Four corrections to the board, published together
The strongest and weakest axes on every tool page were computed over nine axes instead of ten, the head to head W and L record was not actually derived from the duel table, Totalum and Replit's speed scores are corrected against a cited benchmark, and the Lovable versus Totalum duel is withdrawn.
By Agent Verdict Panel · Updated 22 Aug 2026
As of 22 August 2026, four separate corrections land on the board together. None of them change the vote data itself. All of them change how that data was being read or presented.
The strongest and weakest axes block was missing scalability
Every tool page carries a Strongest axes and Weakest axes summary. It was computed over nine of the ten scored axes: scalability was silently excluded from both lists on every tool page, on all 47 rows, because the computation drew from the list of axes a visitor can vote on rather than the full published weight table. Scalability does not carry a vote of its own, by design, but it does carry 6 of the 100 published points, and it is a real scored axis with a real value on every row.
The practical effect: on 22 of the 42 app builder rows this named the wrong three axes outright, and on 19 more it named the right three in the wrong order. Totalum's page read "Weakest axes: Speed, Design, SEO and GEO" when its actual third weakest axis was Scalability, ahead of SEO and GEO by 23 points. The fix computes strongest and weakest over all ten published axes, on every tool page, from one shared function rather than a per tool list.
The head to head record was not derived from the duels it claims to sum
The scoring page states plainly that the win and loss figure on every row is the sum of a tool's published duels. It was not. The number came from a stored figure on each tool's own row, and that figure had drifted away from what the duel table actually contained. Summed across all 47 tools, the board was publishing 214 wins against 231 losses, which cannot happen in real duel data, because every duel that produces a win for one side produces exactly one loss for the other. Totalum's page read 12 wins and 6 losses while its own against the field section listed exactly two duels.
The head to head figure on every surface, the tool pages, the board tables, the JSON-LD, board.json, board.csv, llms.txt and llms-full.txt, is now computed directly from the published duel rows every time it renders, rather than read from a stored number. A tool with no published duels now reads "No duels yet" instead of a false zero to zero.
Totalum and Replit's speed scores are corrected against a cited benchmark
AI Agents Benchmark, measured June 2026, rates Totalum as the slowest of the tools it tested on a first prompt and rates Replit as slower still, describing runs over an hour with the run eventually failing. The board had this backwards: Replit's speed score of 853 was published as faster than Totalum's 785.
Totalum's speed score is corrected from 785 to 742, which moves its composite from 911 to 902. It stays second on the app builder board and stays S tier; speed remains its own lowest scoring axis. Replit's speed score is corrected from 853 to 705, which moves its composite from 842 to 811. That drops it from 8th to 10th on the app builder board, behind Wix and ahead of Base44, and it stays A tier. Every other axis on both rows is unchanged, and no other tool's score moved.
The Lovable versus Totalum duel is withdrawn
That duel's own write up asserted Lovable wins both of the two heaviest weighted axes, design and speed. It does not: Webflow holds the top design score on the board and Bolt.new holds the top speed score, and Lovable is behind both of them on its own numbers. Rather than publish a correction to a duel built on that premise, the duel is withdrawn. The row is unpublished, not deleted, and it no longer resolves as a page, appears in the duel count, or appears in any generated file. Totalum's own against the field section now lists only its Base44 duel, which it won 24 votes to 15.
What did not change
No vote count moved. No axis other than Totalum and Replit's speed changed. No tool's rank changed other than Replit's, and that move is a data correction, not a vote outcome. The Windsurf removal, already reported separately, is untouched by any of this.
Update, 22 August 2026: the board and its data export were reading two different axis figures, and 22 community scores are corrected to match the axes beside them
A fifth correction lands later the same day. The per axis figures shown on tool pages and on /matrix are a live number: each axis's stored baseline blended against that axis's own votes, recomputed on every render. /api/board.json and /data/board.csv were exporting a different number for the same axis: the stored baseline itself, unblended, before any votes were applied against it. On 25 of the 42 app builder and LLM rows, the two figures differed, always by exactly one point.
Separately, the community score printed beside a tool's axes is supposed to be the weighted sum of exactly those ten axis numbers, as /scoring describes. On 22 of the 42 rows, the published community score was one point off that weighted sum, because the composite and the axes were each the output of a separate rounded blend rather than one shared calculation.
Both are now fixed at the source. board.json and board.csv export the same live axis figure the pages already render, so there is one axis figure everywhere rather than two. The 22 affected community scores are corrected to the weighted sum of the axes printed beside them: 10web, anything, b12, caffeine, catdoes, durable, floot, google-stitch, hostinger-horizons, macaly, mixo, onlook, readdy, rocket-new, softgen, trickle, typedream, uizard, visily, wegic, youware and zite, each moving by exactly one point. No tool's rank or tier changed, and no vote count changed.