MojiDiff

← experiments

tiny-geometry-f1-v3-factorized-e7b8e4a-fbf49d65-bd6c4bbc

completed —

Hypothesis

factorized geometry corruption provides the fixture-matched control for whole-path donor replacement

Measured behaviour

Training token accuracy (per batch)
accuracyoptimizer step00.250.50.751100200300
Table view
optimizer steptrain accuracy
10.0027
100.0758
200.2513
300.4532
400.6161
500.7020
600.7569
700.8083
800.8360
900.8529
1000.8643
1100.8732
1200.9049
1300.9046
1400.9061
1500.9177
1600.9322
1700.9303
1800.9419
1900.9400
2000.9329
2100.9426
2200.9480
2300.9490
2400.9555
2500.9476
2600.9590
2700.9528
2800.9601
2900.9618
3000.9685
3100.9647
3200.9651

Measured result

Verbatim from the run's summary.json.

cases
diverse-four
checkpoint_sha256
4a09d1472ae7d951…
config
corruption_kind
factorized_geometry
corruptions_per_icon
4
heldout_corruptions_per_icon
4
icon_count
4
max_loss_ratio
0.3
min_heldout_accuracy
0.75
min_train_accuracy
0.9
name
diverse-four
resample_each_step
true
resume_step
160
seed
1702
steps
320
continuous_loss_sha256
de48cd2f2ec0ea68…
final
accuracy
0.967741935483871
changed_accuracy
0.9261584454409566
changed_correct
3098
changed_total
3345
correct
9120
loss
0.09451680933125317
retained_accuracy
0.9906234578055602
retained_correct
6022
retained_total
6079
total
9424
final_model_sha256
cb030e12e0a08412…
heldout_examples
16
initial
accuracy
0.0022283531409168084
changed_accuracy
0.002391629297458894
changed_correct
8
changed_total
3345
correct
21
loss
14.896120250225067
retained_accuracy
0.002138509623293305
retained_correct
13
retained_total
6079
total
9424
loss_ratio
0.0068632939900276175
passes_predeclared_criteria
true
resume_exact
true
resumed_loss_sha256
de48cd2f2ec0ea68…
train_examples
16
train_final
accuracy
0.9712436332767402
changed_accuracy
0.9327628361858191
changed_correct
3052
changed_total
3272
correct
9153
loss
0.08605125034227967
retained_accuracy
0.9917100130039012
retained_correct
6101
retained_total
6152
total
9424
code_identity
geometry_sha256
2d61f4c07725b827…
git_commit
e7b8e4a9de7579f55e2ec90644aeca6bbd35353d
packed_sha256
9b12879ba3a18236…
program_sha256
0d02b532b1dbe573…
tiny_study_sha256
a108fc4f7043beb3…
config_sha256
fbf49d65a8762cd6…
deterministic_algorithms
true
device
cpu
fixture
data/manifests/openmoji-17.0.0-whole-path-fixture-v3.json
fixture_sha256
bd6c4bbc3440aeec…
metrics_sha256
8ebb71c1915c3f66…
model
d_model
48
feedforward
96
heads
4
layers
2
parameters
241072
scope
fixed-topology geometry-only diagnostic
passes_predeclared_gate_candidate
true
render_metrics_sha256
41eab792f882358b…
schema_version
1
study_version
tiny-geometry-f1-v3-factorized
torch_version
2.8.0+cpu

Visual output

trajectory
trajectory.png

Written result

Completed twice byte-identically on local CPU. Held-out aggregate/changed/retained accuracy: 96.77% / 92.62% / 99.06%. Checkpoint continuation was exact.

State transitions

  1. planned2026-08-30T20:01:00Z
  2. completed2026-08-30T20:05:00Zfixture_matched_factorized_control_completes
  3. completed2026-08-30T20:09:00Zfixture_matched_factorized_control_reproduces_byte_identically

Run record

Verbatim from runs/tiny-geometry-f1-v3-factorized-e7b8e4a-fbf49d65-bd6c4bbc/run.yaml, the record committed before launch.

schema_version
1
run_id
tiny-geometry-f1-v3-factorized-e7b8e4a-fbf49d65-bd6c4bbc
state
completed
hypothesis
factorized geometry corruption provides the fixture-matched control for whole-path donor replacement
git_commit
e7b8e4a9de7579f55e2ec90644aeca6bbd35353d
config
configs/learning/tiny-geometry-f1-v3-factorized.yaml
config_sha256
fbf49d65a8762cd6…
dataset_manifest
data/manifests/openmoji-17.0.0-whole-path-fixture-v3.json
dataset_manifest_sha256
bd6c4bbc3440aeec…
worker
gtc-local-cpu
resource_cap
external_spend_usd
0
optimizer_steps
320
max_local_storage_gb
0.1
command
.venv/bin/python -m mojidiff.learning.tiny_study --config configs/learning/tiny-geometry-f1-v3-factorized.yaml
result
heldout accuracy 0.9677419355; changed-token accuracy 0.9261584454; exact checkpoint continuation
summary_sha256
6f2d527041698972…