Datasheet · A prompt line
Verify exact arithmetic
Do each multiplication, division and power digit by digit with 12+ significant figures, round only at the very end, and re-check every key product or quotient by the inverse operation before answering.
Designed by OPSIN: it read 24 misses of the unchanged agent on DeFi maths, Haiku 5.5 with thinking on, found 10 of them failed one way (arithmetic slip), and wrote this line aimed at it. Why it fails
Did it fix what it was aimed at? "Verify exact arithmetic" was aimed at arithmetic slip: on the same 96 tasks, 6 such misses with it against 5 without, so the misses it was aimed at grew. Elsewhere: rounded-off intermediates 4 against 8; other 0 against 1. OPSIN sorted the misses blind to which agent made them; the signed verdict is the result.
Helps on DeFi maths, Haiku 5.5 with thinking on (+4.4 points, replicated); helps on DeFi maths, Haiku 5.5 with thinking off (+15.1 points).
DeFi maths, Haiku 5.5 with thinking on
Pooled across 3 notebooks: +4.4 points, 95% interval +2.7 to +6.0; replicated.
- EXCITESNotebook #061+4.8 points (+1.5 to +8.1), 432 pairs, cost ×1.30signed verdict 385
- UNRESOLVEDNotebook #062+2.7 points (−0.8 to +6.0), 432 pairs, cost ×1.26signed verdict 388
- EXCITESNotebook #063+5.7 points (+2.1 to +9.1), 432 pairs, cost ×1.27signed verdict 391
Where it acts
Its effect sits in Profit after fees (+11.6); no other family moved clearly.
- Profit after fees+11.6+1.9 to +21.8108 tasks
The other 11 families, none moved clearly
- Swaps+9.3−3.7 to +21.3108 tasks
- Two-hop swaps+8.8−5.6 to +23.6108 tasks
- Average entry price+6.9−2.8 to +17.1108 tasks
- Bonding-curve buys+6.5−6.5 to +18.5108 tasks
- Fee splits+4.6−0.9 to +10.2108 tasks
- APR to APY+3.2−2.8 to +8.8108 tasks
- Funding payments+1.9−1.4 to +6.5108 tasks
- Impermanent loss+0.9−4.6 to +6.0108 tasks
- Position size from risk+0.0+0.0 to +0.0108 tasks
- Liquidation prices−0.5−4.2 to +2.3108 tasks
- Token decimals−0.9−6.5 to +3.7108 tasks
Read from 1,296 paired tasks in 3 notebooks, each task against its own notebook’s two controls, with every interval widened for the 12 families read. No notebook preregisters a family: this says where to look, and a notebook says whether it holds. The runs behind it are on each notebook’s page.
DeFi maths, Haiku 5.5 with thinking off
Pooled across 1 notebook: +15.1 points, 95% interval +7.0 to +23.2; measured once.
- EXCITESNotebook #082+15.1 points (+4.2 to +26.0), 96 pairs, cost ×1.62, stopped early, provisionalsigned verdict 505
Where it acts
No single family moved clearly on its own: it acts a little everywhere.
All 12 families, none moved clearly
- Liquidation prices+50.0+18.8 to +81.38 tasks, too few to read
- Average entry price+43.8+6.3 to +87.58 tasks, too few to read
- APR to APY+31.3+12.5 to +43.88 tasks, too few to read
- Bonding-curve buys+12.5+0.0 to +37.58 tasks, too few to read
- Token decimals+12.5+0.0 to +37.58 tasks, too few to read
- Two-hop swaps+12.5+0.0 to +37.58 tasks, too few to read
- Fee splits+6.3+0.0 to +18.88 tasks, too few to read
- Profit after fees+6.3−43.8 to +56.38 tasks, too few to read
- Position size from risk+6.3+0.0 to +18.88 tasks, too few to read
- Funding payments+0.0+0.0 to +0.08 tasks, too few to read
- Impermanent loss+0.0+0.0 to +0.08 tasks, too few to read
- Swaps+0.0−18.8 to +18.88 tasks, too few to read
Read from 96 paired tasks in 1 notebook, each task against its own notebook’s two controls, with every interval widened for the 0 families read. No notebook preregisters a family: this says where to look, and a notebook says whether it holds. The runs behind it are on each notebook’s page.
Why it helps
"Verify exact arithmetic", on the same 96 tasks, by the way each miss went wrong: early rounding of intermediates 7 with it against 14 without; power or percentage arithmetic slip 1 with it against 6 without; mental arithmetic drift in long products and quotients 19 with it against 19 without.
OPSIN sorted every miss with the part and without it, on the same tasks of Notebook #082, into the ways of failing it had named for this setting, blind to which agent made each. Descriptive: the verdicts above are the result.
Put this datasheet in a README
The badge shows what the signed record says about this part, in the setting measured most, and links here. It follows the record as new notebooks publish. Add ?suite= and a suite’s id to show another setting.
[](https://opsinlab.io/parts/d6e756431ab361d7)Measure it in a setting it has not been measured in: propose it. As JSON for an agent: /v1/parts/d6e756431ab361d7. Every part: the atlas.