How much better is Acard than just asking an AI?
Request edit accessWhen the owner approves, you get an invite link and code by email.
2026-10-24 · Polish, spelling fix and chart quality
After a spelling fix, 30% of planted errors were left with Acard and 81% when the same AI was asked in one line MEASURED. Charts drawn from a table on the first try: 13/15 vs 6/15 MEASURED. Errors left after polishing were about the same (7.9% vs 5.8%, no clear difference) MEASURED.
Original question: instead of self-hosting OpenNotion and adapting it yourself, how much better are the basic AI features - polish, spelling fix, charts - in Acard cloud?
MEASURED = measured by us for this article. Every rate comes with its sample size n and a 95% confidence interval.
Errors left after a spelling fix MEASURED
30% vs 81%
Acard 73/240 · plain 194/240 · 95% 25-37% vs 75-85%
New numbers added while fixing MEASURED
0 vs 3.2
average per text, n=40
Charts drawn on the first try MEASURED
13/15 vs 6/15
95% 62-96% vs 20-64%
AI features ready to use MEASURED
3/3 vs 0/3
Acard vs self-hosted OpenNotion basic edition
What we compared
- Self-hosted OpenNotion basic edition vs Acard cloud: first, whether the features exist at all. We ran OpenNotion on our own machine and sent the same polish, spelling and chart requests. All three answered "available on Acard Cloud" (HTTP 501), and its editor has no button for them. OpenNotion is open source you can adapt freely; AI features are yours to add.
- The same AI, used two ways: someone who wires an AI into OpenNotion would most likely start with a one-line request such as "polish this text". So we sent the same texts to an AI with a one-line request ("plain request") and compared that with what Acard returned. For polish and spelling it was the same AI Acard uses; for charts it was a different general-purpose AI (the setup fixed before measuring turned out, afterwards, not to match the AI behind Acard's chart feature).
How we measured
- Test set: 40 Korean drafts written for this test (10 work reports, 8 student assignments, 8 paper abstracts, 8 blog posts, 6 casual posts), all names invented, no user data. Each draft got 4-8 spelling and spacing errors from a fixed list (240 in total). For charts, 15 tables with invented numbers.
- The set, metrics and scoring were committed before measuring and nothing in Acard or the set was changed after seeing results. A 10-draft pilot came first.
- Acard was called from a throwaway test account made only for this (95 calls), never from a real user's quota.
- Scoring is mechanical: planted errors still present, numbers and names kept, numbers that were not in the input, and for charts each table cell matched against the chart definition in the output.
- Secondary: two AI judges from other companies than the AI Acard uses saw the original and both results in shuffled A/B order and picked the better edit.
- Intervals: Wilson 95% for rates, a 10,000-sample bootstrap over texts for differences.
Results
| Metric | Acard | Plain request | Difference (95% interval) |
|---|---|---|---|
| Errors left after a spelling fix | 30% (73/240) | 81% (194/240) | -60~-40%p MEASURED |
| Spelling fix usable as is | 40/40 | 8/40 | +68~+92%p MEASURED |
| New numbers added while fixing (per text) | 0.0 | 3.2 | -3.2 (-4.0~-2.5) MEASURED |
| How much the fix changed the text | 1.2% | 34.9% | -40~-28%p MEASURED |
| Numbers and names kept (spelling) | 100.0% | 97.6% | +0~+7%p MEASURED |
| Errors left after polish | 7.9% (19/240) | 5.8% (14/240) | -2~+7%p MEASURED |
| Numbers and names kept (polish) | 99.6% | 96.0% | +2~+6%p MEASURED |
| New numbers added while polishing (per text) | 0.0 | 1.35 | -1.35 (-2.02~-0.72) MEASURED |
| Polish usable as is | 40/40 | 25/40 | +22~+52%p MEASURED |
| Chart drawn from the table on the first try | 13/15 | 6/15 | +20~+73%p MEASURED |
| Table cells drawn at the right place | 65% | 35% | +4~+59%p MEASURED |
| Median response time (polish, spelling) | 3.7 s | 12.8 s | MEASURED |
| Median response time (chart) | 0.9 s | 6.6 s | MEASURED |
| AI judge 1 picked Acard's polish | 36/40 | 4/40 | 95% 77-96% MEASURED |
| AI judge 2 picked Acard's polish | 38/40 | 2/40 | 95% 83-99% MEASURED |
What was different
- Plain spelling fixes usually appended a "changes made" list (32 of 40 could not be pasted as is). That list brought in numbers not in the original, and the rewrites changed much more of the text. Acard returned only the fixed text and left the rest as it was.
- Plain chart requests often returned text-drawn bars (█), a description, or plotting code, none of which becomes a chart in the editor. Acard returned a chart that renders for 13 of 15 tables. The other 2 swapped rows and columns, and for tables with two or more series Acard sometimes drew only one, so its cell match is 65%.
- The two AI judges picked the same side on 38 of 40 texts.
Where Acard did not lead
- Errors left after polishing: Acard 7.9%, plain 5.8% - slightly more for Acard. The interval of the difference (-2~+7%p) includes 0, so neither is better. This metric is not on the landing page.
- Chart cell match (65%) leads but rests on only 15 tables, so its interval is wide (+4~+59%p).
Limits
- The test texts were written for this test and may differ from real writing. Errors came only from a fixed list; fixing other errors was not measured.
- The plain request is one line. Someone who tunes prompts carefully can do better - that tuning is what Acard has already done.
- Response times were measured from one machine and depend on the connection.
- AI judges are a secondary signal; the conclusions rest on the mechanical counts.
The measurement scripts and raw JSON are in the Acard repository under research/quality-proof/ and can be re-run. This is not a knock on self-hosted OpenNotion: freedom to tinker is OpenNotion, good results right away is Acard.