tiny-geometry-v1-3493eff-bf71ca62-32a80ab5
failed declared
Hypothesis
A compact fixed-topology geometry denoiser can overfit one OpenMoji icon, learn a four-icon diverse fixture well enough to recover disjoint held-out corruptions, and continue bit-exactly after a checkpoint reload.
Why run it
Separates basic packed-geometry learnability and checkpoint correctness from the harder unresolved questions of topology generation and the final corruption process.
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
35d62dcfe460f574…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
correct
3390
loss
1.3494370430707932
total
4592
final_model_sha256
bd0b0814a19c7402…heldout_examples
16
initial
accuracy
0.002395470383275261
correct
11
loss
14.943475902080536
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
correct
4592
loss
0.0007702835428062826
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
correct
1527
loss
0.2106002140790224
total
1632
final_model_sha256
8e0a9878092058a0…heldout_examples
8
initial
accuracy
0.001838235294117647
correct
3
loss
15.358022212982178
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
correct
1632
loss
0.001458597937016748
total
1632
code_identity
geometry_sha256
537d4e25da7352c8…git_commit
3493effb6b860c40be075ef0c546935c0a18d660
packed_sha256
9b12879ba3a18236…program_sha256
0d02b532b1dbe573…tiny_study_sha256
842908765b9f651b…config_sha256
bf71ca62d7f3e3a3…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-v1
torch_version
2.8.0+cpu
Visual output
Written result
Failed the predeclared learning criterion and the artifact-idempotency rerun.
Both cases reached 100% training-token accuracy. One-icon held-out accuracy was 93.57% against a 95% threshold; diverse-four held-out accuracy was 73.82% against an 85% threshold. The diverse checkpoint continuation reproduced every loss and final model tensor exactly.
The identical full rerun then failed closed because legacy torch.save checkpoint container bytes differed. The original checkpoint and all first-execution evidence are preserved. This does not invalidate the learning metrics, but it prevents v1 from passing the project's reproducible-artifact contract.
State transitions
- planned2026-08-30T05:36:21.749863Z
- failed2026-08-30T05:42:45.573520Z
Run record
Verbatim from runs/tiny-geometry-v1-3493eff-bf71ca62-32a80ab5/run.yaml, the record committed before launch.
schema_version
1
run_id
tiny-geometry-v1-3493eff-bf71ca62-32a80ab5
state
failed
hypothesis
A compact fixed-topology geometry denoiser can overfit one OpenMoji icon, learn a four-icon diverse fixture well enough to recover disjoint held-out corruptions, and continue bit-exactly after a checkpoint reload.
expected_information_gain
Separates basic packed-geometry learnability and checkpoint correctness from the harder unresolved questions of topology generation and the final corruption process.
parent_run
packed-render-stress-v1-d5ca696-df13aed7-005dc6b4
git_commit
3493effb6b860c40be075ef0c546935c0a18d660
dirty_patch_sha256
e3b0c44298fc1c14…config
configs/learning/tiny-geometry-v1.yaml
config_sha256
bf71ca62d7f3e3a3…dataset_manifest
data/manifests/openmoji-17.0.0-tiny-learning-fixture-v1.json
dataset_manifest_sha256
32a80ab576a6d5a4…source_revision
f9fc506a3f913be9897ab0181d611d4c910a4104
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
driver
None
cuda
None
cudnn
None
tensorrt
None
container_image
None
precision
float32
seeds
1701, 1702
determinism
deterministic_algorithms
true
dropout
0.0
fixed_full_batch_order
true
predeclared_criteria
one-icon
min_train_accuracy
0.99
min_heldout_accuracy
0.95
max_loss_ratio
0.1
diverse-four
min_train_accuracy
0.98
min_heldout_accuracy
0.85
max_loss_ratio
0.2
exact_checkpoint_continuation
true
resource_cap
external_spend_usd
0
optimizer_steps
800
train_examples
24
heldout_examples
24
isolated_renders
24
max_local_storage_gb
0.1
checkpoint_policy
Keep the single compact resume checkpoint under ignored reproducible processed data; record and verify its hash in the compact report.
command
.venv/bin/python -m mojidiff.learning.tiny_study --config configs/learning/tiny-geometry-v1.yaml
outputs
report
reports/learning/tiny-geometry-v1
derived
data/processed/tiny-geometry-v1
run_record
runs/tiny-geometry-v1-3493eff-bf71ca62-32a80ab5
artifact_durability
compact-report-pending-local-git;checkpoint-local-only
failure
stage
reproducibility-rerun
reason
The first execution completed and produced deterministic learning metrics, but the identical rerun generated a semantically equivalent legacy PyTorch checkpoint with different container bytes. The create-or-identical guard rejected replacement.
first_execution_exit_code
0
reproducibility_rerun_exit_code
1
scientific_hypothesis_passed
false