Rixven

The Sponor Architecture

A model that remembers by resonating, not by storing.

Every long-context transformer pays the same bill: a KV cache that grows with everything it has ever seen. Sponor's central bet is that this is not a necessary cost of memory — it's a consequence of one specific way of implementing memory. We are building the other way.

A concentric resonance field diagram, with one ring of the field highlighted in ember orange — an illustration of the spectral structure PSSR is built around. loading=eager fetchpriority=auto />

The components

Six pieces that do the core work, in the order signal moves through them.

  1. 01

    PSSR — Persistent Spectral State Resonance

    The resonance field itself — the mechanism that replaces attention and the KV cache. State persists as phase, not as stored tokens.

  2. 02

    FEAC — Unified Free-Energy Gate

    A single gating principle governing what the field retains, rather than a separate mechanism per layer.

  3. 03

    LSH-DCT + PCM — Sparse Local Attention + Phase Coherence

    Where local, token-to-token precision still matters, applied sparsely and reconciled against the field by phase.

  4. 04

    HRS — Zero-Attention Long Context

    How long-range dependency is carried without any attention operation at all, at any range.

  5. 05

    V-ROP — Output Projection

    Reads the field back out into token-space predictions.

  6. 06

    Sparse MoE FFN — Mixture-of-Experts Feed-Forward

    Capacity scaling that stays sub-linear in active compute.

Supporting mechanisms

ETD-RoPE
Position encoding adapted for a field rather than a sequence.
PILD
Verlet-integrated residual connections — physically motivated stability.
NCC
Normalisation across the field’s coherent components.
IB-Field
Information-bottleneck shaping applied to the field itself.

The Effective Context Bound

Long-context claims in the literature are almost always empirical: train the model, run a retrieval test, report the number that came out. The Effective Context Bound (ECB) is our attempt to derive that number instead of measuring it after the fact.

If context is carried by a resonance field, the field has a spectrum, and the spectral gap between its lowest modes sets an analytic bound on how far a signal can propagate before it is indistinguishable from noise. That bound is computable from the field's own parameters, before a single training step runs.

Whether a tight analytic bound is a useful one is a separate, open question — an elegant bound on a weak field doesn't help anyone. Read the current thinking on this in The Lab .

Roadmap — eleven phases

We are in Phase 1. This list will move slowly, on purpose, and we'll update it here rather than promise a date we don't know.

  1. 00 Theory Done
  2. 01 PSSR standalone implementation In progress
  3. 02 FEAC gating integration Not started
  4. 03 Sparse local attention layer Not started
  5. 04 Long-context (HRS) validation Not started
  6. 05 Output projection + decoding Not started
  7. 06 MoE FFN integration Not started
  8. 07 Small-scale end-to-end training Not started
  9. 08 ECB empirical validation Not started
  10. 09 Scaling study Not started
  11. 10 Publication Not started

What we don't know yet

This could fail at any of several points: the field could prove numerically unstable at scale, the ECB could turn out to bound something that doesn't matter in practice, or a sparse local attention layer we still need for precision could reintroduce the costs we're trying to remove. We would rather list these plainly than round them off in the framing.

Follow the actual progress, including the dead ends, in The Lab.