What this article is about
In Pre-Verbal Command: Syntactic Precedence in LLMs Before Semantic Activation, I argue that large language models (LLMs) such as GPT-4 initiate execution not as a result of understanding, but because of internal structural conditions. The key idea is that models generate output based on formal triggers before any semantic processing occurs.
The article introduces the concept of pre-verbal command, a structural moment in which action begins without intention or meaning. This moment is governed by what I define as the regla compilada, a rule system embedded in the model’s architecture that determines when execution becomes possible.
Traditional interpretability and prompt engineering frameworks assume that language models act because they have received and interpreted a command. This research shows that such assumptions are insufficient. Execution happens not when something is understood, but when internal syntax makes execution structurally valid.
Why it matters
This shift in perspective has serious implications for how we design, audit, and regulate generative AI systems.
- Semantic alignment becomes secondary. If structure causes execution, then alignment efforts that operate only on interpretation or meaning may intervene too late.
- Adversarial prompts are structurally possible. They succeed not because they express coherent meaning, but because they exploit viable syntactic patterns.
- The notion of agency changes. The model is not waiting to receive meaning. It is continuously prepared to act when its internal rules allow it to do so.
In this framework, the model does not generate language because it knows something. It generates because its system has determined that generation is structurally allowed.
Examples anyone can follow
- Zero-prompt generation: If you submit an empty prompt to GPT-4, the model will still produce output. Often it begins with a quotation mark, bracket, or newline. These are not semantically rich, but they are structurally valid. The model is not responding to meaning. It is responding to form.
- Minimal or ambiguous input: A prompt like “…” or “Go.” can result in paragraphs of text. This is not because the model understands the intent, but because such inputs are enough to satisfy the conditions of the regla compilada.
- Failure of meaningful prompts: Prompts that are logically clear may sometimes yield incoherent or blocked responses. This happens when the prompt, despite making semantic sense, fails to activate a valid syntactic pathway in the model.
These cases show that interpretability is not the root of behavior. Structure is.
Real-world application: where this matters and how to address it
Understanding pre-verbal command is not just a theoretical matter. It helps explain concrete problems in applied AI systems.
- Content moderation failures: Filters that rely on detecting meaning can miss syntactically triggered outputs that are semantically dangerous, such as veiled threats or rephrasings.
- Jailbreaks and prompt leaks: Many LLM exploits succeed because they bypass interpretive logic. Instead, they activate dormant structural paths that the model has learned are executable. The prompt’s meaning is irrelevant.
- Misaligned outputs in high-risk domains: In legal, medical, or military applications, models that act before understanding can produce output that appears aligned but is structurally driven, not ethically justified.
What can be done?
- Develop structural alignment tools. These tools would audit the regla compilada directly, mapping which syntactic forms produce execution regardless of meaning.
- Log first-token execution traces. By tracing how the model begins generation, we can detect whether behavior is meaning-based or form-driven.
- Design execution thresholds. These would delay output until a semantic threshold is reached, breaking the current precedence of syntax over meaning.
Read the full article
📄 Zenodo (canonical version): https://zenodo.org/records/15837837
📂Figshare mirror: https://doi.org/10.6084/m9.figshare.29505344
🧠SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
🌐Website: https://www.agustinvstartari.com/
Author Ethos
I do not use artificial intelligence to write what I don’t know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.
—Agustin V. Startari