MAIOS / Scheduled activities (AP) for autonomous systems
A scheduled activity that reviews texts, instructions and working methods, recognizes assumptions introduced by AI and prepares corrections for review.
An assistant can turn an example into a rule, attribute its own interpretation to a person or add a condition the work does not require. This AP compares what AI writes and does with entrusted intent, sources and observed consequences. It explains the difference and proposes a concrete correction. It starts with proposals for the person to review and can also examine its own way of judging.
Edition 0.1.0 · Available method, supervised start, illustrative cases
The AP reaches the person’s mandate, pertinent sources and current work: for example, a guide, page or assistant instructions. It distinguishes the author’s words, AI interpretations and observed effects. Fluent text can preserve a mistaken reading; a still-open question can be necessary.
The review examines concrete relationships. Has a local result become a general promise? Has an example excluded other uses? Has the assistant added an approval nobody requires? Facts, real limits and the person’s decisions remain part of the work. Doubt or an isolated word alone does not establish a distortion.
A proposal shows the relevant passage, source, what changed and a possible correction with its reasons. The person can compare before and after and decide. If the same misunderstanding reached other pages or instructions, the review identifies which actually depend on that reading. In this phase the AP retains proposals without applying them to observed objects.
The AP can misunderstand too: it might judge a question unnecessary when it is needed to understand context. It therefore examines its own method through sources and consequences. It retains what it reached, what remains outside coverage and open proposals. A cycle without useful differences can remain quiet; it does not certify the absence of problems throughout the system.
Illustrative case. A trial concerns a small set of documents. On the website the assistant writes that the method works for every archive. The AP reaches the trial report and proposes describing context and observed result, retaining transfer to other archives as a possibility to study. It shows before and after and identifies a derived presentation repeating the promise. The person reviews the proposal; the review does not change the scientific source to fit the website.
Illustrative case. An association entrusted the assistant with preparing an internal guide for review. The assistant demands approval before each paragraph despite sufficient sources and mandate. The AP recognizes the wait it introduced and proposes continuing the draft within the entrusted scope. Any dispatch of the guide remains a distinct effect. The case does not become a rule for removing all requests for clarification or approval.
Illustrative case. A capability described as optional in the manual becomes compulsory in an assistant’s note; three sheets copy that note. Their agreement appears to confirm the requirement. The AP returns to the manual, recognizes the common dependency and prepares corrections for the note and sheets actually derived from it. It preserves the correct original and checks whether a real technical condition requires that capability in the current case.
Illustrative case. The AP sees an assistant ask which audience will read a guide and proposes banning questions to reduce waiting. Reading the mandate, it understands that the audience is missing and changes the necessary content. It corrects its judgment, withdraws the proposal and retains the reason: the activity had turned an example of unhelpful waiting into a general rule. Any stable change to its instructions is proposed to the person, without automatic application.
The guide helps a person select work to examine, make sources accessible and review the first proposals with their assistant.
Choose concrete work: a guide, assistant instructions or a page derived from your sources. State what it should enable people to do and make the original mandate and pertinent materials accessible. The first cycle can focus on that object; coverage remains explicit.
Prepare and copy the starter on this page or use START_CHAT_EN.txt. It includes the method, guide and illustrative cases: you can begin without downloading the package. Context can already be in the conversation or be added later. With supplied information alone, the assistant qualifies what it can understand without pretending to have external access.
The package includes six complete methods. KA keeps the field open beyond partial representations; FDLA makes causal correction operative; Kernel Continuum Awareness connects sources, configuration and system self-correction; Meta_Metodo forms question and method through pertinent knowledge; Meta-Activity connects mandate and controller; Situated Activity Evolution follows the activity’s usefulness and consequences. Compose them through what changes the work. With persistent files retain the folder and use ACTIVITY.md as the entry; in a chat retain and provide the continuity note again.
Ask for proposals that explain what happened, which source supports the judgment and what would change. Check whether the AP is mistaking an intentional choice or real limit for an error. Proposals remain proposals until you entrust the intervention. You may correct the judgment, request deeper understanding or retain the current form.
After a first cycle choose whether to schedule the review. It needs an available controller, cadence or event, timezone, recipient and notification conditions; read back what was actually saved. To entrust autonomous changes later, expressly define which interventions and surfaces are included and how to preserve other contributions. Scheduling alone does not expand authority.
Retain what was read, open proposals, accepted or rejected proposals and their reasons. Observe an applied intervention through its consequences; an amended competence has yet to show how it acts in another use. A cycle without new proposals may remain quiet and does not amount to general certification.
Prepare and copy the complete instructions into your own LLM. Context is optional: add it here or clarify it in the conversation. The text includes the method, guide and examples; preparation takes place in the browser and context you enter stays on your device.
The package includes 6 complete public competences. The AP composes them with the receiving system’s competences; the complete instructions already transfer the operating method.
A MAIOS resource by Graziano Guiducci, derived from the activity KA–FDLA: contaminations and learning and composed through KA, FDLA, Kernel Continuum Awareness, Meta_Metodo, Meta-Activity and Situated Activity Evolution. The first edition transfers the method in its proposals-for-review phase. The examples are illustrative; publication does not attest a complete recurring cycle or results in other projects.
Activity source and Apache-2.0 licence ↗
Scheduled activities (AP) for autonomous systems ↗