openmoji-g1-corruption-process-corpus-76f41a3-2arms-9b9b1699

cancelled declared

Hypothesis

Gate F selected factorized role-uniform corruption as the primary process from four-icon fixtures where held-out recovery ran at 96-99%, which the Gate G sequence has since shown to be memorization rather than denoising. On the 2,681-icon split, path-correlated corruption defines a materially more learnable problem than factorized, because its corrupted fields are more identifiable: a training-free local-continuity statistic separates them at AUC 0.9333 against factorized's 0.7698. A model trained on path-correlated corruption should therefore learn a corrupted-field detector that beats the trivial statistic, where the factorized model did not.

Why run it

This re-runs a selection the project made on evidence that no longer supports it, at corpus scale and with every improvement the Gate G sequence produced. It either identifies a corruption process on which this model class works, or shows that none of them does - in which case the model class, not the task, is the problem and PROJECT_PLAN.md section 12's fallback interpretations become the live branch.

Scope limits

Two arms, not three. Whole-path replacement is excluded because it needs a donor pool and a compatible-coverage audit: under single-donor compatibility the detectability probe replaced only 32 of 13,128 held-out fields, a 0.24% rate, so including it as configured would be a null treatment rather than a comparison. Its Gate F fixture already measured only 24.74% of legal fields as having an exact compatible donor. It is deferred with that reason recorded, not dropped. Fixed-topology and geometry-only, single seed, one corruption draw per icon.

State transitions

planned2026-09-20T22:05:00Z
cancelled2026-09-20T22:46:00Zpremature_no_trained_model_can_yet_learn_detection_on_any_process

Run record

Verbatim from runs/openmoji-g1-corruption-process-corpus-76f41a3-2arms-9b9b1699/run.yaml, the record committed before launch.

schema_version

run_id

openmoji-g1-corruption-process-corpus-76f41a3-2arms-9b9b1699

state

cancelled

parent_runs

openmoji-g1-edit-mask-v6-0228c3b-0f6ce115-9b9b1699, openmoji-g1-detectability-877feff-4processes-9b9b1699

supersedes_selection_from

tiny-geometry-f1-v3-seed2701-factorized-5eca9c1-e252438f-bd6c4bbc

hypothesis

Gate F selected factorized role-uniform corruption as the primary process from four-icon fixtures where held-out recovery ran at 96-99%, which the Gate G sequence has since shown to be memorization rather than denoising. On the 2,681-icon split, path-correlated corruption defines a materially more learnable problem than factorized, because its corrupted fields are more identifiable: a training-free local-continuity statistic separates them at AUC 0.9333 against factorized's 0.7698. A model trained on path-correlated corruption should therefore learn a corrupted-field detector that beats the trivial statistic, where the factorized model did not.

expected_information_gain

scope_limits

arms

factorized

config

configs/learning/openmoji-g1-process-factorized.yaml

sha256

7bfd76b675ed7b91…

role

control, and the process Gate F selected

path_correlated

config

configs/learning/openmoji-g1-process-path-correlated.yaml

sha256

1ca993912b923805…

role

treatment

matched

The two configs differ only in training.corruption_process. Both carry the v5 slot binding and the v6 edit mask, the full 2,681-icon split, the 128-icon validation draw and its corruption seeds, seed 3101, corruption probability 0.35, batch size 16, learning rate 0.001, eval_every 60, the 6,300-step cap, and the held-out-loss selection policy with patience 8.

code

git_commit

76f41a3

execution_mode

native-local

dataset

hybrid_sha256

9b9b1699677a6f97…

bucket

bucket-p32-t128

train_samples

2681

validation_samples

128

baselines

identity

note

emit x_t unchanged; a legal policy under argmax decoding

aggregate_at_p035

0.6506198760247951

render_recovery_fraction

0.0

training_free_detector

note

parameter-free local-continuity statistic, at a matched 8.573% flag rate

factorized

auc

0.7698

precision

0.7451

lift

2.13

recall

0.1824

path_correlated

auc

0.9333

v6_trained_detector_on_factorized

precision

0.4312

lift

1.23

recall

0.1058

execution

seed_set

3101

deterministic_algorithms

true

cublas_workspace_config

:4096:8

optimizer_step_cap

6300

early_stopping

objective

held_out_loss

patience_evals

min_delta

0.0

resource_cap

max_steps

20000

max_storage_gb

network

none

predeclared_criteria

note

Declared before either arm runs. Held-out aggregate accuracy is deliberately NOT the gated quantity: the identity policy scores 0.6506 and dominates every model so far, so an accuracy criterion would reward collapsing to identity. The gated quantity is whether a trained model learns a corrupted-field detector better than a zero-parameter statistic on the same data.

evaluated_at

the checkpoint selected by held-out loss, per arm

primary

path_correlated_beats_factorized_on_detection

statement

The path-correlated arm's keep-head precision lift over its own base rate exceeds the factorized arm's, both measured at a matched 8.573% flag rate.

a_trained_detector_beats_the_free_one

statement

At least one arm's keep-head precision lift exceeds the training-free statistic's lift on that same process - 2.13x for factorized, to be measured for path-correlated before the arms are compared. This is the criterion that decides whether a trained model adds anything at all.

beats_identity

statement

At least one arm's held-out aggregate accuracy exceeds 0.6506198760247951. Reported for both arms regardless.

structural_safety

statement

Locked-path exactness holds and the canonical checkpoint round-trips, both arms.

reproducibility

statement

Each arm returns an identical JSON result on a second invocation.

degenerate_pass_watch

As in v6, a model predicting keep everywhere scores exactly identity. Report the fraction of fields predicted keep for both arms; an arm that passes beats_identity while its keep fraction approaches 1.0 has collapsed to the trivial policy and must be reported as collapse.

falsification_meaning

If neither arm learns a detector beating the free statistic, the conclusion is about this model class rather than about either corruption process: a bidirectional transformer trained with this objective does not learn a signal that is present and trivially extractable. That is a Gate F and Gate G level result and makes PROJECT_PLAN.md section 12's fallback interpretations the live branch.

reported_not_gated

held-out aggregate, changed and retained accuracy for both arms, beside identity, keep fraction, detector recall and precision for both arms, the four-level render recovery sweep for both arms, beside identity's zero, selected step, wall time, peak memory

outputs

local_metadata

runs/openmoji-g1-corruption-process-corpus-76f41a3-2arms-9b9b1699

report_roots

reports/learning/openmoji-g1-process-factorized, reports/learning/openmoji-g1-process-path-correlated

durable_artifacts

/home/dev/.cache/openmoji-g1-corruption-process-corpus-76f41a3-2arms-9b9b1699

planned_at

2026-09-20 22:05:00+00:00

cancelled_at

2026-09-20 22:46:00+00:00

cancellation

deferred_by

openmoji-g1-detection-only-v7-856230a-9d2db303-9b9b1699

reason

This comparison assumes a trained model can learn corrupted-field detection on some corruption process. The detection-only diagnostic tested that assumption directly and falsified it: with the competing value objective removed entirely, the trained detector reached 1.365x precision lift against a zero-parameter statistic's 2.13x on the same data, and converged there - held-out loss selected step 660 of a 6,300 cap. Comparing processes is only meaningful once a trained model can learn the signal on at least one of them, so running the arms now would compare two numbers that both lose to a heuristic.

status

Not withdrawn. The configs, criteria and matched setup remain valid and this run record stands ready to be re-registered under a new identity once a model can learn detection at all.