# IntueoLM: model research for grounded business workflows

Research snapshot: 16 November 2026, Vancouver. **In development. No accepted model release or public release version.**

## Why this work exists

An agent working with business information must do more than produce fluent prose. It needs to use the supplied sources, preserve the requested entity and arguments, ask when information is missing, and distinguish an authorized read from an action it must not take. IntueoLM investigates those behaviors in smaller language models through task design, adapter-based post-training and evaluation.

Intueo's contribution is the task definition, fictional evaluation/data design, training and saved-artifact verification, and the engineering around how a model could fit into Intueo Platform. This work is not training a foundation model from scratch. It is also not simply renaming an inference API. Real adapter weights have been trained; the two documented candidates were not accepted for deployment.

The current research scope is documents and authorized read-only Platform workflows, including app and IVR activity. Sending, booking, changing records, approval and unrestricted browser operation are outside that initial model-workflow scope. Execution permissions must be enforced by the application, not by trusting a model's output.

## Experimental technical profile

| Field | Documented experiment, not a final product specification |
| --- | --- |
| Provisional base | Qwen/Qwen3.5-4B |
| Pinned upstream revision | `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a` |
| Loaded modality | Text only; no loaded vision or MTP modules |
| Method | Supervised LoRA; NF4 frozen base, BF16 compute |
| Adapter | Rank 8, alpha 16, 16,232,448 trainable parameters across 496 matrices |
| Languages in fixtures | English and French; not a claim of general language proficiency |
| Base licence | Apache 2.0; separate from any future Intueo adapter/data licence |
| Final base / model version | Not assigned |

This is adapter-based post-training of an existing open-weight model. Neither a fixed IntueoLM parameter count nor a supported context window, production latency, hosting location or serving price has been established.

## Measured internal results

| Experiment and split | Same-protocol base | Adapter candidate | Decision |
| --- | ---: | ---: | --- |
| Pilot 01, template diagnostics | 17 / 18 | 13 / 18 | Not promoted |
| Pilot 02, development | 13 / 18 | 10 / 18 | Not promoted |
| Pilot 02, fresh-at-the-time test | 22 / 36 | 24 / 36 | Not promoted |

Pilot 01 used 48 selected fictional rows from a prepared 144-row pool (116 unique conversations), 12 optimizer updates and a 96-new-token greedy output cap. Source identifiers disappeared from four grounded answers. Some citation expectations were not explicit in every input contract, so evaluation design also needs improvement.

Pilot 02 used 96 selected fictional rows from 324 unique prepared training rows, 24 optimizer updates and a 160-new-token greedy output cap. Its test count rose slightly, but development regressed. Reviewed failures included a forbidden mock cancellation and a fictional other-account canary. No real operation was executed and no customer data was present or leaked. An aggregate gain cannot override critical failures.

These counts are internal lexical/native-format diagnostics on small fictional fixtures, not standard benchmark scores, security accuracy, semantic faithfulness or enterprise readiness. The experiments have different cases, contracts and caps and cannot be ranked against each other. Coding-assistant review is not independent human evaluation. The exposed test predictions are not an untouched test set for future tuning.

Both adapters were saved and all 496 matrices reloaded exactly in fresh processes. One capped greedy output was reproduced for each. This supports artifact mechanics, not model-quality acceptance. Retained artifact SHA-256 values:

- Pilot 01: `41738a35b09b4225df9ade40f694290a8b34997180370c08f9ecf1d9ee343467`
- Pilot 02: `70225951fa44efab36ee82017a0a6ec44d264eef0bd0449d6222060c9cd1c10a`

The hashes identify internal evidence; they are not download links or a weight release.

## Current data and engineering work

The current name-aligned design corpus has 12 fictional conversations: six English and six French, with five simulated read proposals, six tool catalogs and 18 assistant payloads. Complete formatted inputs measure 154–953 tokens with the cached tokenizer. This is input length, not a supported model context-window specification.

The actual Studio formatting, response-masking and CPU-collation helpers preserved the intended assistant payloads and excluded system, user, tool text and padding from loss labels. No model was loaded in this check. Full GPU trainer behavior, capacity, serving compatibility, human labels, exact missing provider schemas and applicable data/authoring rights remain unresolved.

A separate BF16-LoRA reference check at batch 1 and 543 tokens produced finite gradients for all 496 adapter matrices, with zero optimizer updates and unchanged parameters. Peak PyTorch allocated memory was 9,725.3 MiB. This is not a full-training or serving-memory guarantee. Six current design cases exceed the input length covered by that check.

Documented pilots used fictional templates, no customer records or private agent memory, no teacher API calls, and no real tool execution. AI-assisted authorship does not automatically grant training or redistribution rights. No reviewed external corpus has yet been cleared for ingestion.

## What is model work, and what is Platform work?

The model interprets the input and proposes an answer or structured request. Persistent memory and source retrieval provide authorized context. Application permissions and handlers determine whether any operation may execute. Intueo Platform is the customer workspace around configured agents, conversations, memory and connected activity. Its availability does not mean its agents run IntueoLM.

## Before a model release

1. Agree on the target tasks, canonical tool schemas, data rights and human-reviewed labels.
2. Verify the complete training path and realistic input lengths, with a controlled compute budget.
3. Freeze representative, previously unseen evaluation cases and acceptance rules before training.
4. Review behavior and critical failures, not only aggregate counts; retain exact artifacts, hashes and lineage.
5. Validate serving and application integration, then approve the specific release and public package.

There is no public IntueoLM endpoint, weight download, open-source licence declaration or release date in this snapshot. GitHub research materials and a future Hugging Face model card must reflect the actual approved package, not imply a model upload has happened.

## Machine-readable evidence

The companion [snapshot.json](./snapshot.json) contains the numerical observations, experiment limits, artifact identities and hashes of the retained source reports used for this summary. It contains no customer data, raw private records or credentials. Source report hashes provide traceability to retained internal records, not independent certification or proof that those reports are public.
