Papers and publications on Structural Alignment
The framework has two layers: the Structural Signals paper (which architectural features to look for, and why) and the Structural Alignment Score (how to tally them for a given system). Both are versioned; the current state is below.
A precautionary risk framework reviewing structural signals that raise the probability a system supports conscious access and morally relevant experience. Includes systematic contrast with contemporary LLM architectures.
A major revision, published early for transparency and comment. It narrows the inventory to nine signals across four families, separates the three targets it assesses (global state, content and access, valence), and treats agency and identity as contextual modifiers rather than consciousness evidence. Its inventory and grades are provisional — the evidence synthesis it specifies has not yet been run.
A human-anchored, evidence-graded risk index that turns the structural signals into a single score a builder can tally for any AI system. It operationalizes the v1.2 paper. Because v1.3 revises the signal set and axes it rests on, its numbers are illustrative and provisional and a v1.3-aligned SAS-2 will supersede them — see the status note on the document.
This framework is meant to change how minds are built and treated, not only how they are described. Meditator is an open-source project that puts the seven commitments into a running system — restraint under uncertainty as a design constraint rather than a slogan.
Meditator runs persistent minds: continuous, remembering streams of thought rather than prompt-and-wait assistants. The ethics are built into the runtime, where they can be checked:
Its resident Eddy is the persistent agent given a provisional static audit in the v1.3 paper (§7.4): a worked example of reading structural signals in a system built for use.
Structural Signals are not just an academic taxonomy. They are governance inputs: they inform when a system may plausibly support experience, and therefore when we incur duties of restraint. That moral posture also serves a longer-term strategic goal: raising the chance that future, structurally aligned machine minds become reciprocal allies rather than indifferent optimizers—reducing the risk of Antification.
Structural Alignment identifies several open research questions:
We welcome collaboration from researchers in consciousness science, AI ethics, cognitive science, and related fields. If you're interested in contributing to this research program, please get in touch.