tiny-geometry-v3-93ed354-ff42f4dd-32a80ab5
completed —
Hypothesis
The diverse-four aggregate held-out accuracy materially overstates genuine denoising: accuracy on actually changed coordinate tokens will trail aggregate accuracy by at least 0.15, while all v2 learning outputs remain otherwise unchanged.
Why run it
Determines whether the next Gate E experiment should target corrupted-token recovery rather than model capacity, additional optimization steps, or topology.
Measured behaviour
Table view
| optimizer step | train accuracy |
|---|---|
| 1 | 0.0012 |
| 10 | 0.2892 |
| 20 | 0.6366 |
| 30 | 0.8676 |
| 40 | 0.9694 |
| 50 | 0.9975 |
| 60 | 1.0000 |
| 70 | 1.0000 |
| 80 | 1.0000 |
| 90 | 1.0000 |
| 100 | 1.0000 |
| 110 | 1.0000 |
| 120 | 1.0000 |
| 130 | 1.0000 |
| 140 | 1.0000 |
| 150 | 1.0000 |
| 160 | 1.0000 |
| 1 | 0.0028 |
| 10 | 0.1141 |
| 20 | 0.4179 |
| 30 | 0.7195 |
| 40 | 0.8920 |
| 50 | 0.9649 |
| 60 | 0.9948 |
| 70 | 0.9993 |
| 80 | 1.0000 |
| 90 | 1.0000 |
| 100 | 1.0000 |
| 110 | 1.0000 |
| 120 | 1.0000 |
| 130 | 1.0000 |
| 140 | 1.0000 |
| 150 | 1.0000 |
| 160 | 1.0000 |
| 170 | 1.0000 |
| 180 | 1.0000 |
| 190 | 1.0000 |
| 200 | 1.0000 |
| 210 | 1.0000 |
| 220 | 1.0000 |
| 230 | 1.0000 |
| 240 | 1.0000 |
| 250 | 1.0000 |
| 260 | 1.0000 |
| 270 | 1.0000 |
| 280 | 1.0000 |
| 290 | 1.0000 |
| 300 | 1.0000 |
| 310 | 1.0000 |
| 320 | 1.0000 |
Measured result
Verbatim from the run's summary.json.
cases
diverse-four
checkpoint_sha256
52aee590c65f81f5…config
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
resume_step
160
seed
1702
steps
320
continuous_loss_sha256
240bd2ca34fd1b1a…final
accuracy
0.7382404181184669
changed_accuracy
0.5817073170731707
changed_correct
954
changed_total
1640
correct
3390
loss
1.3494370430707932
retained_accuracy
0.8252032520325203
retained_correct
2436
retained_total
2952
total
4592
final_model_sha256
bd0b0814a19c7402…heldout_examples
16
initial
accuracy
0.002395470383275261
changed_accuracy
0.0024390243902439024
changed_correct
4
changed_total
1640
correct
11
loss
14.943475902080536
retained_accuracy
0.0023712737127371273
retained_correct
7
retained_total
2952
total
4592
loss_ratio
5.3330747233888907e-05
passes_predeclared_criteria
false
resume_exact
true
resumed_loss_sha256
240bd2ca34fd1b1a…train_examples
16
train_final
accuracy
1.0
changed_accuracy
1.0
changed_correct
1622
changed_total
1622
correct
4592
loss
0.0007702835428062826
retained_accuracy
1.0
retained_correct
2970
retained_total
2970
total
4592
one-icon
checkpoint_sha256
None
config
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
resume_step
None
seed
1701
steps
160
continuous_loss_sha256
None
final
accuracy
0.9356617647058824
changed_accuracy
0.8858695652173914
changed_correct
489
changed_total
552
correct
1527
loss
0.2106002140790224
retained_accuracy
0.9611111111111111
retained_correct
1038
retained_total
1080
total
1632
final_model_sha256
8e0a9878092058a0…heldout_examples
8
initial
accuracy
0.001838235294117647
changed_accuracy
0.005434782608695652
changed_correct
3
changed_total
552
correct
3
loss
15.358022212982178
retained_accuracy
0.0
retained_correct
0
retained_total
1080
total
1632
loss_ratio
9.539973069621893e-05
passes_predeclared_criteria
false
resume_exact
None
resumed_loss_sha256
None
train_examples
8
train_final
accuracy
1.0
changed_accuracy
1.0
changed_correct
599
changed_total
599
correct
1632
loss
0.001458597937016748
retained_accuracy
1.0
retained_correct
1033
retained_total
1033
total
1632
code_identity
geometry_sha256
8f30cf4c834873bd…git_commit
93ed3542a69f1697c4d3ca13ba1982cc2cda5828
packed_sha256
9b12879ba3a18236…program_sha256
0d02b532b1dbe573…tiny_study_sha256
578ec1d402fdb88f…config_sha256
ff42f4dd0b8915f0…deterministic_algorithms
true
device
cpu
fixture
data/manifests/openmoji-17.0.0-tiny-learning-fixture-v1.json
fixture_sha256
32a80ab576a6d5a4…metrics_sha256
c4e0d9a8c7873466…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
d4c36c75381689fb…schema_version
1
study_version
tiny-geometry-v3-corruption-split
torch_version
2.8.0+cpu
Visual output
Written result
Completed reproducibly; the predeclared diagnostic passed.
For diverse-four, actually-changed-token accuracy is 58.17%, 15.65 points below the 73.82% aggregate. Retained-token accuracy is also limited at 82.52%, so the model often overwrites coordinates already correct in x_t. The one-icon case is much stronger: 88.59% changed and 96.11% retained accuracy. A complete rerun reproduced all artifacts.
This supports testing broader deterministic corruption coverage at the same batch size, model, probability, and optimizer-step count before increasing capacity or steps.
State transitions
- planned2026-08-30T05:55:54.629673Z
- completed2026-08-30T06:01:12.713902Zaggregate_accuracy_materially_overstates_changed_token_recovery_and_diverse_model_also_overwrites_retained_coordinates
Run record
Verbatim from runs/tiny-geometry-v3-93ed354-ff42f4dd-32a80ab5/run.yaml, the record committed before launch.
schema_version
1
run_id
tiny-geometry-v3-93ed354-ff42f4dd-32a80ab5
state
completed
hypothesis
The diverse-four aggregate held-out accuracy materially overstates genuine denoising: accuracy on actually changed coordinate tokens will trail aggregate accuracy by at least 0.15, while all v2 learning outputs remain otherwise unchanged.
expected_information_gain
Determines whether the next Gate E experiment should target corrupted-token recovery rather than model capacity, additional optimization steps, or topology.
parent_run
tiny-geometry-v2-47811d0-75d558c9-32a80ab5
git_commit
93ed3542a69f1697c4d3ca13ba1982cc2cda5828
dirty_patch_sha256
e3b0c44298fc1c14…config
configs/learning/tiny-geometry-v3-corruption-split.yaml
config_sha256
ff42f4dd0b8915f0…dataset_manifest
data/manifests/openmoji-17.0.0-tiny-learning-fixture-v1.json
dataset_manifest_sha256
32a80ab576a6d5a4…worker
gtc-local-cpu
environment
hostname
gtc
architecture
x86_64
python
3.12.3
torch
2.8.0+cpu
numpy
2.5.2
device
cpu
training_threads
2
gpu
None
precision
float32
seeds
1701, 1702
controlled_factor
evaluation metric partition only
held_constant
fixture, model and initialization, corruption draws, optimizer and batches, training steps, aggregate acceptance thresholds
predeclared_diagnostic
diverse_four_changed_accuracy_gap_minimum
0.15
resource_cap
external_spend_usd
0
optimizer_steps
800
train_examples
24
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-v3-corruption-split.yaml
outputs
report
reports/learning/tiny-geometry-v3-corruption-split
derived
data/processed/tiny-geometry-v3-corruption-split
run_record
runs/tiny-geometry-v3-93ed354-ff42f4dd-32a80ab5
artifact_durability
compact-report-pending-local-git;checkpoint-local-only
result
diagnostic_hypothesis_passed
true
artifact_identity_rerun
true
diverse_four
aggregate_accuracy
0.7382404181184669
changed_accuracy
0.5817073170731707
retained_accuracy
0.8252032520325203
aggregate_minus_changed
0.1565331010452962
one_icon
aggregate_accuracy
0.9356617647058824
changed_accuracy
0.8858695652173914
retained_accuracy
0.9611111111111111