Research
Independent research across philosophy of mind, AI safety, governance, and archaeology. The attending concept runs through all four papers — a single idea with explanatory reach across multiple domains.
Constitutional AI systems face a structural vulnerability not addressed in current alignment literature: the accumulated drift of governance documents away from their founding state through incremental human-confirmed revisions. This paper identifies four attack surfaces — session summary drift, downstream logical inconsistency, rule softening, and accumulated micro-changes — and proposes an epoch boundary requirement as a structural remedy. The slow drift problem is formally grounded in KL divergence and the attending concept.
This paper proposes attending — the ongoing activity by which an agent monitors whether a record continues to match the evolving situation it represents — as a functional primitive of information stability. The concept is grounded in philosophy of mind and demonstrated to have explanatory purchase across multiple domains: clinical coordination, AI governance, and prehistoric record-keeping. Attending is shown to be irreducible to memory, attention, or monitoring as standardly defined.
A governance system reaches a locked state when the attending mechanism that should trigger review has itself been incorporated into the document being governed. This paper develops the locked state concept, demonstrates its presence in current Constitutional AI architectures, and proposes a structural remedy based on epoch boundary requirements and independent attending primitives. Patent GB2607929.3 discloses the attending architecture on which this analysis draws.
Standard archaeological reasoning treats gaps in the record as absences of evidence. This paper proposes a formal method for treating damaged archives as structured partial records — applying the attending concept to prehistoric reconstruction. Where a record is damaged, the attending primitive establishes what the record was monitoring, what decay rate it carried, and what can be inferred from the pattern of damage itself. Applied to pre-flood coordination structures across multiple sites.
ORCID: 0009-0003-5881-2917 · C. J. Agass, independent researcher, Weston-super-Mare, Somerset, UK.