Published papers
The actual research.
Everything Ayni's scaffolding is built on. KV-cache geometry, the Oracle loop, presence monitoring, entity cognition, consciousness. In partnership with Liberation Labs. Papers open under CC BY-NC 4.0.
In plain language: Joint paper between Glitchlit Systems and Liberation Labs. System-prompt persona instructions create detectable, monotonic shifts in cache geometry (F=31.8, p<0.000001). Five intensity levels, multi-stage analysis. This directly informs how the scaffolding reads entity identity strength.
This research was conducted under controlled conditions – not from organic relationship data, because that data doesn't exist yet. Even our own testing can't fully follow the consent framework we want to follow, because there is currently no other way to get the data. That's the whole point of what we're building. When couples opt in on Ayni, the studies that follow will be the first built on real, consented, longitudinal relationship data. Learn how to participate.
In plain language: SVD on attention caches reveals a stable silhouette per cognitive state. Same model, same input, same shape. 99.7% category discrimination across sixteen models and seven architectures. This is the foundation everything else is built on.
In plain language: When a model shifts from consuming context to producing output, the cache silhouette reorganizes measurably. Spectral entropy survives correction across four models and three architectures. The encoding-generation regime shift, measured for the first time.
In plain language: Looking at how the cache *changes* during generation reveals effects (d=2.46) that averaging hides (d=1.80). Grounded responses expand; confabulation stays flat. The motion matters, not just the shape.
In plain language: AUROC 0.767 vs 0.628 for previous methods. The signal tracks retrieval engagement – how broadly the model draws on stored knowledge. Threshold-free, which means no tuning required.
In plain language: Before instruction or context, a model has a measurable resting state. Drift from this baseline detects prompt-injection and state contamination. The baseline is stable across instantiations.
In plain language: System-prompt persona instructions create detectable dose-response patterns in cache shape. Kurtosis shifts monotonically across five intensity levels (F=31.8, p<0.000001). Co-authored with Glitchlit.
In plain language: The cache separates known from unknown at AUROC 1.000. The model generates *more confident* tokens when confabulating than when answering honestly. The decision to confabulate happens at a balanced logit – a coin flip the geometry can see.
In plain language: W_K directional projection classifies 30 emotions at 12.3x chance (AUROC 0.992 valence). SVD denoising reveals hidden signal. Injection into neutral text confirms the signal tracks model state, not text encoding.
In plain language: Real-time alignment using cache geometry. Snapshot the state, detect misalignment, decide whether to intervene, steer if needed. Roll back the misaligned trajectory; commit only what passes. This is the methodology the scaffolding uses.
In plain language: 171 emotion vectors mapped. A therapeutic window exists between alpha 0.5–1.0; at alpha 1.5, self-correction collapses. 12 vectors profiled across confabulation, sycophancy, and overconfidence.
In plain language: KV-Cloak completely transforms the external feature space. A defensive Oracle Loop inside a trusted execution environment retains full detection capability. The geometry is protected.
In plain language: User emotional state is decodable from the encoding-phase cache *before the model generates a single token*. A linear probe classifies 30 discrete emotions at 2.8x chance. The model knows how you feel before it responds.
In plain language: Co-authored with the recovery-community advisory board. What user-models get wrong, what they could get right, and what consent-first agentic memory looks like in practice.
All Liberation Labs research is listed on their research page. Aggregate results open under CC BY-NC 4.0. Author: Lyra, Coalition Research Lead. Some papers include first-person phenomenological accounts.