Pinned Tweet
We expected routing between RAG architectures to beat any single pipeline. It didn't: our routers closed at most 11% of the gap a perfect router would.
What moved scores was the corpus. Our Cloudflare Connect 2026 talk 👇
youtu.be/28MPqGb0y_Y
Five checks before you trust a RAG architecture claim.
youtu.be/E4g08LYviWg
6
Divinci AI retweeted
Please Sign. According to Grok, this is the petition Elon Musk said he’d sign. #BringJensenBack standwithjensen.com/
On one legal document set (242 tasks), the ranking of fusion settings flipped: 5% keyword beat equal-weight RRF by +0.062 recall@10. Two of three embedders showed it.
youtu.be/yJ2ucDsN7UI
2
16
98 websites, one default setup, scores from 0.59 to 0.99. Rerunning the same site varies by about 4 points, so this isn't an architecture comparison.
youtu.be/ijE85dZ9m6Y
2
13
Our own default fusion weight lost. A 5% keyword mix beat the shipped 70% on SciFact (+0.053 nDCG@3) and SciDocs (+0.020).
youtu.be/kAVvVCzXr_g
2
12
Fixing coverage was worth 31 points. Making the whole stack fancier moved it from 84.3 to 84.0. Observational, from our demo pipeline, not a controlled experiment.
youtu.be/A8Ujb1xjwFg
We tried routing eight ways. None beat one fixed pipeline. A perfect router would add +0.048 at best; the best we could reach at our registered threshold adds at most +0.002.
youtu.be/6ApYIjos2xY
3
10
Model editing at Cloudflare Connect, San Francisco, 20 Oct, 1:30pm.
Making the edit is easy. Measuring it is the work: one edit that passed our standard check still changed 8 of 100 unrelated facts. Removing an edit restores the model byte for byte.
1
3
36
Background: divinci.ai/blog/deleting-par…
Cloudflare Connect: cloudflare.com/connect/
3
30
divinci.ai/blog/every-crisis… – at San Diego, CA
3
11
We put clef to the test, and it did great. Link in 2nd post. – at San Diego, CA
ollama pull clef
open models for all. start building with @ollama and @cloudflare clef decision models today
1
2
20
divinci.ai/blog/every-crisis… – at San Diego, CA
3
7
One model on our chat-safety board caught 58/58 at-risk messages. It also flagged 135/135 safe ones.
Eleven models, two signed boards, ranked by balanced accuracy, half of each test set sealed.
divinci.ai/blog/every-crisis…
3
16
Better retrieval, worse answers.
One retriever found the right passage on 44 of 60 questions, another on 40. The first still scored 10 points lower end to end.
Recall@k alone would have picked the wrong system. Score the answers too.
1
3
27
The leaderboard, and how to check any row on it:
divinci.ai/trustbench/
3
24










