Română

Ioan Daniel Ionici

Independent researcher · Oradea, Romania

About

I work on one question: what can a persistent AI agent know about elapsed time and about its own change, when its whole state can be saved, copied and restored? The works below propose a way to ask that question and test it against conventional alternatives.

This is a personal, free effort, without the claim of changing a field. Where something is not known, the works say so; where I was wrong, they were corrected in new versions with a change note.

Works

All works are open access on Zenodo. The DOIs below are concept DOIs: they always lead to the latest version.

A Time of Their Own: Agent-Held Temporal Evidence for Persistent AI Agents

Perspective · preprint · version 1.3 · 19 pages

A conceptual proposal and a programme of tests. It classifies the temporal evidence an agent can hold (order, duration, sequential computation, continuity, its own change) by who can write, reset, forge or copy it; states an elementary result (an agent whose whole restorable state is restored cannot detect the restoration from that state alone); separates a continuity anchor from a dual clock; and pairs each of six tests with the conventional competitor that could make it unnecessary.

DOI: 10.5281/zenodo.23671714One-page fact sheet (PDF)
Cite · BibTeX

Ionici, I. D. (2026). A Time of Their Own: Agent-Held Temporal Evidence for Persistent AI Agents. Zenodo. https://doi.org/10.5281/zenodo.23671714

@misc{ionici2026perspective,
  author    = {Ionici, Ioan Daniel},
  title     = {A Time of Their Own: Agent-Held Temporal Evidence for Persistent AI Agents},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.3},
  doi       = {10.5281/zenodo.23671714},
  url       = {https://doi.org/10.5281/zenodo.23671714}
}

Does Explicit Episode–Trace Binding Help? An Exploratory Study with Information-Matched, Competence-Audited Controls

Exploratory study · preprint · version 1.3 · 38 pages

A small, synthetic, offline benchmark of source attribution. An apparent advantage of explicit binding (+0.38 Brier skill score) disappeared once the information-matched control became competent; the signal came from local update statistics, not from accumulated history. The contribution is methodological: a decoding ceiling, a control-competence audit, equivalence tests and construct-validity diagnostics. Version 1.3 corrects the claim that the benchmark is closed.

DOI: 10.5281/zenodo.23666335One-page fact sheet (PDF)
Cite · BibTeX

Ionici, I. D. (2026). Does Explicit Episode–Trace Binding Help? An Exploratory Study with Information-Matched, Competence-Audited Controls. Zenodo. https://doi.org/10.5281/zenodo.23666335

@misc{ionici2026study,
  author    = {Ionici, Ioan Daniel},
  title     = {Does Explicit Episode–Trace Binding Help? An Exploratory Study with Information-Matched, Competence-Audited Controls},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.3},
  doi       = {10.5281/zenodo.23666335},
  url       = {https://doi.org/10.5281/zenodo.23666335}
}

Episode–Trace Binding Pilot: Code, Frozen Configurations, and Per-Run Results

Software · code and data · version 1.3

Simulation code, frozen and hashed configurations, per-run results (533 runs), reports, 20 automated checks and the scripts that regenerate every reported table. Version 1.3 adds a hashed lockfile for the exact environment and one command, reproduce_lite.py, for three levels of verification. These checks establish internal consistency, not independent replication.

DOI: 10.5281/zenodo.23667153One-page fact sheet (PDF)
Cite · BibTeX

Ionici, I. D. (2026). Episode–Trace Binding Pilot: Code, Frozen Configurations, and Per-Run Results. Zenodo. https://doi.org/10.5281/zenodo.23667153

@misc{ionici2026archive,
  author    = {Ionici, Ioan Daniel},
  title     = {Episode–Trace Binding Pilot: Code, Frozen Configurations, and Per-Run Results},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.3},
  doi       = {10.5281/zenodo.23667153},
  url       = {https://doi.org/10.5281/zenodo.23667153}
}

Un timp al lor: despre ceasuri, urme și memorie la inteligențele artificiale care trăiesc mai mult decât o conversație

Popular-science brochure · Romanian · version 1.0 · 15 pages

A Romanian-language brochure for general readers that retells the idea of the Perspective, the closed study and the scientific reality about the risks and benefits of AI agents. It contains no new results. 25 sources.

DOI: 10.5281/zenodo.23703417One-page fact sheet (PDF)
Cite · BibTeX

Ionici, I. D. (2026). Un timp al lor: despre ceasuri, urme și memorie la inteligențele artificiale care trăiesc mai mult decât o conversație. Zenodo. https://doi.org/10.5281/zenodo.23703417

@misc{ionici2026brosura,
  author    = {Ionici, Ioan Daniel},
  title     = {Un timp al lor: despre ceasuri, urme și memorie la inteligențele artificiale care trăiesc mai mult decât o conversație},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.0},
  doi       = {10.5281/zenodo.23703417},
  url       = {https://doi.org/10.5281/zenodo.23703417}
}

Status and limits

  • Exploratory. One experiment is completed; it ended in a practically null result against a competent control. No confirmatory data have been collected.
  • A proposal, not a demonstration. The Perspective proposes a hypothesis and the tests that could refute it; it does not demonstrate it. It also states that the accompanying study does not test the programme.
  • Mechanisms are not new. Trusted counters, rollback protection, machine unlearning, data attribution and update-norm clocks are established or adjacent. The claimed contribution is the separation of three arrangements and the test programme.
  • No independent verification. The code was written with AI assistance; the checks establish internal consistency only. No human review or independent retraining has been carried out.
  • Versions. Each version carries a change note. Corrections narrow claims; the hypothesis is unchanged.

How to cite

Use the concept DOI of the work, which resolves to the latest version, and name the version you used. Each work above has a ready citation and BibTeX entry.

Series for general readers

A six-article series in Romanian, “Cum funcționează un LLM”, explains how a language model works, from tokens to agents, with English terms next to the Romanian ones.

Series page (Romanian) →