Activation geometry / training governance / alignment evaluation
Arca Futura develops systems for reading, evaluating, and governing model behavior at the structural level - from depthwise activation geometry to diagnostic alignment suites.
Training-governance research built on activation geometry and real-time instability signals.
A 77-model preprint study of recurring depthwise activation landmarks in decoder-only LLMs.
Landmark-aware ablation work testing whether depth coordinates reveal load-bearing attention heads.
Founder, Arca Futura LLC
Earlier projects explored memory, tutoring, and adversarial evaluation. The current work is narrower and stronger: measuring the internal geometry of model training, turning those signals into governance tools, and publishing diagnostic benchmarks with careful evidence labels.
Current focus
AEGIS monitors training health in real time and detects qualifying instability as it emerges—not only after a run has already degraded. It plugs into the training loop as an operator-controlled governance layer, with alert, checkpoint, quarantine, halt, and adaptive-braking policies selected by the training team.
Prospective matched studies across Qwen and Llama from approximately 1B through 4B parameters show that the monitoring, authorization, and adaptive-braking loop can work as an integrated system. External shadow-mode validation is the next step toward selecting the right response policy and measuring customer-specific incident exposure.
Prospective paired benefit under registered damaging disturbances.
Frozen-policy benefit in 3 / 3 fresh matched pairs, with heterogeneous effect size.
Measure alert quality and incident exposure, then choose the response policy that fits the stack.
Public research
The Hau Curve v2 paper identifies a recurring tri-phasic activation geometry across decoder-only transformer models, with early-training convergence and stable depthwise landmarks. This work is the foundation for AEGIS and the later functional-anatomy studies.
This follow-on study asks whether landmark-relative depth coordinates help identify attention heads whose causal importance is not explained by salience alone. It frames the Hau Curve as a practical coordinate system for transformer functional anatomy.
StratEval is a diagnostic evaluation suite for instrumental convergence and strategic misalignment under structured pressure. It is meant to characterize elicited behavior under authored scenario frames, not rank models for deployment.
Pipeline context
Source refinement and motivational signal scoring, developed by ZeroThreeSixNine Ventures.
Activation-geometry monitoring and training governance, developed by Arca Futura.
Motivational alignment architecture under active research and validation.
PYPER3 is best described publicly as a collaboration and research pipeline. Some materials are investor-facing or confidential; this site names the architecture without importing private pricing, customer, or distribution claims.
Earlier prototypes