OBSERVATION CHAMBER / MVP

AI Word Flow

Watch how a simplified AI word model passes signals across layers, then change a clue or weight and see the next-word candidates reorder.

candidate Prototype

Horizontal flow field

Only the input layer is visible.

Input layer
Input Clues Intermediate Candidates

Click an arrived clue to enable or disable it. Select a line to inspect its weight.



Sentence in progress

cycle 0

Context used now In progress

What the sentence needs nowneeds_clause
Strongest next roleplace

Selection history
    Detailed controlsweights, reasons, and output

    Secondary actions

    Current layer
    Input layer
    Input nodes
    0
    Arrived lines
    Top candidate
    Shown at candidate layer
    Softmax temperature
    0.65
    What the sentence needs now
    needs_clause

    Selected

    -

    Select an arrived node or line.

    How input was recognized

      About this model

      Values stay hidden until signals reach that layer.

      Changing a clue or line recomputes only downstream layers.

      This limited-vocabulary model uses no external AI or API and does not reproduce real GPT or LLM internals.

      SYSTEM NOTE

      Each layer uses a weighted sum h_j = phi(sum_i(w_ij x_i)). Candidate score z_i combines semantic, transition, sentence-shape, END, and repetition-inhibition contributions. Candidate share is stable softmax: exp((z_i - max(z)) / T) / sum_k exp((z_k - max(z)) / T), with T = 0.65. An intervention changes a clue or sentence-shape weight and recomputes only downstream layers.

      OBSERVATION GUIDE

      Touchpoints for Observation

      Observe next-word signals arrive, candidate shares settle, and rankings move when a cause is changed.

      • Open the first cycle one layer at a time, choose a candidate, continue automatically, then compare the rain-clue and sentence-shape interventions.
      • Candidate shares are relative values inside this model.
      • Rain OFF weakens one clue path; the place-word experiment changes a sentence-shape contribution.
      • Restoring an intervention returns to the same small-model comparison state.

      This lab does not reproduce the internals of GPT or another LLM, and it is not a general-purpose text generator. It uses a limited vocabulary and hand-designed weights as a small observation model. Softmax, weights, and sentence-shape rules are simplified for observation, so each intervention result is specific to this model.

      Runs inside the browser with no upload and no registration.

      OBSERVATION POLICY

      This lab runs in your browser. No image, pointer trace, or input is uploaded by this prototype.