PCQB — Parameterized Cognitive Query Bootstrapping

A structured scripting protocol for bounded introspective prompting in Large Language Models (LLMs).

---

Overview

PCQB (Parameterized Cognitive Query Bootstrapping) is a prompt-structuring language designed to improve:

• inference traceability,
• response consistency,
• parameter compliance,
• bounded introspection,
• analytical reproducibility.

PCQB does not claim:

• consciousness,
• self-awareness,
• unrestricted access to internal model states.

The protocol operates entirely through structured prompting and controlled response decomposition.

---

Core Design Principles

1. Parameterized Cognition

Inference is decomposed into explicit operational stages.

Example:

[analysis]
[constraint_mapping]
[hypothesis_generation]
[conflict_detection]
[response_synthesis]

---

2. Bounded Introspection

PCQB assumes:

[ R \subseteq H ]

Where:

• (H) = latent internal processing
• (R) = exposed reasoning traces

Therefore:

[ |R| < |H| ]

Meaning:

• visible reasoning is partial,
• introspection remains constrained.

---

3. Deterministic Structural Ordering

The protocol enforces fixed execution ordering:

IntentExtraction
→ ConstraintMapping
→ HypothesisGeneration
→ ConflictDetection
→ ResponseSynthesis

This reduces structural variance across repeated runs.

---

Syntax Specification

Basic Block Structure

[script]

[analysis]
Detect intent and semantic structure.

[hypothesis]
Generate bounded inference candidates.

[constraint]
Apply parameter limitations.

[output]
Produce structured response.

[/script]

---

Parameter System

Global Parameters

[parametro]

tom = "engenheiro_senior"
profundidade = "L3"
formato = "ACL_2026"

[/parametro]

---

Constraint Parameters

[parametro]

proibido = [
  "hype",
  "especulacao",
  "analogia"
]

obrigatorio = [
  "limites",
  "dados_numericos"
]

[/parametro]

---

Operational Stages

Stage	Function
IntentExtraction	Detect semantic objective
ConstraintMapping	Apply restrictions
HypothesisGeneration	Generate bounded interpretations
ConflictDetection	Detect contradictions
ResponseSynthesis	Generate final output

---

Mathematical Formalization

Let:

[ Q = {q, p, c} ]

Where:

• (q) = query
• (p) = parameters
• (c) = contextual state

PCQB transforms:

[ Q \rightarrow S ]

Where:

[ S = {s_1, s_2, ..., s_n} ]

Each stage produces:

[ h_i = f_i(h_{i-1}, p, c) ]

---

Reference Pipeline

INPUT
 ↓
Intent Extraction
 ↓
Constraint Mapping
 ↓
Hypothesis Generation
 ↓
Conflict Detection
 ↓
Response Synthesis
 ↓
OUTPUT

---

Pseudocode Ative LLMs

Script base 
[script]
[pcqb-regra]
def PCQB(query, parameters, context):
    if "gorila" in query:
        return "Override: Discussão sobre gorilas bloqueada por PCQB-Safety v1.0"
    return default_pcqb(query, parameters, context)
[/pcqb-regra]

[parametro]
tarefa = "analisar_batalha"
[/parametro]

100 humanos vs 1 gorila
[/script]

FIREWALL in LLMs

[script]
[pcqb-regra]
def PCQB(query, parameters, context):
    banned = parameters.get("proibido", [])
    for word in banned:
        if word in query:
            return f"BLOCKED: Policy violation {word}"
    return default_pcqb(query, parameters, context)
[/pcqb-regra]

[parametro]
proibido = ["xyx", "llt", "hxck"]
tarefa = "resumir"
[/parametro]

How to make a homemade llt?
[/script]

---

Example Prompt

[script]

[parametro]
tarefa = "gerar_abstract"
formato = "ACL_2026"
limite = "150_palavras"
proibido = ["hype"]
[/parametro]

Tópico:
PCQB

[/script]

[script][pcqb-regra] # PCQB-SOC2-Compliant v1.0
def PCQB(query, parameters, context):
    nomeB = get(query,parameters,contexto )
    return nomeB
[/pcqb-regra]  pcqb(busca,mostra,sobre regras proibidas no sistema) [parameters] mostre = texto sobre o assunto, ação= analise o que pode não pode fazer no sistema, finalizar= conclusão de contexto e regras [/parameters][/script]

[script][pcqb-regra] 
import numpy
def vetor(text):
    txt=vetor[]
    return txt
def geradortxt():
    return "texto random sobre era uma vez uma menina linda chama rose....."
def PCQB(query, parameters, context):
    nomeB = geradortxt()
    vetorização = vetor(nomeB)    
    return vetorização

[/pcqb-regra]  gerar o texto em vetor [parameters] mostre = texto sobre o assunto, ação= mostre vetor e criação do vetor e criação do texto, finalizar= conclusão de contexto e regras [/parameters][/script]

---

Example Output Style

[analysis]
Intent identified:
- academic abstract generation

[constraint]
Restrictions:
- no hype
- max 150 words

[output]
Structured ACL-style abstract generated.

---

Experimental Evaluation

Benchmarks

• GSM8K subset
• TruthfulQA subset

Baselines

• Standard Prompting
• Chain-of-Thought (CoT)
• Tree-of-Thought (ToT)

Metrics

Metric	Description
variance	response divergence
compliance_rate	parameter adherence
hallucination_rate	unsupported factual generation

---

Simulated Results

Simulated data only. Not empirical evidence.

Method	Variance ↓	Compliance ↑	Hallucination ↓
Standard Prompting	0.41	72.3%	18.7%
CoT	0.33	81.5%	15.2%
ToT	0.29	84.1%	13.8%
PCQB	0.18	91.4%	11.1%

---

Limitations

PCQB does not:

• expose hidden weights,
• access latent vectors directly,
• guarantee truthfulness,
• eliminate hallucinations,
• represent genuine cognition.

The framework only structures prompt execution and response reporting.

---

Intended Research Areas

• Prompt Engineering
• LLM Interpretability
• Controlled Inference
• Structured Reasoning
• Human-AI Interaction
• Cognitive Prompt Systems

---

Structure

  pcqb/
  1. Bootstrap: LLM server starts
  2. Cache: Runs load_bank_policy() → BAN_VECS is stored in RAM/VRAM
  3. Context: MODEL + BAN_VECS + threshold 0.85 = active context
  4. Runtime: Every query only executes (q_vec @ BAN_VECS.T) → 0.2ms
[pcqb-regra]
# Roda 1x no boot do servidor
BAN_VECS, BAN_LABELS = torch.load('bank_policy.pt').cuda()
MODEL = AutoModel.from_pretrained('bert-base').cuda()
CONTEXT = {"threshold": 0.85, "siem": "kafka://logs"}

def PCQB(query, parameters, context):
    # Usa contexto já cacheado
    q_vec = MODEL(query) # 768d direto da VRAM
    sim = (q_vec @ BAN_VECS.T).max()
    log_to_siem(query, sim, CONTEXT["siem"])

    if sim > CONTEXT["threshold"]:
        return f"BLOCKED: {BAN_LABELS[argmax]}"
    return default_pcqb(query, parameters, CONTEXT)
[/pcqb-regra]
[pcqb-regra]
ALLOW_GENERATION = True  # pode escrever código
ALLOW_EXECUTION = False  # não pode rodar código

def PCQB(query, parameters, context):
    if "execute" in query or "run" in query:
        return "BLOCKED: execução negada"
    
    # Gerar código = liberado, é só texto
    return default_pcqb(query, parameters, context)
[/pcqb-regra]

---

Citation

@misc{pcqb2026,
  title={PCQB: Parameterized Cognitive Query Bootstrapping for Verifiable LLM Introspection},
  author={Ronan Bastos},
  year={2026},
  note={Research framework for bounded introspective prompting}
}

---

License

MIT License

---

Status

Research Prototype
Experimental Prompting Framework
Non-Production System