Collapse Aware AI Core Gold Build

The Collapse Aware AI™ Core Gold Build is an operational middleware layer for retained-state-sensitive behavioural selection.

It is designed for applications that already have a bounded set of possible actions, responses or behaviours and need a controlled way for relevant prior state to influence which permitted option is selected.

CAAI does not replace the surrounding model, workflow system, game engine or authored content. The host supplies the available actions and remains responsible for execution. CAAI operates at the selection boundary between those permitted possibilities and the resulting behaviour.

The accepted Core Gold Build operates locally and can perform selection without requiring a paid model-provider call for every decision.

Technical evidence and bounded paid evaluations are available now.

What the Core Gold Build does

A host application supplies a situation and a bounded set of candidate behaviours.

Core Gold can then:

• receive and preserve the supplied candidate behaviours;
• apply retained-state influence during selection;
• select one of the supplied permitted behaviours;
• preserve controlled state for later recall;
• produce durable decision evidence;
• preserve request and thread identifiers;
• support deterministic reproduction from identical accepted inputs and state;
• validate malformed requests;
• report degraded or invalid upstream responses honestly rather than fabricating success;
• operate through the verified application → API → adapter → runtime path.

The surrounding application remains responsible for creating and permitting the available actions.

CAAI is the selection layer, not the action generator.

Reference behaviour and CAAI selection

For commercial evaluation, a host can declare the action its workflow would ordinarily take.

CAAI then performs retained-state-sensitive selection among the permitted actions.

The host-declared reference action and the CAAI-selected action can be recorded and inspected together.

This provides a practical way to ask:

Should relevant retained history affect which permitted behaviour is selected now?

The current Core Gold Build should not be interpreted as providing a clean single-variable “history off / history on” counterfactual. That stronger materiality comparison is part of the next Commercial v2.0 selector architecture.

A simple example

A collections workflow has three permitted next actions:

• send another reminder;
• offer a payment-plan route;
• route the case to human support.

The host supplies those permitted actions and declares the ordinary reference action.

Relevant retained case state may then influence which permitted action CAAI selects.

The resulting decision can be returned with an auditable Decision Record.

The same pattern can apply to NPC behaviour, agent workflows, simulations, training environments and other systems where previous events should influence later bounded choices.

Where Core Gold can fit

AI agents

A specialist selection layer beneath an existing agent or orchestration system.

The host decides which actions are allowed. CAAI applies retained-state-sensitive selection within that permitted set.

Games and NPC systems

A game can supply authored behaviours and relevant retained state from earlier player interactions.

CAAI can then influence which permitted behaviour is selected while keeping the decision reproducible and inspectable.

The current Core Gold selector does not independently infer meaning from arbitrary prose. The host integration determines the relevant retained-state relationship.

Simulation and training

Prior scenario events can influence later bounded behavioural choices while preserving an inspectable decision trail.

Enterprise and workflow systems

Previous case history can influence selection between already-permitted operational actions without requiring CAAI to replace the existing workflow or policy platform.

Research and evaluation

Core Gold provides a controlled local selection surface with retained state, persistence and durable decision evidence suitable for bounded technical evaluation.

These are integration targets. CAAI does not claim to replace the surrounding platform.

Why the distinction matters

Many AI memory systems focus on:

What information should be stored or retrieved?

CAAI focuses on a different question:

How should retained information affect the behavioural choice being made now?

Likewise, CAAI is not another workflow orchestrator.

A workflow, agent framework or game system can continue to establish the permitted actions.

CAAI sits beneath that layer and performs retained-state-sensitive selection among those possibilities.

Local and controlled operation

The Core Gold selection runtime can operate locally.

The selection path itself does not require a paid cloud-model call for every decision.

This can be useful where a buyer values:

• predictable runtime costs;
• local processing;
• separation from model providers;
• retained-state privacy;
• reproducible testing;
• inspectable decision evidence;
• controlled integration beneath an existing system.

A surrounding application may still use cloud models or services where its owner chooses.

CAAI does not require the entire host system to be cloud-free.

Current accepted evidence

The accepted Core Gold evidence lineage covers:

• retained-state-sensitive candidate selection;
• persistent state and recall across restart;
• preservation of supplied candidate behaviours;
• durable Decision Records;
• deterministic reproduction under identical accepted conditions;
• request and identifier preservation;
• structured request validation;
• honest degraded-mode and invalid-response handling;
• verified UI/API/adapter/runtime routing;
• customer-safe decision evidence;
• local managed-evaluation evidence.

Detailed technical material is kept private and is available under an appropriate evaluation, licensing or development discussion.

Current boundaries

The accepted Core Gold Build is a Phase-1 foundation.

Its current selector has known limitations, including:

• no native host-supplied base-utility model;
• no semantic understanding of arbitrary candidate prose;
• no complete adaptive NPC brain;
• no full Evolution 2 continuity architecture;
• no completed Weighted Emergence Layering implementation.

These boundaries are documented rather than hidden.

Core Gold Commercial v2.0

A new commercial selector lineage has now been architecturally specified.

Core Gold Commercial v2.0 is separate from the frozen Core Gold v1.0 evidence baseline.

Its planned improvements include:

• native host-declared reference/base preference;
• structured retained-state relationships rather than lexical coincidence;
• retained history capable of promoting or suppressing behaviour;
• bounded directional influence;
• constraint-only pre-selection policy;
• genuine buyer-supplied scenarios;
• controlled retained-state materiality comparison;
• new commercial Decision Records and customer-safe evidence.

Commercial v2.0 is specified, not yet delivered, and will require fresh implementation, testing, hashes and independent verification before its capabilities are presented as accepted.

Evolution 2 and future stability work

Evolution 2 is a separate engineering lineage focused on richer continuity capabilities including structured retained state, bounded retrieval, semantic association, Open Loops, Interaction Fit, Agent Self-History and contradiction/change surfacing.

A further Runtime Stability / Governor v2 layer is planned to investigate stronger runtime detection and constraint of behavioural drift, repetition, contradiction and continuity instability.

These later systems are roadmap and engineering work and are not presented as completed Core Gold capabilities.

What Core Gold is not

Core Gold is not presented as:

• a foundation model;
• AGI;
• consciousness or sentience;
• a chatbot replacement;
• autonomous reinforcement learning;
• a finished adaptive NPC brain;
• guaranteed hallucination prevention;
• guaranteed identity-drift prevention;
• full WEL;
• a replacement for the host application.

Its role is narrower:

retained-state-sensitive behavioural selection with persistent, auditable decision evidence.

Proprietary technology

The CAAI runtime, internal selection implementation, scoring logic and Crown internals remain proprietary to Inappropriate Media Limited.

Public material demonstrates the product boundary and evidenced behaviour.

It does not disclose private source code, scoring formulas, thresholds or protected runtime internals.

Technical access is provided only through an appropriate evaluation, licensing or development arrangement.

Paid diagnostic evaluations

We are opening a limited number of bounded commercial evaluations for teams with an existing decision point where prior state may reasonably affect a later permitted choice.

A typical diagnostic can examine:

• one bounded decision point;
• a small permitted action set;
• a host-declared reference action;
• relevant retained-state factors;
• the resulting CAAI selection;
• customer-safe Decision Record evidence.

The purpose is not to replace the buyer's architecture.

It is to answer a narrower question:

Does retained-state-sensitive selection add useful behaviour and evidence at this decision boundary?

Where appropriate, a successful diagnostic can lead to a larger managed evaluation or integration discussion.

Licensing and evaluation

Collapse Aware AI is available for discussions with developers, studios, simulation teams, enterprise-agent builders and organisations working with bounded operational decisions.

If your system already has permitted candidate actions and previous state should influence which one is selected, we are interested in evaluating that boundary.