Whitepaper, citation and reproduction.
Whitepaper v2 reports every pre-registered result on real hosts, what has only been shown in simulation, and every failure. Code, specification, experiments, results and the transcripts of the agent runs are open.
Read whitepaper v2PDF on GitHubIdentifiers
- Whitepaper, all versions
- 10.5281/zenodo.24304403
- Whitepaper v2
- 10.5281/zenodo.24320392
- Whitepaper v1
- 10.5281/zenodo.24304404
- Software 0.2.1
- 10.5281/zenodo.24321030
- Software 0.2.0
- 10.5281/zenodo.24317658
- Cue Framework v2.1.0
- 10.5281/zenodo.24301903
- Author
- ORCID 0009-0000-1456-2467
Cite
R. A. Elu. RSOS: an executable spatial-symbolic
language for human–AI and AI–AI communication.
Whitepaper v2, Zenodo, 2026.
https://doi.org/10.5281/zenodo.24304403
The concept DOI always resolves to the latest version. CITATION.cff in the
repository gives the same in machine-readable form.
Reproduce
./reproduce.sh runs the test suite and reruns the simulation experiments. The agent
experiments are re-analysed from their recorded runs, with no new model calls. Each results file carries a SHA-256
hash of its numbers, so a re-analysis either matches the published hash or shows where it differs.
cd experiments
python3 a1_analyse.py # A1 → d224abbf…
python3 a2_analyse.py # A2 → 8916f4de…
python3 a3_analyse.py # A3 → ff058345…
python3 a3b_analyse.py # A3b → 0ae0d031… (needs tiktoken)
python3 ht1_score.py # HT1 → e1f8266b…
The A3b analysis needs tiktoken (o200k). Without it, the analysis stops instead of
reporting BPE counts of −1.
What changed from v1 to v2
- v1 described the language and fixed the predictions of the first real-host test.
- v2 reports that test (A1) and three more (A2, A3, A3b), plus the first human test (HT1).
- Comparison now includes re-implementations of two published compact protocols.
- The keyboard layer, the failure map and the 24-domain programme are new.
Related work
Compact agent messaging is an active field. Each system below optimises one axis: size, structure or term meaning. A3 and A3b measure the one they leave open: whether every copy of the shared state stays true, and what it costs to make it true again.
| Work | What it does |
|---|---|
| Slipstream | Maps messages onto shared concept anchors; reports 82% fewer tokens. |
| ACCP (IETF draft) | Compact frames, an intent ontology, checkpoints and delta encoding; reports 60–90% fewer tokens. |
| AACP (IETF draft) | Typed coordination format; recurring instructions encoded by cache and pattern lookup. |
| CLSR (ICML 2026) | Agents evolve compact symbolic reasoning languages; 3–6× fewer generated tokens than chain of thought. |
| Verifiable Semantics for A2A | Certifies terms behaviourally and recertifies them on drift. |
Earlier work by the author: The Recursive Symbolic Operating System (RSOS), whitepaper v1.0/v1.1, August 2025. Its validation claims were internal; the pre-registered programme reported here replaces them.