APBSC · Phase I
Ten papers
Phase I of the series, all dated 2026-09-21, all at version 0.1. The full drafts are still internal research documents; these pages state what each paper argues, what it rests on, and what it explicitly does not claim.
- 00FrameworkAgentic publication boundary and security closure: the integrated frameworkSplits AI-assisted delivery into five information states and argues that the step from persisted to published has no automatic closure today.
- 01ConceptThe green-light fallacy: when agent verification succeeds and the real postcondition does notA pile of local PASS results supports only the predicates those checks encode. It does not add up to the postcondition you actually care about.
- 02Concept · taxonomyAgent context as an egress surfaceThe context an agent needs in order to work well becomes, once committed, a new way for information to leave.
- 03ConceptFrom secret scanning to semantic privacySecret scanners answer whether a known secret form is present. They do not answer whether private information can be recovered.
- 04HypothesisOrganizational agent context propagationWhen teams share agent rules, templates, skills and scanners, one context error can travel to every repository that inherits it.
- 05HypothesisAI build–attack asymmetryIf building gets cheaper faster than securing does, weak targets multiply; if AI also lowers the cost of finding them, the two sides drift apart.
- 06MethodologyResearcher-blind privacy measurementStudying privacy leaks should not mean reading, keeping and republishing them. The architecture removes the researcher's need to see them.
- 07Study protocolAn empirical study protocol: sampling, matched control and prevalence estimationA preregistrable design for the question the pilots cannot answer: is there a measurable difference, for which classes, and with what uncertainty?
- 08GovernanceProvider and platform responsibility: publication-boundary governance for AI coding toolsCoding agents now observe, execute, persist and publish. Platforms have built strong controls for the first three; publication semantics is still a layer of its own.
- 09Algorithm · closes Phase IThe researcher-blind publication-boundary measurement algorithmThe series' requirements turned into a pipeline a machine can run while the researcher stays blind.