 | 0.2635 |
| 20 | 0.5833 |
| 30 | 0.7806 |
| 40 | 0.8964 |
| 50 | 0.9314 |
| 60 | 0.9626 |
| 70 | 0.9651 |
| 80 | 0.9663 |
| 90 | 0.9516 |
| 100 | 0.9810 |
| 110 | 0.9786 |
| 120 | 0.9902 |
| 130 | 0.9816 |
| 140 | 0.9773 |
| 150 | 0.9884 |
| 160 | 0.9853 |
| 1 | 0.0024 |
| 10 | 0.0849 |
| 20 | 0.3081 |
| 30 | 0.5640 |
| 40 | 0.7117 |
| 50 | 0.7844 |
| 60 | 0.8369 |
| 70 | 0.8772 |
| 80 | 0.9055 |
| 90 | 0.9209 |
| 100 | 0.9273 |
| 110 | 0.9281 |
| 120 | 0.9482 |
| 130 | 0.9477 |
| 140 | 0.9654 |
| 150 | 0.9645 |
| 160 | 0.9684 |
| 170 | 0.9641 |
| 180 | 0.9680 |
| 190 | 0.9691 |
| 200 | 0.9787 |
| 210 | 0.9717 |
| 220 | 0.9743 |
| 230 | 0.9804 |
| 240 | 0.9719 |
| 250 | 0.9804 |
| 260 | 0.9795 |
| 270 | 0.9813 |
| 280 | 0.9793 |
| 290 | 0.9837 |
| 300 | 0.9797 |
| 310 | 0.9858 |
| 320 | 0.9874 |
Measured result
Verbatim from the run's summary.json.
cases
diverse-four
checkpoint_sha256
504e0880889449b1…config
corruption_kind
factorized_geometry
corruptions_per_icon
4
heldout_corruptions_per_icon
4
icon_count
4
max_loss_ratio
0.2
min_heldout_accuracy
0.85
min_train_accuracy
0.98
name
diverse-four
resample_each_step
true
resume_step
160
seed
1702
steps
320
continuous_loss_sha256
83d82ab6c855a815…final
accuracy
0.9875871080139372
changed_accuracy
0.9695863746958637
changed_correct
1594
changed_total
1644
correct
4535
loss
0.04504280601395294
retained_accuracy
0.9976255088195387
retained_correct
2941
retained_total
2948
total
4592
final_model_sha256
505579f99c46b827…heldout_examples
16
initial
accuracy
0.002395470383275261
changed_accuracy
0.0030413625304136255
changed_correct
5
changed_total
1644
correct
11
loss
14.920220911502838
retained_accuracy
0.0020352781546811396
retained_correct
6
retained_total
2948
total
4592
loss_ratio
0.0031178024965364862
passes_predeclared_criteria
true
resume_exact
true
resumed_loss_sha256
83d82ab6c855a815…train_examples
16
train_final
accuracy
0.9878048780487805
changed_accuracy
0.9704978488014752
changed_correct
1579
changed_total
1627
correct
4536
loss
0.042252493440173566
retained_accuracy
0.9973018549747049
retained_correct
2957
retained_total
2965
total
4592
one-icon
checkpoint_sha256
None
config
corruption_kind
factorized_geometry
corruptions_per_icon
8
heldout_corruptions_per_icon
8
icon_count
1
max_loss_ratio
0.1
min_heldout_accuracy
0.95
min_train_accuracy
0.99
name
one-icon
resample_each_step
true
resume_step
None
seed
1701
steps
160
continuous_loss_sha256
None
final
accuracy
0.9932598039215687
changed_accuracy
0.9873188405797102
changed_correct
545
changed_total
552
correct
1621
loss
0.028571574424859136
retained_accuracy
0.9962962962962963
retained_correct
1076
retained_total
1080
total
1632
final_model_sha256
626cd212c55ba27b…heldout_examples
8
initial
accuracy
0.001838235294117647
changed_accuracy
0.0036231884057971015
changed_correct
2
changed_total
552
correct
3
loss
15.33237636089325
retained_accuracy
0.000925925925925926
retained_correct
1
retained_total
1080
total
1632
loss_ratio
0.002342729000340221
passes_predeclared_criteria
false
resume_exact
None
resumed_loss_sha256
None
train_examples
8
train_final
accuracy
0.9797794117647058
changed_accuracy
0.9601328903654485
changed_correct
578
changed_total
602
correct
1599
loss
0.06699059216771275
retained_accuracy
0.9912621359223301
retained_correct
1021
retained_total
1030
total
1632
code_identity
geometry_sha256
1ce82851e51941e4…git_commit
333e1492e94dfc87df23a907e72c9ca941515342
packed_sha256
9b12879ba3a18236…program_sha256
0d02b532b1dbe573…tiny_study_sha256
e1dd8c00e3aa0398…config_sha256
e57646b7f70d51e4…deterministic_algorithms
true
device
cpu
fixture
data/manifests/openmoji-17.0.0-tiny-learning-fixture-v1.json
fixture_sha256
32a80ab576a6d5a4…metrics_sha256
c9b85e5582d0947e…model
d_model
48
feedforward
96
heads
4
layers
2
parameters
241072
scope
fixed-topology geometry-only diagnostic
passes_predeclared_gate_candidate
false
render_metrics_sha256
6db25d963f173bd8…schema_version
1
study_version
tiny-geometry-f1-factorized
torch_version
2.8.0+cpu
Visual output
Written result
PyTorch 2.8.0+cpu on deterministic CPU; 241072 parameters.
| case | train accuracy | held-out accuracy | changed accuracy | retained accuracy | resume exact | passes |
|---|---|---|---|---|---|---|
| one-icon | 0.9798 | 0.9933 | 0.9873 | 0.9963 | None | False |
| diverse-four | 0.9878 | 0.9876 | 0.9696 | 0.9976 | True | True |
This is a fixed-topology, geometry-only diagnostic. It is evidence that the packed pipeline can learn and resume; it is not evidence for the final corruption process or unconditional generation.
State transitions
- planned2026-08-30T06:30:00Z
- completed2026-08-30T06:36:00Zmatched_factorized_fixed_topology_control_completes_with_strong_heldout_recovery_and_render_safe_trajectories
Run record
Verbatim from runs/tiny-geometry-f1-factorized-333e149-e57646b7-32a80ab5/run.yaml, the record committed before launch.
schema_version
1
run_id
tiny-geometry-f1-factorized-333e149-e57646b7-32a80ab5
state
completed
hypothesis
Under the Gate E fixed-topology setup, the exact factorized role-uniform control provides a reproducible baseline for comparing within-path corruption correlation.
expected_information_gain
Establishes matched learning curves, changed-token recovery, valid paired trajectories, checkpoint continuation, and artifact identity before interpreting the correlated arm.
parent_run
tiny-geometry-v5-44ce3de-a34456a8-32a80ab5
git_commit
333e1492e94dfc87df23a907e72c9ca941515342
dirty_patch_sha256
e3b0c44298fc1c14…config
configs/learning/tiny-geometry-f1-factorized.yaml
config_sha256
e57646b7f70d51e4…dataset_manifest
data/manifests/openmoji-17.0.0-tiny-learning-fixture-v1.json
dataset_manifest_sha256
32a80ab576a6d5a4…worker
gtc-local-cpu
environment
device
cpu
torch
2.8.0+cpu
deterministic_algorithms
true
training_threads
2
controlled_factor
independent legal geometry-token gates
held_constant
fixed topology, styles, and typed padding, fixture, model, initialization, optimizer, and step budgets, corruption probability, per-step resampling, held-out draws, and render protocol, checkpoint boundary and canonical checkpoint format
resource_cap
external_spend_usd
0
optimizer_steps
800
heldout_examples
24
isolated_renders
24
max_local_storage_gb
0.1
command
.venv/bin/python -m mojidiff.learning.tiny_study --config configs/learning/tiny-geometry-f1-factorized.yaml
outputs
report
reports/corruption/tiny-geometry-f1-factorized
derived
data/processed/tiny-geometry-f1-factorized
run_record
runs/tiny-geometry-f1-factorized-333e149-e57646b7-32a80ab5
artifact_durability
compact-report-pending-local-git;checkpoint-local-only
result
artifact_identity_rerun
true
checkpoint_resume_exact
true
one_icon
heldout_accuracy
0.9932598039215687
changed_accuracy
0.9873188405797102
retained_accuracy
0.9962962962962963
online_train_probe_accuracy
0.9797794117647058
diverse_four
heldout_accuracy
0.9875871080139372
changed_accuracy
0.9695863746958637
retained_accuracy
0.9976255088195387
online_train_probe_accuracy
0.9878048780487805