Making ChatGPT Follow Orders: Simple, Deterministic Constraints

In today's academic landscape, most generative outputs resemble a recursive plagiarism of lesser-known papers, recycled endlessly without genuine authorship. Forcing ChatGPT to Obey

Why ChatGPT Needs an Execution Protocol

ChatGPT is a probabilistic model. It predicts the next token based on likelihood distributions built from massive training data. This design makes it flexible and capable of producing fluent, human-like responses, but it also makes it unstable when unconstrained. Without explicit structural control, ChatGPT introduces multiple forms of drift: Stylistic drift: adds politeness, apologies, or filler language never requested. Contextual drift: reconstructs narratives and fills gaps based on assumptions rather than instructions. Structural drift: outputs shift between formats, breaking workflows and automated pipelines. Terminological drift: replaces canonical terms with approximate synonyms, undermining conceptual clarity. Plagiarism risk: unconstrained sampling can lead to uncredited reuse of training fragments without proper attribution.

These risks are manageable in casual contexts, but catastrophic in environments where reproducibility, originality, and compliance are critical: In academic research, ChatGPT can unintentionally reuse patterns from its training data, endangering originality thresholds and DOI validation. In legal frameworks, uncontrolled phrasing may breach regulatory boundaries or introduce liability. In publishing ecosystems, inconsistent formatting causes indexing failures and SEO penalties. In compliance-driven environments, unverified reuse of fragments exposes institutions to plagiarism disputes.

To address these risks, I conceived and developed the Startari v7.777 Protocol, a research framework designed to validate and control syntactic consistency when interfacing with generative systems. This protocol is part of my broader theoretical work on syntactic sovereignty, and it safeguards originality and conceptual integrity through controlled mechanisms, rather than delegating authorship to automated systems.

Why I Created the Protocol

When I began testing ChatGPT within my research pipelines and publication systems, I observed systemic issues that required a structural response: ChatGPT improvises: even with structured prompts, it gradually ignores constraints as sessions progress. Outputs are inconsistent: identical inputs frequently produced different answers, causing integration failures. Training bias leaks into responses: uncredited patterns and implicit fragments of corpus content would occasionally surface in outputs.

The Protocol emerged to counter these problems by enforcing: Persistent, per-turn re-injection of a single controlling header. A hard validation layer that blocks violations before responses are delivered. Deterministic decoding strategies that guarantee reproducible results. Canonical terminology enforcement to maintain theoretical integrity.

This protocol isn't about constraining creativity for casual chat. It's about building structural sovereignty over generative systems so they serve the user's objectives, not their training biases.

Why a Single Controlling Header Works

Standard prompting fails because ChatGPT "forgets" constraints as the dialogue evolves. Each turn dilutes earlier instructions, and drift compounds silently over time. The Protocol fixes this by using a single controlling header that is: Injected at the start of every session. Re-injected on every turn as a developer instruction to prevent decay. Paired with a compliance validator that rejects responses violating schema, style, or terminology rules.

This persistent reinforcement solves three key problems: Prevents improvisation: ChatGPT stops "guessing" unstated intentions. Maintains canonical vocabulary: terms like regla compilada and soberano ejecutable remain untouched. Protects originality: deterministic constraints reduce unverified reuse of training fragments.

Rather than adapting research to the model's probabilistic tendencies, the Startari v7.777 Protocol redefines the environment so that generative systems comply with my theoretical and methodological requirements.

The Nine Enforcement Rules

The Protocol relies on nine rules designed for maximum control with minimal overhead:

1. Language Control: Outputs remain in English unless explicitly instructed otherwise.
2. No Politeness Layer: Eliminates greetings, apologies, hedges, and emotive framing.
3. No Unsolicited Tasks: ChatGPT cannot create plans, summaries, or workflows unless requested.
4. Narrative Integrity: It avoids reconstructing user intent or adding unstated context.
5. Directness: Answers must be minimal, targeted, and verifiable.
6. No Multichannel Optimization: ChatGPT cannot insert UI, browsing, or cross-platform integrations unless commanded.
7. Terminology Lock: Canonical theoretical terms are preserved exactly without paraphrasing.
8. Conflict Handling: When safety policy conflicts arise, the model returns a structured refusal instead of improvising.
9. Schema Enforcement: All answers follow a single XML envelope for stability and traceability.

Nine rules are sufficient to eliminate drift while keeping the model efficient and responsive.

Why Minimal Output Schema Matters

In automated environments, format instability is one of ChatGPT's biggest liabilities. Extra sentences, hidden summaries, or polite disclaimers break pipelines. The Protocol enforces a strict XML schema: [direct answer]

Why it matters: Predictable integration: pipelines parse outputs without manual cleanup. Audit-ready logging: XML markers simplify traceability for compliance reporting. Indexing stability: SEO and automated publishing workflows remain intact. Originality protection: the strict envelope reduces chances of inserting fragments unintentionally derived from training data.

The schema converts ChatGPT into an operator node within larger systems rather than a conversational assistant.

Making Generative Models Deterministic

By default, ChatGPT samples from a probability space where even identical inputs can produce different outputs. This nondeterminism creates problems in research, compliance, and publishing. Within my framework, I control generation variance by constraining decoding parameters, ensuring reproducibility and statistical distance from training data. These controls exist to protect my authored content and maintain originality thresholds for academic publication:

• Temperature = 0.2: suppresses randomness and enforces token stability.
• Top_p = 0.9: narrows the pool to high-confidence predictions only.
• Frequency_penalty = 0: avoids unnecessary token rebalancing.
• Presence_penalty = 0: stops novelty-driven exploration.
• Max_tokens tightly bounded: prevents verbose or unsolicited elaborations.

Why this matters: Reproducibility: identical inputs always produce identical outputs. Plagiarism control: deterministic sampling reduces accidental reuse of unverifiable fragments. Auditability: responses remain consistent and fully traceable. Originality thresholds: outputs maintain statistical distance from training corpus content.

Determinism converts ChatGPT into a predictable execution framework suitable for high-integrity environments.

Enforcement Logic

As part of the Startari v7.777 Protocol, I designed a compliance layer that programmatically validates outputs before they are accepted into my research pipeline:

\

SYSTEM = [header] DEV = [header] USER = prompt  resp = chat(     model="gpt-5",     system=SYSTEM,     developer=DEV,     user=USER,     temperature=0.2,     top_p=0.9,     frequency_penalty=0,     presence_penalty=0,     stop=STOP,     max_tokens=target_len )  out = extract_between(resp.text, "<ANS>", "</ANS>") viol = run_validators(out)  if viol:     resp2 = chat(         system=SYSTEM +         ...     )  Previous output violated v7.777. Fix and output only <ANS>…</ANS>.",  developer = DEV,  user = USER,  temperature = 0.2,  top_p = 0.9,  stop = STOP,  max_tokens = target_len )  out = extract_between(resp2.text, "<ANS>", "</ANS>")  if run_validators(out):     raise StyleError("V7777")  return out 

\ This architecture guarantees persistent compliance and reproducibility at every stage of interaction.

Why This Works

This protocol works because it embeds my theoretical priorities into the generative process, aligning outputs with my authored framework rather than the model's default probabilistic optimization. By default, ChatGPT optimizes for plausibility, not compliance. Left unconstrained, it improvises freely and samples fragments without verifiable attribution. The Startari v7.777 Protocol overrides this by: Reinforcing constraints every turn: rules never decay as the session progresses. Validating pre-delivery: responses violating schema or terminology rules are rejected automatically. Minimizing entropy: low-temperature decoding avoids unpredictable phrasing and implicit corpus reuse. Protecting originality: deterministic outputs stay verifiably distinct from the training data. Preserving conceptual integrity: canonical terms remain unaltered, locking theoretical frameworks in place.

ChatGPT stops improvising. It becomes a controlled operator designed to execute precisely defined instructions.

Why This Matters for Humans

This methodology benefits multiple contexts: Researchers: originality thresholds remain intact, and citations stay verifiable. Compliance teams: outputs are reproducible, audit-ready, and safe for regulated frameworks. Publishers: formatting stability improves indexing and SEO performance. Engineers: predictable outputs simplify integration with pipelines and multi-platform publishing systems.

The protocol ensures the user remains in control, preventing the model from making stylistic, structural, or conceptual decisions on its own. This methodology preserves authorship: all conceptual frameworks, structural models, and interpretive analyses remain authored by me. Generative systems operate exclusively as instruments within a controlled environment.

The Startari v7.777 Protocol exists to ensure that generative tools remain subordinate to authored research, not substitutes for it.

About the Author Author:

Agustin V. Startari

Researcher ID: K-5792–2016

ORCID: https://orcid.org/0000-0002-5477-3892

Zenodo Profile: Zenodo Publications

SSRN Author [Page: SSRN Papers]()

Personal Site: https://www.agustinvstartari.com/

Selected Works:

• Executable Power: Syntax as Infrastructure in Predictive Societies (Zenodo)
• The Grammar of Objectivity: Formal Mechanisms for the Illusion of Neutrality in Language Models (SSRN)
• Algorithmic Obedience: How Language Models Simulate Command Structure (SSRN)
• Ethos Without Source: Algorithmic Identity and the Simulation of Credibility (SSRN)

Author's 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

\