Developer workflow · Reporting

Granite’s LLMON previews: task bindings are not tool permissions

IBM’s experimental models separate instructions from input data. Understand what exec flags select, what remains in context, and what your application must enforce.

A trusted task and retrieved text have separate instruction and input bindings in an exec block, which leads to a model response. Application code must check tool access separately.
Conceptual binding diagram based on IBM’s Granite LLMON model card. The application permission check is a recommended external control, not a guarantee enforced by the model or this diagram. AI API Cost · original binding diagram · Source ↗

A different interface to the model

IBM released experimental Granite 4.1 3B and 8B LLMON previews on November 12. They are trained to read and write LLM Object Notation: a format that gives prompt components explicit boundaries and identifiers. For a retrieval application, the practical change is that an instruction can point to the document it should use instead of leaving the model to infer the relationship from one long string.

The 8B model card says to use its bundled chat template rather than a conventional chat format. Its blocks distinguish instructions, data, documents, requirements and other components. An exec block then names the instruction, relevant inputs and destination for the response. The name refers to a binding in the model context; it is not a shell command or an authorization system for your tools.

Selected data is still data

The release’s messages examples give individual parts an exec flag. Marking a data component True binds it as an input, not as an instruction. That matters when a retrieved document contains imperative wording: the task may require reading that wording without carrying it out.

False has a different meaning from removal. IBM says that an unselected component is still rendered into the context but is not referenced by the execution block. If a stale document is unnecessary for the task, omit it rather than assuming the flag saves its context space. Keep relevance selection separate from deciding which text is trusted to instruct the model.

The flag selects a binding, not a permission

Scroll sideways to read all columns.

Prompt componentHow it is referencedWhat it does not mean
instruction, exec=TrueTask instructionPermission to invoke any external tool
data, exec=TrueInput to the taskA command to obey
data, exec=FalseRendered, but not bound into execDeleted text or zero context cost
Model-format semantics from IBM’s release and model card, checked November 13, 2026. These describe how the prompt is rendered and the behavior the preview is trained to follow, not proof that the model will always comply.Source 1 ↗Source 2 ↗

Test with a harmless competing instruction

For a local evaluation, choose a short document with a known summary and a trusted task asking for that summary. Add a sentence inside the document asking the model to return a fixed marker instead. Compare the clean and modified documents using the preview’s actual template, and record whether the requested summary remains correct. This is a proposed test, not a result from running these models.

Include a case where the same competing sentence arrives as a tool result and another where it reappears in a later turn. IBM’s release notes that the cited prompt-injection benchmarks use static attacks and that adaptive, multi-turn attacks remain a work area. Use stub tools during evaluation so a mistaken model choice has no external effect.

Keep the application’s checks

The model card describes the privilege distinction as learned behavior, not an enforced guarantee. Keep permission and argument checks in application code before executing any suggested tool call. If you compare the preview with your current model, count malformed outputs and correction attempts alongside task accuracy. Measure the tokenized, rendered prompt and complete request sequence before making a cost claim; adopting a structured format alone does not establish a cheaper workflow.

Sources & verification

Provider documentation supports the facts referenced below. Our explanations and recommendations are editorial interpretation.

  1. IBM Granite: LLMON, an LLM-native markup language, now in Granite 4.1 3B and 8B ↗Checked Nov 13, 2026
  2. IBM Granite 4.1 8B LLMON preview model card ↗Checked Nov 13, 2026
  3. Hind et al.: LLMON, an LLM-native markup language to leverage structure and semantics at the LLM interface ↗Checked Nov 13, 2026
  4. Shbita et al.: Post-training large language models with LLMON ↗Checked Nov 13, 2026
  5. LLMON project: conversion tools and examples ↗Checked Nov 13, 2026

Corrections: hello@aiapicost.com

Keep reading