The Adaptation Engine

The New Skies Initiative is almost entirely automated, as the distance between the probes and the Solar System would require years of waiting for communications to cross, giving catastrophes plenty of time to appear. These distances so far and question marks so many mean that even the best-designed mission would require some ability to change to overcome challenges the designers could not have known. Thus, through tireless design and testing, the New Skies Initiative sports a powerful yet safe method for the machines to change their own designs. Self-modifying machines were old news at this point; making them robust enough for a long-term mission bore the real novelty. The Initiative soon featured a software framework so robust, forged through decades of trials, that it nigh guaranteed high productivity through time while quelling fears of losing control beyond a reasonable doubt. The international community revealed it to the public under a name so simple as to be aggrandizing, the Adaptation Engine.

The Adaptation Engine is not any physical thing nor any single piece of software. Instead, it describes the behavior the New Skies Initiative exhibits simply as a result of the divisions' designs and their interaction with other divisions and the base AI's. It is an emergent property of the system, verified through supercomputers running simulations for weeks on end, ensuring that the behavior of the entire Initiative will maintain certain behaviors. It is the sum of all behaviors of all the mission's parts, as well as their safeguards.

The Adaptation Engine, despite its complexity, is summarized easily. In essence, the sensors aboard Ternary Orbiters, modules, and especially robots will feed into base AIs as constant observations, where they may further travel to other bases. The base AIs are capable of recognizing when occurrences do or do not align with their goals, both in the near and long term. With these recognized, they have the option to modify the design of something they plan to manufacture, the extent of change varying with a calculated risk. This complex process may include trial and error, original insight, concept generation and self-scrutinizing, and decision making — all dependent on the base AI's personality, another changeable set of factors. This makes base AIs, summed with the efforts of the divisions feeding them data and feedback, sophisticated learners, able to plan ahead, learn from past mistakes or those of other bases, and generalize or reframe their approaches for new scenarios.

Of the three divisions feeding the base AI's data, not to mention base AI's sharing data with each other, robots do the lion's share. As they are only partially under their base's control, they are intelligent enough to react to unexpected scenarios and take long, arduous sequences of actions to reach their assigned objective, instead of relying on base AI's to tell them precisely what to do and how. Through robotic eyes, ears, and other sensors, bases observe their own progress. Every time a robot returns to its base, its memory logs will provide valuable fodder for the base's AI.

Finally, not all components of a design are as easy for a base to change as others. Some are immutable, such as the NSISO and its Verdant Objective; not even thousands of years of mutation could touch them. Others are extremely stable, like error detection and correction algorithms that allow the machines to communicate despite radio noise. Most, however, are somewhere between fairly stable, such as intertool sizing and standardization, and quite liberal traits, such as most robot and module parts besides computation, communication, and energy storage components.

The Classification and Terminology System

With the ability to modify themselves granted to them, inevitably some machine will deviate significantly from its original design. Thus, the machines are also able to reclassify themselves. Natural language processing is by now dirt cheap and present in gadgets large and small, and computation is faster and more efficient than ever, so every player in humanity's great historic reach into the stars, whether seen by human eyes or not, shall be fitted with their own name. This motivated the creation of the Classification and Terminology System, a database and set of algorithms that takes up practically no space on the high-tech storage of the Founders.

The highest classification of the Classification and Terminology System is the Primary System Division, as previously explained. Each division has within it various models of machine. A model is a research and development team's final product, a machine designed to fulfill a certain purpose in the best way the modern world will allow. Further, each model contains an array of sub-types known as variants, each differing in finer details from each other. Several of these variants come from home pre-designed, but others the Adaptation Engine can create to label significantly modified machines, turning each model's handful of variants into a rainbow of them. Lastly, individual machines themselves of the same model and variant are almost never identical; with cheap computation power, data storage, and extremely flexible manufacturing methods, base AI's essentially run a constant experiment. By producing slightly different machines and remembering what differs between them and why, base AI's can use feedback from their performance to measure the effectiveness of certain traits. Of course, the threshold AIs use to assign machines different variants or the same is arbitrary, but the AI's can judge it more objectively than any human ever could. Further, the AIs strongly tend to create changes in an unnatural, but very human, manner: Instead of smooth gradients of change between machines, the Adaptation Engine will, upon crossing the "new variant" threshold, allow itself to make more extreme changes, creating a pattern of variation that fits neatly into categories, just as humans please. This ensures the AI's create categories of things in an easy to understand way to a human, unlike the messy gradations that permeate the natural world.

As the variants the Adaptation Engine creates cannot be named by the humans watching so far away, the Classification and Terminology System can also name its new variants. To do this, the System takes everything new about the variant, from the reasoning behind its changes to its appearance from all angles, then generates a fitting name. The actual name generation is trivially simple by the standards of the day, based on a large language model trained specifically to do so. These usually are one or two words each, but may include acronyms or strings of digits, with the meaning behind them stored alongside the name.