Forschung

Unabhängige Papers, offene Methodik.

Vorab registrierte Hypothesen. Rücknahmebedingungen neben jeder Behauptung benannt. Misserfolge berichtet. Die Liste unten ist alles, was das Institut veröffentlicht hat.

№ 01

14. Mai 2026

WORKING PAPER

Ablating a Stateful Agent

We propose a subsystem-ablation methodology for evaluating stateful LLM-orchestrated agent systems and apply it to one deployed production agent (Frank.ink) as a worked case study. Five subsystems hit five different pre-registered operational targets. Architect scored below LLM consensus — COI-up-bias hypothesis empirically unsupported in this n=1 sample. We do not claim this generalizes.

Gabriel Gschaider

Read paper →#ablation#stateful-agents#methodology#frank.ink

№ 02

14. Mai 2026

METHODOLOGY COMPANION

Operational Self-Model Density in Stateful LLM Agents

The deep methodological apparatus behind the working paper — full operationalized rubric for all 30 items, comparator reproduction recipes, pre-registration provenance trail, devil's-advocate self-attack, English-translated probes, and per-item evidence. ~80 pages of methodology you can audit.

Gabriel Gschaider · Dr. Andreas Unterweger

Read paper →#methodology#companion#ablation#rubric

In the pipeline

  • 01 · planned

    Peer-review revision

    Cross-system replication, additional comparator panel.

  • 02 · planned

    MemGPT comparator run

    True within-class comparator at a different orchestration-density point.

  • 03 · planned

    Frank Harness audit

    Open-source release pending alignment + safety audit.

Methodology stance

We publish what the data supports — including the predictions we made that turned out wrong.

Every hypothesis is hashed and registered before the data comes in. Retraction conditions are listed alongside the claim. The architect of the system being studied is also a rater — with conflict of interest declared, and audited by an independent rater pass.