tiny-geometry-v5-44ce3de-a34456a8-32a80ab5
completed declared
Hypothesis
Applying the v4 deterministic per-step coverage treatment to one-icon as well will make both predeclared cases pass while preserving diverse recovery, exact checkpoint continuation, recognizable renders, and byte-identical artifacts.
Why run it
Final controlled check for Gate E's one-icon overfit, tiny-diverse recovery, resume, and rendering exit evidence before beginning corruption-family comparisons.
Measured behaviour
Table view
| optimizer step | train accuracy |
|---|---|
| 1 | 0.0018 |
| 10 | 0.2672 |
| 20 | 0.5858 |
| 30 | 0.7855 |
| 40 | 0.8848 |
| 50 | 0.9161 |
| 60 | 0.9583 |
| 70 | 0.9645 |
| 80 | 0.9675 |
| 90 | 0.9626 |
| 100 | 0.9877 |
| 110 | 0.9755 |
| 120 | 0.9835 |
| 130 | 0.9810 |
| 140 | 0.9828 |
| 150 | 0.9896 |
| 160 | 0.9920 |
| 1 | 0.0020 |
| 10 | 0.0804 |
| 20 | 0.3095 |
| 30 | 0.5623 |
| 40 | 0.7269 |
| 50 | 0.7879 |
| 60 | 0.8441 |
| 70 | 0.8789 |
| 80 | 0.9031 |
| 90 | 0.9262 |
| 100 | 0.9294 |
| 110 | 0.9299 |
| 120 | 0.9503 |
| 130 | 0.9456 |
| 140 | 0.9608 |
| 150 | 0.9639 |
| 160 | 0.9682 |
| 170 | 0.9671 |
| 180 | 0.9658 |
| 190 | 0.9678 |
| 200 | 0.9793 |
| 210 | 0.9699 |
| 220 | 0.9680 |
| 230 | 0.9795 |
| 240 | 0.9760 |
| 250 | 0.9811 |
| 260 | 0.9791 |
| 270 | 0.9782 |
| 280 | 0.9782 |
| 290 | 0.9854 |
| 300 | 0.9837 |
| 310 | 0.9863 |
| 320 | 0.9852 |
Measured result
Verbatim from the run's summary.json.
cases
diverse-four
checkpoint_sha256
e975f17c15497260…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
resample_each_step
true
resume_step
160
seed
1702
steps
320
continuous_loss_sha256
1c909c13042da5ba…final
accuracy
0.9856271777003485
changed_accuracy
0.9664634146341463
changed_correct
1585
changed_total
1640
correct
4526
loss
0.04804584034718573
retained_accuracy
0.9962737127371274
retained_correct
2941
retained_total
2952
total
4592
final_model_sha256
d7972e2a714662fd…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
0.0029922704661883926
passes_predeclared_criteria
true
resume_exact
true
resumed_loss_sha256
1c909c13042da5ba…train_examples
16
train_final
accuracy
0.985191637630662
changed_accuracy
0.9636251541307028
changed_correct
1563
changed_total
1622
correct
4524
loss
0.05133143730927259
retained_accuracy
0.996969696969697
retained_correct
2961
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
resample_each_step
true
resume_step
None
seed
1701
steps
160
continuous_loss_sha256
None
final
accuracy
0.9926470588235294
changed_accuracy
0.9836956521739131
changed_correct
543
changed_total
552
correct
1620
loss
0.02712903532665223
retained_accuracy
0.9972222222222222
retained_correct
1077
retained_total
1080
total
1632
final_model_sha256
7029c413d5bdce0c…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
0.0019676517358534536
passes_predeclared_criteria
false
resume_exact
None
resumed_loss_sha256
None
train_examples
8
train_final
accuracy
0.977328431372549
changed_accuracy
0.9515859766277128
changed_correct
570
changed_total
599
correct
1595
loss
0.06208717590197921
retained_accuracy
0.9922555663117134
retained_correct
1025
retained_total
1033
total
1632
code_identity
geometry_sha256
8f30cf4c834873bd…git_commit
44ce3de6f0f76c5815cf82c5b6d3dda0d28a8ae0
packed_sha256
9b12879ba3a18236…program_sha256
0d02b532b1dbe573…tiny_study_sha256
8e09bccf8d3c1b90…config_sha256
a34456a8e5087700…deterministic_algorithms
true
device
cpu
fixture
data/manifests/openmoji-17.0.0-tiny-learning-fixture-v1.json
fixture_sha256
32a80ab576a6d5a4…metrics_sha256
2805101702954e39…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
8eafdf8ce49f84c2…schema_version
1
study_version
tiny-geometry-v5-all-resampled
torch_version
2.8.0+cpu
Visual output
Written result
Completed reproducibly with a mixed per-run result and positive cumulative Gate E exit.
One-icon held-out accuracy rises to 99.26% (98.37% changed, 99.72% retained), and diverse-four remains at 98.56%. The fixed one-icon probe is 97.73%, below its legacy 99% memorization threshold, so v5's literal both-criteria hypothesis is false. That probe is an unseen batch under online resampling, not the batch v1 overfit.
Across the controlled sequence, Gate E is complete: v1 records 100% one-icon overfit; v5 records strong one-icon held-out recovery; v4/v5 record strong diverse recovery and recognizable renders; v2-v5 record exact checkpoint continuation and artifact identity.
State transitions
- planned2026-08-30T06:13:13.901491Z
- completed2026-08-30T06:19:09.284898Zone_icon_and_diverse_heldout_recovery_are_strong_but_v5_fixed_probe_misses_legacy_memorization_threshold_gate_e_closes_from_controlled_sequence
Run record
Verbatim from runs/tiny-geometry-v5-44ce3de-a34456a8-32a80ab5/run.yaml, the record committed before launch.
schema_version
1
run_id
tiny-geometry-v5-44ce3de-a34456a8-32a80ab5
state
completed
hypothesis
Applying the v4 deterministic per-step coverage treatment to one-icon as well will make both predeclared cases pass while preserving diverse recovery, exact checkpoint continuation, recognizable renders, and byte-identical artifacts.
expected_information_gain
Final controlled check for Gate E's one-icon overfit, tiny-diverse recovery, resume, and rendering exit evidence before beginning corruption-family comparisons.
parent_run
tiny-geometry-v4-ede1009-7ccbc64e-32a80ab5
git_commit
44ce3de6f0f76c5815cf82c5b6d3dda0d28a8ae0
dirty_patch_sha256
e3b0c44298fc1c14…config
configs/learning/tiny-geometry-v5-all-resampled.yaml
config_sha256
a34456a8e5087700…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
one-icon corruption draw coverage
held_constant
fixture and held-out draws, model and initialization, batch sizes and corruption probability, optimizer and step budgets, diverse-four v4 treatment, checkpoint boundary
predeclared_criteria
one_icon_train_accuracy_minimum
0.99
one_icon_heldout_accuracy_minimum
0.95
one_icon_loss_ratio_maximum
0.1
diverse_four_train_accuracy_minimum
0.98
diverse_four_heldout_accuracy_minimum
0.85
diverse_four_loss_ratio_maximum
0.2
checkpoint_resume_exact
true
artifact_identity_rerun
true
resource_cap
external_spend_usd
0
optimizer_steps
800
batch_size_one_icon
8
batch_size_diverse
16
distinct_corruptions_per_icon_continuous
1280
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-v5-all-resampled.yaml
outputs
report
reports/learning/tiny-geometry-v5-all-resampled
derived
data/processed/tiny-geometry-v5-all-resampled
run_record
runs/tiny-geometry-v5-44ce3de-a34456a8-32a80ab5
artifact_durability
compact-report-pending-local-git;checkpoint-local-only
result
run_hypothesis_passed
false
gate_e_exit_evidence_complete_cumulatively
true
artifact_identity_rerun
true
checkpoint_resume_exact
true
one_icon
fixed_probe_accuracy
0.977328431372549
heldout_accuracy
0.9926470588235294
changed_accuracy
0.9836956521739131
retained_accuracy
0.9972222222222222
diverse_four
fixed_probe_accuracy
0.985191637630662
heldout_accuracy
0.9856271777003485
changed_accuracy
0.9664634146341463
retained_accuracy
0.9962737127371274