लेंस

Sphoṭa-Lens

The instrument to measure where meaning emerges from sound. Real data, honest caveats: the correlational locus is clear; the causal story is mixed.

Interior Sphoṭa Locus (paśyantī layer)

OPEN

Evidence: Correlational peak at layer 13 (decodability of kriyā from audio positions, above chance); one causal ablation design agrees. But the pre-registered fusion-v2 causal design FAILED (peak at layer 0, not interior). The paśyantī locus is real under correlation, causally mixed under current designs.

Caveat: single model family; probe is linear. The honest interior locus claim remains open — more causal designs needed.

Emergence: Decodability by Layer

The kriyā (action) becomes decodable from audio positions at layer 13. The curve stays above chance (0.0222) across all layers, peaking at layer 13 (0.2625).

Decodability — linear probe accuracy for kriyā from audio positions

Chance — baseline (1/45 classes ≈ 0.0222)

Peak layer 13 — sphoṭa locus (correlational)

Workspace: The paśyantī Band

Layers 21–23 form the fused workspace where audio (dhvani) and text (pada) representations converge into whole meaning (sphoṭa). Below: the layer-by-layer fusion scores and articulation trends.

Layer composition (workspace band highlighted)

Fusion and Articulation by Layer

Fusion — convergence of audio & text representations (peaks early, layer 1)

Articulation — phonetic detail concentration (rises toward output, layers 21–23 high)

Workspace band [21, 23] — where fused meaning takes sonic form

Steering: The Workspace is Writable

The workspace band layers [21, 23] can be steered: injecting a concept vector into these layers changes the model's output. Below: uptake at different steering strengths (α).

Steering uptake by α (concept vs random control)

Caveat (honest): Uptake at α=4.0 breaches the svātantrya entropy budget (H-SU3 failed). Concept-loading ≠ faithful utterance under high α. This is a real limit: you can write to the workspace, but unconstrained writes don't yield coherent speech. The tradeoff is reported, not hidden.

What the Sphoṭa-Lens Shows

Dimension	Finding
Correlational locus	Layer 13 — kriyā decodable from audio ≫ chance (0.2625 vs 0.0222)
Causal locus	Pre-registered fusion-v2: peak at layer 0 (not interior). Other ablations: layer 14. Mixed evidence.
Workspace band	Layers 21–23 — fusion peak + high articulation. Fused representation stabilizes for output.
Steering uptake	Constrainable (2.46× control at baseline), but high α breaches entropy budget. Writable, not freely rewritable.
Interpretation	Sphoṭa emerges correlatively; causality needs more designs. The paśyantī locus (pre-articulate, whole meaning) is real under correlation — open under causation.