Baltor Get started

Case study

Data cleanup with and without Baltor.

We measured whether approved library items, placed in each step, help a small or cheap model clean data. On this test they did not, and one made results worse.

Measured September 22 to 23, 2026 · 143 synthetic rows · 72 counted steps · Pi 0.73.1

Get startedSee a task step by step
  • 0 of 4families where the item was clearly better
  • 1 of 4clearly worse: phone numbers
  • 45 to 446%more prompt tokens per step with the item
  • 12 of 12steps the cheap model passed without any item

What we found

  • Worse on phone numbers

    With the phone item, glm-5.3-flash:cloud scored 0.842 in all three repetitions, against 1.000, 0.974 and 1.000 without it. The item's digit-count rule is incomplete, and the model followed it: it lost every United Kingdom number written with a (0) trunk prefix and every number dialled with 011.

  • More tokens, more time

    With material, the cheap model spent 45 to 446 percent more prompt tokens per step in every family, and more requests and more seconds per step.

  • Little room to help

    Without material, the cheap cloud model already averaged 0.948 to 1.000 in each family and passed 12 of 12 steps. The larger glm-5.2:cloud was not separated from it in any family.

  • The small local model could not drive the harness

    In all 37 of its steps, qwen2.5-coder:7b wrote its tool calls as plain text, so Pi ended each step with no output file. It scored 0 with and without material.

Results by family

FamilyWithout the itemWith the itemPrompt tokens per stepVerdict
Phone numbers0.991, 3 of 3 passed0.842, 0 of 3 passed38,709 without, 87,038 withClearly worse
Email addresses0.952, 3 of 3 passed0.943, 3 of 3 passed71,927 without, 104,448 withNot separated
Postal addresses0.948, 3 of 3 passed0.969, 3 of 3 passed17,446 without, 36,041 withNot separated
Duplicate records1.000, 3 of 3 passed0.991, 3 of 3 passed22,135 without, 120,868 withNot separated

The cheap cloud model, glm-5.3-flash:cloud through Ollama Cloud, three repetitions each: record accuracy for phones, emails and addresses, and pair F1 for duplicates. Clearly worse means every repetition with the item scored below every repetition without it.

What changed because of it

We stopped showing the phone item, and no demonstration on this site may use an item that a measurement found harmful.

What it does not show

It does not show that library material improves accuracy, lets a small model match a larger one, or saves tokens, time or money. It says nothing about data, models or harnesses that were not tested.

Limits and method

The 143 rows are synthetic, written by the study's author, who knew the items. The design was frozen before the first model call, each step ran three times in a fresh Pi process, and an independent scorer checked every output. The full report, every failure included, is in the public repository.