Block-level LLMs

Block-Level LLMs, the Same Authority

AI embedded in each block of a lesson, on your own model or ours, answering from your field’s own library.

Syllabium is an enterprise learning management system (LMS) for regulated industries, made by Athantic AI LLC.

An LLM in an LMS is a large language model working inside a learning management system: an AI model that reads text and writes text in reply, put to work on the system’s courses. In Syllabium, LLMs are embedded at the level of each block of a lesson, and every request to one passes the same authority checks as every other action.

What Is a Block of a Lesson?

A block is the smallest unit of content in Syllabium: one piece of a slide, such as a passage of text, an image or embedded media.

Content is built in levels, each inside the next: curriculum, course, module, lesson, slide and block.

LevelWhat it is
CurriculumA program of courses, each drawn from the library at a pinned version, with access, display and pacing set per entry.
CourseA unit with its own outline, objectives, final assessment and survey.
ModuleA part of a course, which can carry its own quiz.
LessonRead slide by slide; it completes when every slide has been shown.
SlideOne screen of a lesson, made of blocks on a layout.
BlockOne piece of a slide: text, an image or embedded media.

Authors insert, edit and remove blocks on the slide canvas, and versions of library content can be compared block by block. Platform describes the studio and the other products.

How does an LLM work inside a block?

Every AI request in Syllabium is a defined action on named content, down to a single block, and the platform assembles what the model is sent.

Authority can reach down to the block: a grant can be limited to one library or to one block. Each AI act on a block writes its own row in the action log, and staff can see a block’s AI activity in a view of its own. In the authoring studio, the assistant sits beside the editors in one of four modes, and each author chooses the mode.

ModeWhat the assistant does
OffThe assistant is switched off.
AskIt answers, and writes nothing.
Suggest, the defaultIt shows each change before and after, and the person applies it.
ActIt writes into the author’s own draft, recorded, with undo.

The assistant works on a block’s text and makes no bulk changes. It can also propose a new order for courses, modules, lessons and slides, and each change it files waits for a person by default.

Which Checks Does a Block-Level Request Pass?

A block-level request passes the same twelve checks, in a fixed order, as every other AI act, and it takes the same path a person’s action takes.

StepWhat it means for a request on a block
Emergency stopIt is read before anything else, and a stop in force halts the request.
AuthorityThe agent must hold this capability, at the level of authority the act needs.
ScopeThe grant must cover this library, or this block.
Safety-critical ruleOn content flagged safety-critical, the work waits for a person.
LimitsThe act must fall within its rate and spend limits.
DataOnly data of a kind allowed for this model’s provider may be sent.
The actionThe act runs by the same path a person’s action takes.
The recordThe act is recorded, attributed to its agent, and chained.

All twelve checks run on every request, and a refusal names the check that decided it, is recorded and never reaches a model. AI agents in training lists all twelve, in order.

What happens to a block the AI has changed?

It stays a draft until a person takes it forward, and like every change to library content it then passes the one publishing gate, which publishes, holds or refuses it and records the decision.

  • The author is told the gate’s outcome before clicking.
  • The studio marks content the assistant changed, and its content tree marks AI activity, safety-critical content and other people’s changes.
  • Staff see “(AI)” and the agent’s role beside its work.
  • Held changes are compared word by word, then approved, rejected with a reason, withdrawn or revised.
  • A change to content flagged safety-critical stays hidden from learners until a second person approves it, and the person who submitted it cannot approve it.
  • The assistant in the studio drafts and does not publish; a publishing path of its own is planned.

AI review before publish explains the review in full.

How Is Each Model Call Recorded?

Every model call is logged with its model, tokens and cost, in tamper-evident records kept seven years.

EntryWhat it holds
Who actedThe agent, by its own account and role.
The checkThe check that decided the attempt.
The outcomeAllowed, refused or failed.
The modelThe model that answered the call.
TokensThe tokens the call used.
CostWhat the call cost.
Left outNo prompt, no answer and no person’s words.

Tokens are the small pieces of text a model reads and writes, and model use is counted in them. Each entry carries a fingerprint of the one before it. Change one, and every later fingerprint stops matching. Every chain is recomputed nightly and on demand, and once a day the newest fingerprints go to write-once storage outside the database, locked for seven years. Evidence explains the records and the export for an auditor.

How Are Model Costs Limited?

Each agent has limits on actions and tokens an hour and on spend a day and a month, and every call carries a cost cap, estimated before the call is made.

A cap across all agents applies at each level: the platform, each tenant and each organization. A warning email goes out at 80 percent of a budget, and calls are refused at 100 percent. Administrators see AI spend on the console’s dashboard.

Which Models Run Inside the Blocks?

Syllabium uses Claude models, through Amazon Bedrock and Anthropic's API. Where your organization brings its own model, its block-level requests go to that model instead.

The offering is one whole: embedded block-level LLMs (BYOM) with integrated agentic RAG. A request goes by the route set for its organization, else its tenant’s, else the platform’s, and where no route is set, no call is made. Your own models explains bring your own model, and Agentic RAG how answers are cited.

What Do Learners See of the AI?

A learner can ask the assistant to explain a slide, and before any learner page, each learner accepts the documents in force, among them a notice of how AI is used.

A learner’s identity is never sent to a model. A learner’s reflection goes to a model only under the enrollment consent, and without the learner’s name, email, user ID or cohort.

What is on the roadmap for block-level AI?

AI drafts of questions, of alt text and captions for media, and suggestions while a curriculum is composed are planned. Governed AI shows the one route every AI action passes.

Get Started

Enroll your organization online, see the pricing, or talk to us. Syllabium is made and operated by Athantic AI LLC.

Contact us