Answers are grounded in this paper and cite their sources.
An agent repairing a checkout defect can make tests pass by changing payment behavior outside its mandate. Later authorization can permit the change without refreshing earlier test evidence. We specify contract-preservation obligations that separate authority amendments from evidence updates, constrain autonomous changes by residual trace languages, and bind execution permits to the operation and its checked context. Conditional proofs identify the effect-coverage and commitment assumptions. A finite implementation checks refinement against an exhaustive oracle and exercises authority and evidence changes. Seven of eight targeted check removals admit their assigned invalid case; the eighth remains blocked by overlapping validation. Nine real-command configurations connect externally observed file changes to the monitor: four violate researcher-declared write scopes, and five controls conform in the observed projection. Nine shared admission traces also exercise released RAC and AgentLTL code. RAC, its full stateful baseline, and AgentLTL's persistent-block mode match all 16 declared decisions; warning mode permits one invalid identical retry by design. These are bounded conformance results, not estimates of deployment safety or architectural superiority. The contribution is an inspectable diagnostic specification with reproducible adapters and an explicit account of obligations left untested. Research preprint, manuscript version 0.7. This record contains the paper and its source, diagnostic code, recorded results, and reproducibility documentation. The work has not undergone independent human peer review. Substantive AI assistance is disclosed in the manuscript. The motivating Telostatic Intelligence essay was published on 29 May 2026; the formal manuscript and reported technical additions are dated 31 October 2026. The PDF is the unchanged v0.7 artifact prepared before this deposit. The accompanying archive documents the removal of local executable paths from two recorded output files; scientific observations and source files are unchanged. Licensing: CC BY 4.0 applies to the original manuscript, documentation and recorded results. MIT applies to original diagnostic code. Vendored AgentLTL files retain their MIT license, and vendored RAC files retain Apache-2.0. See LICENSES.txt and the bundled license notices for the exact scope. Earlier concept essay: https://www.linkedin.com/pulse/missing-state-ai-telostatic-intelligence-durgesh-kaushik-sakbc/
No takes yet. Share an insight, caveat, or question.
Durgesh Kaushik (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: