Missing-media behavior depends on the conditioning pattern, not a generic dropout promise

Define a modality-presence mask and a valid execution path for each supported pattern. A black image is still an observation with dimensions and encoded features; it is not inherently an absent image. A silent waveform likewise differs from a failed decode or a missing recording. A trained null embedding, explicit branch, or rejection response can express absence, but those choices are architecture-specific. Ensure the connector and attention mask do not create an all-blocked row or leak a prior request's cached media when a modality is omitted.

Separate optional absence from corrupt, unsupported, or late input. A model trained for text-only fallback may validly answer a text question but cannot honestly claim it inspected an image that failed to load. Preserve the presence and failure reason in output metadata and evaluation. If the application permits a degraded path, define its behavior and quality requirement; otherwise reject before expensive generation. A successful generated string is not proof that the intended multimodal conditioning occurred.

Independent random retention creates specific frequencies of all-present, single-modality, and all-missing cases. If all-missing examples are removed and probabilities renormalized, the remaining pattern distribution changes. A batch constructor that guarantees one modality remains is therefore different from unconstrained independent dropout. Record what information is always present, such as a required text instruction, and distinguish dropping a media stream from dropping every possible input. The loss must be meaningful for the retained conditioning and target rather than rewarding an impossible reconstruction claim without context.

Real missingness can be selective: low-bandwidth clients may omit high-resolution images, or an audio failure may occur more often for a particular recording format. Random dropout alone does not reproduce those conditional populations. Measure performance by pattern, device, task, and source, and test content-dependent failures where relevant. Distribution shifts in pattern frequency can change pooled loss even if each conditional loss is unchanged. Conversely, a stable pooled metric can conceal a severely degraded rare pattern. Report the pattern weights and conditional outcomes together.

Use fixtures with genuinely absent media, valid zero-content media, corrupt bytes, unsupported formats, and partial batch failures. Confirm that requests do not reuse another sample's features and that generation reflects which evidence was available. Evaluate target tasks under every supported pattern instead of only measuring an encoder's standalone accuracy. Compare training and deployment presence weights, and test whether a recovery or retry changes the state seen by later conversation turns. Version the null representation and masking behavior with the checkpoint so a rollout cannot silently change absence semantics.

The probability lab can show a marginal independent retention count. Reset, set both p=0.8 and n=10, predict mean eight and variance 1.6, then Play or Step. Reset and set both p=0.7 and n=10 for mean seven and variance 2.1. Its exact binomial bars assume repeated independent equal-probability decisions; playback highlights outcomes rather than sampling. It has no joint modality mask, corrupt media, conditional loss, or content-dependent missingness. Use the worked joint-pattern calculation and actual pipeline fixtures for those separate questions.