DocumentationArchitecture

OutputProjection

A zoomed-in view of one operation in Piro’s stateful inference loop.

Architecture detail

OutputProjection

Deferred CTM architecture: combines specialized retrieval values into recurrent context.

OutputProjection(
    contentScoresₖ,
    timeBiasₖ,
    syncBiasₖ,
    valuesₖ,
    weights
):
    retrievalScoresₖ = contentScoresₖ + timeBiasₖ + syncBiasₖ
    retrievalWeightsₖ = softmax(retrievalScoresₖ)
    retrievedₖ = retrievalWeightsₖ · valuesₖ
    contextₖ = retrievedₖ · weights.attention.outputProjection.W
        + weights.attention.outputProjection.b
    return contextₖ

What it does

Combines content, temporal, and synchronization scores, normalizes them into retrieval weights, uses those weights to combine value vectors, and projects the result into CTM context space.

Why it exists

OutputProjection owns the complete boundary from retrieval evidence to contextₖ, so Attention does not need to carry retrievalWeightsₖ and retrievedₖ as separate intermediate outputs.

← Back to architecture