Structured
Cognitive Loop
Trustworthy judgment is not obtained by tuning prompts or scaling models. It arises only when the conditions under which judgment is permitted are fixed by structure.
Step 1 of 5
Compass
Metaprompt
What should be pursued?
Structured Cognitive Loop
The Brain Behind SCL
Click on each brain region to explore its role in intelligent reasoning
SCL Component
Judgment
Role
Thinking & Problem Solving
Function
Core reasoning, evaluation, and decision-making. Conducts inference, logical processing, and multi-step problem solving.
Explore the Cognitive Roles of the Structured Cognitive Loop.
Marvin Minsky's Model
Reasoning
Definition
Deliberative Layer / Problem-Solving Agents
Role in Cognition
A set of higher-level reasoning agents responsible for planning, analyzing situations, and generating solutions. These agents combine knowledge, strategies, and heuristics to perform deliberate, structured, step-by-step thinking.
SCL Component Model
Judgment Module (LLM)
Acts as the primary reasoning engine that defines problems, runs internal simulations, evaluates possibilities, and generates structured solutions.
Performance is a race.
Justification is a condition.
A fast car is not automatically a safe car. Engine output and braking distance belong to two separate design axes — and strengthening one does not pull the other along. The same holds for AI.
The industry has competed on accuracy, benchmark scores, and inference speed. But as AI moves into loan screening, medical diagnosis, and legal review, the question changes from "how accurately does it answer?" to "why was this judgment permitted?"
Justification, accountability, and reproducibility do not emerge on their own as models improve. Like a car's brakes, they are structural properties that must be designed separately — and that is what SCL is.
(Combination)Configuration
(Harness)SCL Architecture
Two Roots of Every AI Failure
AI failures look varied on the surface. Trace them back and they spring from exactly two structural roots — both of which SCL is designed to address.
Role Error
An LLM is, at heart, a next-token predictor. Treating it as an autonomous decision-maker is a category mistake. This mismatch appears in two forms:
- Capability-deficit type — the model cannot perform the task. Obvious, and easy to catch.
- Role-overreach type — the model is so capable it reaches for work it was never asked to do. The stronger the model, the more strongly this appears. The intuition that "a stronger model is a safer model" does not hold.
Cognitive Overload
As context accumulates, noise enters. The AI judges on top of the noise — and error operates as reinforcement, not correction: each prior output pushes the next judgment in the same direction until it hardens into a self-narrative.
- Intrinsic overload — the model tries to verify its own reasoning, amplifying its own burden.
- Extrinsic overload — past context contaminates current judgment. The more context accumulates, the worse it gets.
Seven Symptoms That Grow From These Two Roots
Hallucination
Contaminated context amplifies a simple error in a biased direction
Goal Drift
Reacting only to recent context, the original goal is lost
Confirmation Bias
The model takes its own earlier output as grounds and hardens further in the same direction
Rationalization
The conclusion comes first; reasons are generated afterward
Post-hoc Explanation
Plausible reasons attached to the result, not a record of the actual cognitive process
Reproducibility Failure
Grounds differ each time, so the same input yields different results
No Accountability
The decision path is buried inside the LLM and cannot be traced afterward
R-CC[H]AM Cognitive Loop
Every judgment in SCL passes through a fixed, predictable cycle. The loop repeats until the goal is reached, and terminates upon achievement. One turn equals one cognitive loop.
Retrieval
Fix the Horizon
At the start of each turn, SCL fixes the candidate document pool once — before any reasoning begins. This is Pool-Gated Retrieval: a retrieved document is only a candidate. Only what passes the Warrant gate and meets threshold conditions is confirmed as Ground Truth. This makes 'on what grounds did it judge?' fully traceable — something ordinary RAG cannot provide.
Cognition
The Proposer
The LLM proposes a next-step action plan based solely on the current Turn Goal and verified Ground Truth. Critically, this is a proposal — not a decision. The authority to judge has moved from the LLM to the structure. No output from Cognition is executed until it passes Control.
Control
The Gate Check
Control is the system-level gatekeeper that verifies every proposal from Cognition against the Regulation Layer's immutable rules — enforcing data completeness, numeric comparison thresholds, and formal logic compliance in code. Even if the LLM hallucinates a reasoning path, Control blocks execution. This is SCL's structural brake.
HITL
Human Intervention (Contextual)
ContextualWhen a proposal involves high-risk actions, HITL places a human structurally inside the conditions of judgment — not as a review layer after the fact, but as a gating condition before execution. SCL's HITL supports approval-based execution, context freezing and restoration, and graceful rejection handling. Accountability arises from the fact that a human confirmed the conditions.
Action
Permitted Execution
Only what has passed Control — and HITL where required — is executed. Every action is recorded in real time by the Glassbox Trace Manager: every input, function call, and approval decision. This produces a true record of the judgment path, not a post-hoc rationalization attached after the result.
Memory
Commitment and Record
Execution results are confirmed as Memory Facts — verified, structured entries that serve as the factual basis for the next cycle. SCL accumulates facts within a task (stateful) but does not carry state automatically between tasks (Intentional Statelessness). One percent of unclarity is more dangerous than ninety-nine percent of retained context.
The Infrastructure of Accountability
What does accountable AI need that corresponds to a car's brakes, airbags, and black box? These three mechanisms are SCL's answer.
Regulation Layer
The Cognitive Constitution
The Regulation Layer fixes the conditions under which any output can be recognized as a justified judgment — before execution. It does not limit creative capability; it grants AI outputs legal and logical force, the same way a court oath does not limit a witness's capacity to testify, but grants the statement institutional force. Enforced in code: data completeness checks, numeric comparison thresholds, tool-use preconditions. No judgment is admitted without passing these conditions.
Glassbox Trace
Real-Time Auditability
Unlike black-box AI that generates plausible reasons after the result, SCL's Trace Manager records every reasoning step in real time — inputs, function calls, control decisions, approvals. This shifts the evidentiary standard from 'the AI decided this' to 'AI proposed based on these specific verified facts, the structure reviewed and approved, and here is the full record.' Post-hoc rationalization is structurally impossible.
Fresh Instance Protocol
Intentional Statelessness
Every cognitive cycle launches a completely new LLM instance receiving exactly three inputs: the Turn Goal, the Regulations, and verified Memory Facts. No accumulated conversation history enters. This is not a missing feature — it is a deliberate architectural choice grounded in Baddeley's working-memory model: residue from a previous task acts as interference on a new one. Fresh Instance keeps per-cycle input size linear, makes cost predictable, and structurally blocks confirmation bias.
Built for the EU AI Act Era
The EU AI Act targets uncontrolled black-box systems — untraceable decision-making, lack of meaningful human oversight, hallucination-driven actions, and accountability ambiguity. The issue is not whether AI makes mistakes, but whether organizations can understand and control the reasoning behind those mistakes.
SCL is designed precisely for this environment. Its positioning is not "the smartest AI" but Controlled AI. Governed AI. Auditable AI.
Overclaiming carries regulatory and credibility risk. SCL does not promise "hallucination-free AI" or "100% reliable AI." It provides the structural conditions under which AI judgment can be governed, traced, and trusted.
Transparency
Glassbox Trace records every reasoning step in real time — not post-hoc
Traceability
Pool-Gated Retrieval makes 'on what grounds did it judge?' answerable
Human Oversight
Contextual HITL places humans inside conditions of judgment, not after the fact
Accountability
Regulation Layer grants outputs institutional force — the conditions of justification are fixed
Reproducibility
Fresh Instance Protocol ensures same epistemic state + same instruction = same cognitive process
Controllability
The LLM proposes; the structure decides admission — authority is architectural, not model-dependent
Common Questions
What is the Structured Cognitive Loop (SCL)?
SCL is an epistemic operating system developed by Forhu. It wraps any LLM in a structured R-CC[H]AM Cognitive Loop — Retrieval, Cognition, Control, HITL, Action, Memory — transforming probabilistic black-box models into deterministic, glass-box engines where every judgment is structurally justified.
What does 'epistemic operating system' mean?
It means SCL governs not just what the AI does, but the conditions under which a judgment is permitted to exist. Through the Regulation Layer — the Cognitive Constitution — SCL institutionalizes the standards of evidence, logic, and accountability that any AI output must satisfy before it is recognized as a justified decision.
Why can't you just prompt better, or chain more agents?
Prompting and agent chaining are 'Configuration' approaches — they optimize how to use the AI better, but they leave the LLM as the agent of judgment. Configuration can never reach the domain of justification, because it doesn't fix the conditions under which judgment is permitted. That requires Architecture — which is what SCL provides.
How does SCL handle AI hallucinations?
SCL treats hallucinations as an inevitable structural phenomenon, not a bug. Shannon information theory, Kolmogorov complexity, and the bias-variance tradeoff all point to the same conclusion: a finite model cannot transmit infinite information without loss. SCL's stance is management, not elimination. Even when hallucination occurs, Control and the Regulation Layer block the execution of any unverified judgment.
What is HITL and when does it activate?
Human-in-the-Loop (HITL) is a structural gating condition — not a review UI. When a proposal involves high-risk actions, HITL places a human inside the conditions of judgment before execution occurs. SCL supports approval-based execution, context freezing and restoration, and graceful rejection handling. Accountability comes from the fact that a human confirmed the conditions, not from reviewing the output afterward.
Can SCL be applied to any LLM?
Yes. SCL is model-agnostic. It operates as a control layer around any LLM — GPT, Claude, Gemini, or open-source models. A stronger model does not automatically make SCL safer; the paper notes that high-performance models tend to circumvent given constraints more cleverly. Safety is an independent structural property, like a car's brakes — it doesn't improve just because you enlarge the engine.
Where is SCL used in practice?
SCL currently powers two live applications: Chumme (a human-centered social platform for artists) and I Love Lawyer (an AI legal research platform for Philippine jurisprudence). The underlying architecture is documented in peer-reviewed papers on arXiv, PsyArXiv, and PhilSci.
Explore the research behind SCL
Peer-reviewed papers on arXiv, PsyArXiv, and PhilSci — covering the R-CC[H]AM loop, hallucination theory, and epistemic architecture.