tiny-geometry-f1-path-correlated-b0983dd-c0acb759-32a80ab5
completed —
Hypothesis
Path-block-correlated geometry corruption, at the same marginal legal-token change probability as the factorized control, changes learning or trajectory quality.
Why run it
Isolates within-path corruption correlation from model, data, optimizer, marginal token rate, topology, styles, and render protocol.
Measured behaviour
Table view
| optimizer step | train accuracy |
|---|---|
| 1 | 0.0025 |
| 10 | 0.2763 |
| 20 | 0.5839 |
| 30 | 0.7862 |
| 40 | 0.7083 |
| 50 | 0.7286 |
| 60 | 0.7960 |
| 70 | 0.7972 |
| 80 | 0.8542 |
| 90 | 0.7862 |
| 100 | 0.8879 |
| 110 | 0.8873 |
| 120 | 0.8915 |
| 130 | 0.9013 |
| 140 | 0.9259 |
| 150 | 0.9626 |
| 160 | 0.9418 |
| 1 | 0.0022 |
| 10 | 0.1119 |
| 20 | 0.4260 |
| 30 | 0.5760 |
| 40 | 0.6686 |
| 50 | 0.6895 |
| 60 | 0.7853 |
| 70 | 0.7864 |
| 80 | 0.7807 |
| 90 | 0.7509 |
| 100 | 0.7872 |
| 110 | 0.7975 |
| 120 | 0.8151 |
| 130 | 0.8604 |
| 140 | 0.8630 |
| 150 | 0.8700 |
| 160 | 0.8778 |
| 170 | 0.8987 |
| 180 | 0.8976 |
| 190 | 0.9031 |
| 200 | 0.9111 |
| 210 | 0.9268 |
| 220 | 0.9336 |
| 230 | 0.9421 |
| 240 | 0.9506 |
| 250 | 0.9440 |
| 260 | 0.9506 |
| 270 | 0.9458 |
| 280 | 0.9595 |
| 290 | 0.9678 |
| 300 | 0.9736 |
| 310 | 0.9715 |
| 320 | 0.9815 |
Measured result
Verbatim from the run's summary.json.
cases
diverse-four
checkpoint_sha256
82f5c2a8fb785ddf…config
corruption_kind
path_correlated_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
52099a77d0e8b87d…final
accuracy
0.9810540069686411
changed_accuracy
0.9386459802538787
changed_correct
1331
changed_total
1418
correct
4505
loss
0.05304072146827821
retained_accuracy
1.0
retained_correct
3174
retained_total
3174
total
4592
final_model_sha256
041db3d04db45de4…heldout_examples
16
initial
accuracy
0.002395470383275261
changed_accuracy
0.004231311706629055
changed_correct
6
changed_total
1418
correct
11
loss
14.97111314535141
retained_accuracy
0.001575299306868305
retained_correct
5
retained_total
3174
total
4592
loss_ratio
0.003879326103807346
passes_predeclared_criteria
true
resume_exact
true
resumed_loss_sha256
52099a77d0e8b87d…train_examples
16
train_final
accuracy
0.9817073170731707
changed_accuracy
0.9451697127937336
changed_correct
1448
changed_total
1532
correct
4508
loss
0.06225258283666335
retained_accuracy
1.0
retained_correct
3060
retained_total
3060
total
4592
one-icon
checkpoint_sha256
None
config
corruption_kind
path_correlated_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
0.9552696078431373
changed_accuracy
0.8355855855855856
changed_correct
371
changed_total
444
correct
1559
loss
0.1168382684700191
retained_accuracy
1.0
retained_correct
1188
retained_total
1188
total
1632
final_model_sha256
79a9ee3e7c1e58cb…heldout_examples
8
initial
accuracy
0.0012254901960784314
changed_accuracy
0.0045045045045045045
changed_correct
2
changed_total
444
correct
2
loss
15.369378566741943
retained_accuracy
0.0
retained_correct
0
retained_total
1188
total
1632
loss_ratio
0.00921818651754486
passes_predeclared_criteria
false
resume_exact
None
resumed_loss_sha256
None
train_examples
8
train_final
accuracy
0.9350490196078431
changed_accuracy
0.8490028490028491
changed_correct
596
changed_total
702
correct
1526
loss
0.17234542884398252
retained_accuracy
1.0
retained_correct
930
retained_total
930
total
1632
code_identity
geometry_sha256
1ce82851e51941e4…git_commit
b0983dd00edfce9bb9c5045671edbf767dae6bea
packed_sha256
9b12879ba3a18236…program_sha256
0d02b532b1dbe573…tiny_study_sha256
e1dd8c00e3aa0398…config_sha256
c0acb759be32561d…deterministic_algorithms
true
device
cpu
fixture
data/manifests/openmoji-17.0.0-tiny-learning-fixture-v1.json
fixture_sha256
32a80ab576a6d5a4…metrics_sha256
c4f97dcd8a5a17be…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
9b41fbcb8870576d…schema_version
1
study_version
tiny-geometry-f1-path-correlated
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 | 0.9350 | 0.9553 | 0.8356 | 1.0000 | None | False |
| diverse-four | 0.9817 | 0.9811 | 0.9386 | 1.0000 | 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
- planned2026-08-30T06:40:00Z
- completed2026-08-30T06:47:00Zpath_correlated_fixed_topology_treatment_completes_but_trails_factorized_control_on_changed_token_recovery
Run record
Verbatim from runs/tiny-geometry-f1-path-correlated-b0983dd-c0acb759-32a80ab5/run.yaml, the record committed before launch.
schema_version
1
run_id
tiny-geometry-f1-path-correlated-b0983dd-c0acb759-32a80ab5
state
completed
hypothesis
Path-block-correlated geometry corruption, at the same marginal legal-token change probability as the factorized control, changes learning or trajectory quality.
expected_information_gain
Isolates within-path corruption correlation from model, data, optimizer, marginal token rate, topology, styles, and render protocol.
parent_run
tiny-geometry-f1-factorized-333e149-e57646b7-32a80ab5
git_commit
b0983dd00edfce9bb9c5045671edbf767dae6bea
dirty_patch_sha256
e3b0c44298fc1c14…config
configs/learning/tiny-geometry-f1-path-correlated.yaml
config_sha256
c0acb759be32561d…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
one shared corruption gate per active path geometry block
held_constant
fixed topology, styles, and typed padding, fixture, model, initialization, optimizer, and step budgets, marginal 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-path-correlated.yaml
outputs
report
reports/corruption/tiny-geometry-f1-path-correlated
derived
data/processed/tiny-geometry-f1-path-correlated
run_record
runs/tiny-geometry-f1-path-correlated-b0983dd-c0acb759-32a80ab5
artifact_durability
compact-report-pending-local-git;checkpoint-local-only
result
artifact_identity_rerun
true
checkpoint_resume_exact
true
one_icon
heldout_accuracy
0.9552696078431373
changed_accuracy
0.8355855855855856
retained_accuracy
1.0
diverse_four
heldout_accuracy
0.9810540069686411
changed_accuracy
0.9386459802538787
retained_accuracy
1.0