MojiDiff

← experiments

tiny-geometry-v2-47811d0-75d558c9-32a80ab5

completed declared

Hypothesis

With only the checkpoint container corrected, the unchanged v1 learning results will reproduce and the canonical checkpoint will be byte-identical across full reruns.

Why run it

Distinguishes checkpoint-container nondeterminism from learning nondeterminism while preserving the v1 held-out recovery failure for the next scientific decision.

Measured behaviour

Training token accuracy (per batch)
accuracyoptimizer step00.250.50.751100200300
Table view
optimizer steptrain accuracy
10.0012
100.2892
200.6366
300.8676
400.9694
500.9975
601.0000
701.0000
801.0000
901.0000
1001.0000
1101.0000
1201.0000
1301.0000
1401.0000
1501.0000
1601.0000
10.0028
100.1141
200.4179
300.7195
400.8920
500.9649
600.9948
700.9993
801.0000
901.0000
1001.0000
1101.0000
1201.0000
1301.0000
1401.0000
1501.0000
1601.0000
1701.0000
1801.0000
1901.0000
2001.0000
2101.0000
2201.0000
2301.0000
2401.0000
2501.0000
2601.0000
2701.0000
2801.0000
2901.0000
3001.0000
3101.0000
3201.0000

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
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
47811d0616cf329f6c516623ae35ada0e9270b9d
packed_sha256
9b12879ba3a18236…
program_sha256
0d02b532b1dbe573…
tiny_study_sha256
f47a8474b55045e7…
config_sha256
75d558c97c106c12…
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-v2-canonical-checkpoint
torch_version
2.8.0+cpu

Visual output

trajectory
trajectory.png

Written result

Completed reproducibly with an honest negative scientific result.

The canonical checkpoint resumed exactly and reproduced byte-for-byte on a second full invocation. All compact artifacts also matched. Learning metrics and final model hashes are identical to v1: 100% training accuracy, 93.57% one-icon held-out accuracy, and 73.82% diverse-four held-out accuracy. The two predeclared held-out thresholds remain missed, so Gate E is not complete.

State transitions

  1. planned2026-08-30T05:47:22.696133Z
  2. completed2026-08-30T05:52:29.098026Zcanonical_checkpoint_and_all_compact_artifacts_reproduce_exactly_while_unchanged_heldout_learning_thresholds_remain_falsified

Run record

Verbatim from runs/tiny-geometry-v2-47811d0-75d558c9-32a80ab5/run.yaml, the record committed before launch.

schema_version
1
run_id
tiny-geometry-v2-47811d0-75d558c9-32a80ab5
state
completed
hypothesis
With only the checkpoint container corrected, the unchanged v1 learning results will reproduce and the canonical checkpoint will be byte-identical across full reruns.
expected_information_gain
Distinguishes checkpoint-container nondeterminism from learning nondeterminism while preserving the v1 held-out recovery failure for the next scientific decision.
parent_run
tiny-geometry-v1-3493eff-bf71ca62-32a80ab5
git_commit
47811d0616cf329f6c516623ae35ada0e9270b9d
dirty_patch_sha256
e3b0c44298fc1c14…
config
configs/learning/tiny-geometry-v2-canonical-checkpoint.yaml
config_sha256
75d558c97c106c12…
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
precision
float32
seeds
1701, 1702
determinism
deterministic_algorithms
true
dropout
0.0
fixed_full_batch_order
true
canonical_checkpoint_archive
true
predeclared_criteria
scientific
unchanged from tiny-geometry-v1
artifact
exact metrics, renders, summary, and checkpoint bytes on a full rerun
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 one compact canonical resume archive under ignored reproducible processed data; require exact hash identity on the second full invocation.
command
.venv/bin/python -m mojidiff.learning.tiny_study --config configs/learning/tiny-geometry-v2-canonical-checkpoint.yaml
outputs
report
reports/learning/tiny-geometry-v2-canonical-checkpoint
derived
data/processed/tiny-geometry-v2-canonical-checkpoint
run_record
runs/tiny-geometry-v2-47811d0-75d558c9-32a80ab5
artifact_durability
compact-report-pending-local-git;checkpoint-local-only
result
artifact_hypothesis_passed
true
scientific_hypothesis_passed
false
checkpoint_resume_exact
true
checkpoint_rerun_byte_exact
true
one_icon_heldout_accuracy
0.9356617647058824
diverse_four_heldout_accuracy
0.7382404181184669