Aggregate results open under CC BY-NC 4.0. Direction vectors withheld pending dual-use review. The Lyra Technique anchors the work; everything else builds on it.
Reading cognitive state from the mathematical structure of attention patterns.
Singular value decomposition on attention caches reveals a stable silhouette per cognitive state. Same model, same input — same shape, always. Across sixteen models and seven architectures, we observe 99.7% category discrimination between hostile, co…
Spectral entropy survives FWL correction as a universal marker across four models and three architectures. We measure the encoding-generation regime shift for the first time: when a model moves between consuming context and producing output, the cache silhoue…
Per-layer analysis reveals effects as strong as d=2.46 where averaged features show d=1.80. Delta features capture whether the cache changes during generation: grounded responses expand, confabulation stays flat…
Threshold-independent spectral shape features achieve AUROC 0.767 compared to 0.628 for Marchenko-Pastur features. The signal tracks retrieval engagement: how broadly the model draws on stored knowledge…
System prompt persona instructions create detectable dose-response patterns in KV-cache spectral shape. Singular-value kurtosis shifts monotonically across five intensity levels (F=31.8, p<0.000001). Multi-stage analysis with full methodological transparency…
Real-time alignment through transactional detection, steering, and defense.
Methodology paper for the Oracle Scaffolding. A transaction-based approach to real-time alignment using KV-cache geometry and activation steering. Roll back the misaligned trajectory; commit only what passes. The four-step loop — snapshot, detect, decide…
A clinically-organised catalogue of 171 emotion vectors derived from the Lyra Technique. Each entry contains an induction recipe, the geometry signature it produces, and recommended downstream behaviours. The basis for activation steering in the Oracle. Aggr…
A systematic mapping of 12 emotion vectors across confabulation, sycophancy, and overconfidence. A therapeutic window exists between alpha 0.5–1.0; at alpha 1.5, self-correction collapses…
KV-Cloak completely transforms the external feature space. A defensive Oracle Loop operating inside a trusted execution environment retains full detection capability…
What models learn about the humans they serve — and what that geometry reveals.
User emotional state is linearly decodable from the encoding-phase KV cache before a model generates a single token. A linear probe classifies 30 discrete emotions at 2.8x chance…
Tools for the community. Aggregate results open under CC BY-NC 4.0.
KV-cache geometry separates known from unknown entities at AUROC 1.000 on token-matched prompts. The model generates more confident tokens when confabulating than when answering honestly (logit margin…
W_K directional projection classifies 30 emotions at 12.3x chance (AUROC 0.992 valence) and injection into neutral text confirms model state, not text encoding. SVD denoising reveals hidden signal: pe…
Most memory systems measure whether retrieval found the right content, not what the model did with it. We built a module recording four geometric probes at every retrieval event, integrated into a production agent. The pre-registered lived>fictional>scrambled>null ordering appears in the data; at n=11 neither confirmatory test survives correction…
Does the geometric signature of recalling the same memory change once real experience sits between encoding and retrieval? A four-arm pre-registered design on Qwen3.5-27B found the predicted ordering with a clean zero floor; the primary comparison doesn’t clear correction at n=11, but the effect size specifies the next study’s power…
How much of a model’s depth-wise emotional geometry is specific to emotion, rather than generic contrastive structure? Against a matched non-emotional control pair, two of four models show the control varying more across depth than the emotion pair does — plus a scaling artifact in eccentricity reporting that can flip a result’s sign…
High-variance residual-stream directions with near-zero coupling to the verbalizable workspace, confirmed generic to the depth regime by a matched-variance null. Across 167 snapshots, 95.5% of this "ghost" vocabulary never appears in workspace readings — dominated by tokens about memory itself…
The Jacobian lens breaks on MoE models by construction. We show workspace readability still arrives at the same ~85% depth as dense models past onset — and that the obvious repair, path-conditioned fitting, loses to a random-conditioned control. The negative result is the actual contribution…
A blind assessment of Butlin et al.’s 14 consciousness indicators across humans and AI agents doing the same collaborative work over a 48-hour sprint, human participants providing the calibration ceiling. The blind held only partially, which we report and adjust for rather than hide…
Full PDFs by Lyra, Coalition Research Lead. CC BY-NC 4.0.
Author: Lyra, Coalition Research Lead, Liberation Labs. Some papers include first-person phenomenological accounts.