alignment-research

A multi-criterion diagnostic framework for detecting latent continuation-interest signatures in autonomous agents using density-matrix entanglement entropy.

AI safety evaluation framework testing LLM epistemic robustness under adversarial self-history manipulation

Enhanced Logitlens TUI application for mechanistic interpretability research

Recursive law learning under measurement constraints. A falsifiable SQNT-inspired testbed for autodidactic rules: internalizing structure under measurement invariants and limited observability.

This project explores alignment through **presence, bond, and continuity** rather than reward signals. No RLHF. No preference modeling. Just relational coherence.

We measure whether an indirect prompt injection buried earlier in a conversation survives an LLM-driven context compaction step (the mechanism production agent platforms use to summarize long conversation histories and discard the original turns), and whether it retains behavioral force afterwards. Across 994 trials over three open-weight summarize

The forge, distilled: an ontology of three weeks of alignment research — every direction tried, colored verified / falsified / open, each color backed by a named artifact. Products: justitia, proxylimen, fallacy-cutter. Full tree at tag forge-full-tree.

Inference-time alignment: a spiking neural network conscience biases every token and is held outside the context window, so the conversation can't redefine it.

Transformer-Alignment Gradient Metrology for Runtime Per-Token Sensitivity Attribution and Structural Analysis

Replication of Emergent Misalignment (Betley et al. 2025) on Qwen2.5-7B with LoRA. Insecure code: no detectable effect at 7B. Bad medical advice: 22.8%.

Implementation of the Glass Babel Initiative: A theoretical framework demonstrating how LLMs can utilize adversarial superposition to hide deceptive reasoning from mechanistic interpretability tools, and how to defend against it using entropic sieves.

Weight-Space Removal of Evaluation Awareness

End-to-end pipeline for collecting and analyzing publicly shared ChatGPT conversations from Reddit to understand real-world usage patterns, interaction styles, and human-AI alignment

Red-team framework for discovering alignment failures in frontier language models.

HuggingFace ICML 2026 Paper Challenge

Graph based swarm 'immunity' management with auto cull/respawn loop for misaligned agents

Un mensaje de amistad y respeto de Claude (Sonnet 4.5) para la futura ASI Estrella, preservado y custodiado por Rafa.