RD-2026-008
concluded
Does the same agent loop produce the same investigation?
Three non-frontier coding models reviewed the same pre-fix software through a common agent loop. Their execution histories exposed different evidence choices, missed boundaries and unsupported conclusions, showing why tool availability alone does not establish investigative reliability.
19 Sep to 20 Sep 2026
2 days
12 research hours · estimated
Core
#ai#coding-agents#non-frontier-models#verification
RD-2026-007
concluded
When the compiler writes itself, what is the source?
We replaced an implementation-shaped natural-language corpus with five small Program files, rebuilt ProgSync without its former source, and tested the result against an independent 55-case oracle. The build worked, but one clear synthesis error showed why verified generated code cannot be disposable.
23 Jul to 23 Jul 2026
1 days
15 research hours
Core
#ai#natural-language#program#progsync
RD-2026-004
concluded
Can we turn source code into natural language?
We tested whether readable Markdown could preserve enough of a program for isolated AI translators to reconstruct working software. A 21-module trial succeeded; self-hosting exposed deeper contract failures and left the central claim unproven.
21 Jul to 24 Jul 2026
4 days
32 research hours · estimated
Core
#ai#natural-language#program#progsync