Experiment · EXP-006

Should the model compute, or should code compute?

Internal model arithmetic versus plan-then-execute on spreadsheet-like tasks.

The Decide study made a general experiment precise: language models can plan transformations, but deterministic execution is a better owner of computation. I treat “LLM plans, code computes, system verifies” as the positive alternative to asking the model to be the spreadsheet.

Hypothesis: The model can both understand the request and perform the calculations internally.

Research question

When an agent understands a spreadsheet or analytical request, should the language model also own the underlying calculations and data transformations?

Architecture studied

The more reliable pattern is to use the language model for intent interpretation and planning, then hand computation to deterministic code. A spreadsheet request can be decomposed into operations, executed against the workbook, and then verified against the requested result.

Computational agent loop

Spreadsheet
  → Preprocessing
  → Structure detection
  → AI planning
  → Code generation
  → Secure execution
  → Verification
  → Updated workbook or analysis

If the agent misunderstands workbook structure, every later operation can be wrong. Giving a model a computational tool can matter more than asking it to simulate arithmetic. This page is an architecture study of Decide as well as a general experiment: the founder publicly responded that the breakdown had captured the system accurately. That is an endorsement of the analysis, not a substitute for Decide’s internal benchmarks, and it is not a claim that I built Decide.

Why this separation matters

Language models are useful at converting ambiguous instructions into plans. Code is better suited to arithmetic, filtering, grouping, sorting, workbook mutation, and repeatable transformations. Asking one component to own both reasoning and computation removes a useful correctness boundary.

Result

LLM plans. Code computes. System verifies.

I now treat that as a general design rule for agentic systems that operate over structured data. The model decides what should happen, tools perform operations that have deterministic implementations, and another layer checks whether the requested state was reached.

Limitations

This result combines my architecture study of spreadsheet agents with a broader engineering principle. It does not reproduce any private benchmark or unpublished metric from a third-party product.

Abubakar, Imam. "Should the model compute, or should code compute?." Imam Abubakar AI Research, 2026. https://imamabubakar.com/research/exp-006-model-as-calculator

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.