openmoji-g1-detectability-877feff-4processes-9b9b1699

completed —

Hypothesis

Descriptive, and a check on the conclusion the previous run drew. Gate G was closed on the reading that factorized role-uniform corruption is not identifiable, inferred from a trained detector reaching 0.1058 recall at 1.23x precision lift. The competing reading is that the corruption is identifiable and the model failed to learn it. A parameter-free local-continuity statistic distinguishes the two without training.

Scope limits

48 held-out icons, one corruption draw per icon per process, one fixed statistic. The statistic is a floor on detectability, not a ceiling: it uses no learned parameters, so a result above chance proves identifiability while a result at chance would not prove the converse.

State transitions

completed2026-09-20T21:45:00Zcorruption_is_identifiable_after_all_a_zero_parameter_heuristic_beats_the_trained_detector

Run record

Verbatim from runs/openmoji-g1-detectability-877feff-4processes-9b9b1699/run.yaml, the record committed before launch.

schema_version

run_id

openmoji-g1-detectability-877feff-4processes-9b9b1699

state

completed

parent_run

openmoji-g1-edit-mask-v6-0228c3b-0f6ce115-9b9b1699

recorded_retroactively

true

hypothesis

Descriptive, and a check on the conclusion the previous run drew. Gate G was closed on the reading that factorized role-uniform corruption is not identifiable, inferred from a trained detector reaching 0.1058 recall at 1.23x precision lift. The competing reading is that the corruption is identifiable and the model failed to learn it. A parameter-free local-continuity statistic distinguishes the two without training.

scope_limits

code

git_commit

877feffaf4f16d659475f40c74a14b953bac434b

execution_mode

native-local-cpu

probe

src/mojidiff/learning/detectability.py

execution

icons

statistic

mean absolute distance to the same slot in adjacent segments, view units

network

none

trains_nothing

true

result

auc_by_process

path_correlated_p0.35

0.9333

factorized_p0.10

0.8575

factorized_p0.35

0.7698

whole_path_p0.35

0.5318

whole_path_caveat

Only 32 of 13,128 fields were replaced, a 0.24% rate, because donor compatibility requires a matching length and segment-kind sequence. Its AUC is not comparable with the others at this sample size, and a donor path is locally smooth by construction, so local continuity is the wrong instrument for it.

matched_operating_point

note

both detectors flag 8.573% of fields, the rate the trained keep head chose

continuity_statistic

precision

0.7451

lift

2.13

recall

0.1824

trained_keep_head_v6

precision

0.4312

lift

1.23

recall

0.1058

conclusion

The Gate G conclusion was wrong on its decisive point. Factorized corruption at p=0.35 IS identifiable: a zero-parameter heuristic reaches AUC 0.7698 and, at the trained detector's own flag rate, 0.7451 precision against its 0.4312 - a 73% improvement with no parameters at all. The failure is one of learning, not of information. Every measurement in v1 through v6 stands; the interpretation of the final step does not.

bears_on_gate_f

Path-correlated corruption is markedly more identifiable than factorized, 0.9333 against 0.7698, which is what a structural argument predicts: a replaced block is inconsistent with its neighbours in a way a single resampled coordinate is not. Gate F chose factorized as primary on four-icon fixtures where recovery was memorization. This is independent evidence that the choice should be revisited.

outputs

local_metadata

runs/openmoji-g1-detectability-877feff-4processes-9b9b1699

report

reports/learning/corruption-detectability.json

completed_at

2026-09-20 21:45:00+00:00