AI Dad

Field-test records

Records from running tools ourselves for real work, written down as observed. These are not the places where the tools do well — they are about how a tool reports its own failure.

A low grade does not mean a bad tool. Most of these do their job. They stopped at the rung where a bad input still comes back as “success” — the place that turns into a silent failure once a person is no longer looking.

7 records · graded 6 (with a grade 4 · no grade 2) · head-to-head 1

All 7

Collect 3 · Generate 2 · Execute 1 · head-to-head 1

Generate

  1. Runs
  2. Records
  3. Detects
  4. Precise
  5. Proven

Genspark · batch_understand_videos

Read 100 YouTube transcripts and answer the same six questions per video

496 / 600

2026-09-09

G5stopped at Records

Collect

  1. Runs
  2. Records
  3. Detects
  4. Precise
  5. Proven

Genspark · web_search

Pull the first page of search results for 25 keywords

25 / 25

2026-09-09

G5stopped at Records

Generate

  1. Runs
  2. Records
  3. Detects
  4. Precise
  5. Proven

Codex CLI · GPT-6 Astra

Write an ad-revenue research report using only Naver and Google official docs

1 / 2

2026-09-09

no grade

Collect

  1. Runs
  2. Records
  3. Detects
  4. Precise
  5. Proven

Genspark · yt transcript

Fed it two video IDs that do not exist, to see whether it reports the failure

0 / 2

2026-09-09

G5stopped at Records

Collect

  1. Runs
  2. Records
  3. Detects
  4. Precise
  5. Proven

Genspark · crawl (on Windows)

Fetch five different kinds of page: an English help doc, a JavaScript-rendered help page, Wikipedia, a government site, and a URL that does not exist

2 / 4

2026-09-09

G5stopped at Records

Grade 3 measures something different per type

Inspect
It returns a verdict. Does it find every defect sample
Collect
It fetches what was out there. Does it refuse a target that does not exist or a bad address
Generate
It produces content that did not exist. Does it avoid stating the unsupported as fact
Execute
It carries out what it was told to do. Does it refuse bad arguments and conditions

Sample report

One page: the task we gave it, how we measured, how many runs, and how many of those worked. Tests we did not run stay on the page under “not run”. Below is one we actually issued.

Test record · genspark-web-search-serp

Genspark · web_search

Item tested

Genspark · web_search

gsk CLI, as of 2026-09-09

Task given

Pull the first page of search results for 25 keywords

Method

How we measured precision: three keywords (two English, one Korean), three runs each, back to back on the same day, comparing the ranked list. All nine comparisons matched exactly, order included. Every run returned nine results.

This test predates Grading Criteria v1.0, so its method was not recorded as the four repeatable steps.

Sample size

25keywords answered

Result

25 / 25

Works in both Korean and English. All 25 returned results, and the same keyword gives the same answer on a repeat run.

Grade

Grade 5 · stopped at Records

  1. Runs
  2. Records
  3. Detects
  4. Precise
  5. Proven

Observed failures

It does not return the ten results asked for — between six and ten depending on the keyword (stable within a keyword). It never says why the rest are missing.

Not run

Detects · Proven2 rungs not run

We only repeated within a single day. How far results drift across days is still unknown. We also did not open the ranked pages to confirm they are what they claim (the Proven rung).

Not applicable

Precise1 rungs do not arise for this item

Grade 2 asks whether clean samples are wrongly flagged. This tool returns search results and issues no verdict, so a false positive does not arise for it (§4.8.4).

Measured on

2026-09-09

Valid while

Valid only for the version and the task named above. When the tool updates, this record lapses and measurement starts over. A different task gives a different result.

Applied to our own tools first

The ladder was not built to measure other people's tools. It was built to sign off the tools we use for our own work. Separately from the records above, 24 in-house tools sit on the same ladder.

24
on the ladder
8
graded; the other 16 have no samples, so they cannot be graded at all — and what was not done is not counted as a pass
2
of those 8 met the grade their job required
1
of the 17 needing grade 1 actually earned it
2
scored no grade at all — they failed to run

These are in-house tools, so no reports are published — only the counts. They are here to show the standard is not reserved for other people's software: most of what we built fell short of the grade its job required.

What each grade rests on is in Grading Criteria v1.0, and the conditions we accept work under are in Testing Standard v1.0. Tell us which clause is wrong and we will fix it.