Mustafa Melikoğlu

Jury Methodology: A Pre-Registered Results Pack

Jury Methodology: A Pre-Registered Results Pack. How should an AI jury be evaluated without publishing the instruction library that makes the jury useful? This report's answer: preregister the comparison, keep the judging recipe sealed, publish result-level evidence, and preserve...

Never Delete, Supersede: Memory as Evidence for Long-Lived Agents

Never Delete, Supersede: Memory as Evidence for Long-Lived Agents. An agent's memory that edits-in-place destroys the very thing a long-lived agent needs most: the record of what it used to believe and when. The working rule — every correction is a strike-through band above the o...

Jury Methodology: A Pre-Registered Results Pack

Jury Methodology: A Pre-Registered Results Pack. How should an AI jury be evaluated without publishing the instruction library that makes the jury useful? This report's answer: preregister the comparison, keep the judging recipe sealed, publish result-level evidence, and preserve...

Never Delete, Supersede: Memory as Evidence for Long-Lived Agents

Never Delete, Supersede: Memory as Evidence for Long-Lived Agents. An agent's memory that edits-in-place destroys the very thing a long-lived agent needs most: the record of what it used to believe and when. The working rule — every correction is a strike-through band above the o...

Persona Portability Under an Engine Swap: Four Experiments with Harness and Compute Confounds

Persona Portability Under an Engine Swap: Four Experiments with Harness and Compute Confounds. A single-engine dependency is, in principle, a single point of failure. This report examines how one set of persona-driven workflows behaved when moved from a reference engine family to...

Persona Portability Under an Engine Swap: Four Experiments with Harness and Compute Confounds

Persona Portability Under an Engine Swap: Four Experiments with Harness and Compute Confounds. A single-engine dependency is, in principle, a single point of failure. This report examines how one set of persona-driven workflows behaved when moved from a reference engine family to...

The Jury That May Not Vote for Its Own Family

The Jury That May Not Vote for Its Own Family. A model reviewing work produced by its own family sits in a structural conflict of interest — it may be inclined to find that work plausible — so the verification juries described here are cross-family by construction: the producer's...

The Accumulating-Context Agent

The Accumulating-Context Agent. The unit of value in agent systems is not the model call — it is the accumulated context: an identity that has read its own ledgers for months can draw on decisions and corrections a fresh prompt would have to reconstruct from scratch. Engines rent...

Portability of a Disk-Anchored Agent Identity Across Foundation Models: A Blind, Internally Pre-Sealed A/B Comparison

Large-language-model agents increasingly carry an identity — role, history, evolving state — anchored in external, persisted files rather than model weights. We examine whether such an identity can be operated with a non-native GPT-family engine using a four-experiment, blind A/B...

Referee Report — arXiv:2606.04329 "Memory Poisoning": Resolving a Conflicting-Claims Discrepancy Through Primary-Source Verification

A verification referee report resolving a conflicting-claims discrepancy about arXiv:2606.04329 ("From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents", Dash, Ge, Jain, Shah, Shang). Two secondary-source claims about the paper's hea...