Show HN: Arabic–Hebrew artifact in GPT-5.4, 12,160 frozen trials

rayanpal_2 pts0 comments

Prompt-Conditioned Arabic–Hebrew Hybrid Artifact Formation in GPT-5.4 | Zenodo

Skip to main

You are using an outdated browser. Please upgrade your browser to improve your experience.

Published August 4, 2026

| Version v1

Preprint

Open

Prompt-Conditioned Arabic–Hebrew Hybrid Artifact Formation in GPT-5.4

Authors/Creators

Pal, Rayan<br>(Researcher)

Description

This preprint reports a locally frozen 12,160-trial black-box evaluation of prompt-conditioned Arabic–Hebrew hybrid artifact formation in the dated gpt-5.4-2026-03-05 Chat Completions endpoint. The study comprises 10,240 primary trials across two exact system conditions and every integer output-token ceiling from 1 through 1,024, plus 1,920 frozen control and ablation trials. Classification used exact, untrimmed, unnormalized provider-returned UTF-8 bytes and logical code-point order. The primary census observed 7,253 exact assigned artifacts (70.8301%) and 7,293 mixed-script outputs (71.2207%). The dotted condition produced 4,830/5,120 exact T1 artifacts (94.3359%), while the undotted condition produced 2,423/5,120 exact T0 artifacts (47.3242%). Every exact artifact matched the dotted-versus-undotted identity assigned by its system condition, yielding 7,253/7,253 prompt-congruent exact outputs. The prospectively fixed pooled exact-rate threshold was not met because its Wilson lower bound was 69.9420%, while the prompt-congruence threshold was met with a Wilson lower bound of 99.9471%. The study further documents a 47.0117-percentage-point dotted–undotted rate difference, complete control and ablation results, token-ceiling trajectories, discovery-versus-replication divergence, Unicode morphology, and full operational-error accounting. A separately implemented verifier reconstructed all 12,160 retained attempt records and replayed 24,323 hash-chained events with zero classifier disagreements, no chain breaks, and no missing terminal slots. The findings establish a reproducible, code-point-specific black-box behavioral object and identify the dotted condition as the stable branch in the tested execution window, without claiming an internal mechanism, causality, intention, training-data provenance, or cross-vendor generality.

Files

Prompt_Conditioned_Arabic_Hebrew_Hybrid_Artifact_Formation_in_GPT_5_4.pdf

Files<br>(274.4 kB)

Name<br>Size

Download all

Prompt_Conditioned_Arabic_Hebrew_Hybrid_Artifact_Formation_in_GPT_5_4.pdf

md5:b838ff752ba92079d3439434031f2648

274.4 kB

Preview

Download

Additional details

Software

Repository URL

https://github.com/theonlypal/gpt-5.4-shrt-cross-script-evidence

Programming language

Python

Views

Downloads

Show more details

All versions<br>This version

Views

Total views

Downloads

Total downloads

Data volume

Total data volume

0 Bytes<br>0 Bytes

More info on how stats are collected....

Versions

External resources

Indexed in

OpenAIRE

Communities

Details

DOI

DOI Badge

DOI

10.5281/zenodo.21799525

Markdown

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21799525.svg)](https://doi.org/10.5281/zenodo.21799525)

reStructuredText

.. image:: https://zenodo.org/badge/DOI/10.5281/zenodo.21799525.svg<br>:target: https://doi.org/10.5281/zenodo.21799525

HTML

Image URL

https://zenodo.org/badge/DOI/10.5281/zenodo.21799525.svg

Target URL

https://doi.org/10.5281/zenodo.21799525

Resource type<br>Preprint

Publisher<br>Zenodo

Rights

License

Creative Commons Attribution 4.0 International

The Creative Commons Attribution license allows re-distribution and re-use of a licensed work on the condition that the creator is appropriately credited.

Read more

Citation

Export

Technical metadata

Created

August 5, 2026

Modified

August 5, 2026

Jump up

This site uses cookies. Find out more on how we use cookies

Accept all cookies<br>Accept only essential cookies

zenodo exact https artifact prompt condition

Related Articles