Connector compatibility requires sequence, width, ownership, and mask contracts

A simple linear projector maps every visual feature from width d_v to language width d_t. Applied independently to each position, it keeps the visual sequence length unchanged. A learned-query module instead can produce a specified number of outputs by attending over the visual sequence, so it changes the length interface as well as the representation. Nonlinear projectors add further choices. Record the actual input and output shapes and feature-selection rule, including whether a class token is retained, before budgeting language positions or loading a connector checkpoint.

The original BLIP-2 experiments use 32 learned queries of width 768, but those are paper-specific architecture choices rather than universal constants. Likewise, a Flamingo cross-attention bridge is not interchangeable with a prefix projector merely because both connect frozen components. Matching widths does not establish a shared learned feature space or compatible training. Freeze policies, position encodings, normalization, and connector weights belong to the same validated artifact set. Test known media with its intended language interface rather than relying on successful tensor allocation as evidence of semantic compatibility.

With text queries and visual keys and values, attention scores have one axis for text positions and another for visual positions, separately for each example and head. Query and key head widths must match, while the value width determines the per-head output. The additive mask must broadcast to the intended axes and block padding or disallowed media. A mask with the right total number of elements but the wrong axis interpretation can silently expose another image or prevent the desired one from being read.

An all-masked query row has no valid probability distribution under the usual negative-infinity additive mask. Choose an explicit valid null-input path, skip rule, or rejection contract rather than accepting a NaN or treating arbitrary zeros as meaningful visual evidence. Interleaved architectures may restrict a text token to particular preceding media; a text causal mask alone does not encode that ownership. Test padding, multiple examples, and media-order changes with distinct content so a cross-example or cross-media leak becomes observable. Verify the actual library's Boolean and additive-mask conventions separately.

Build tiny examples with different media per batch row, padded visual positions, and a known allowed media boundary. Assert shape invariants and perturb only a blocked input to test whether output changes unexpectedly. Such tests diagnose interface leakage but do not prove all learned reasoning is grounded. For memory analysis, distinguish a hypothetical materialized score matrix from fused or tiled implementations that avoid storing all scores. Measure actual peak allocation, kernel behavior, and output equivalence before using a shape formula to predict a deployment's capacity or speed.

The matrix lab shows one fixed two-dimensional linear family, not a connector implementation. Reset, set both a=1 and b=0, predict determinant 1 and condition number 1, then Play. Reset and set both a=1 and b=1 for determinant 1 and condition number(3+√5)/2≈2.618. The selected transform is[[a,b],[0,1]], and readouts describe that final matrix while geometry interpolates. It has no batch, token length, learned weights, attention, media ownership, or masks. Use the analogy only to separate a linear transform from the additional interface semantics required in the real bridge.