Datasheet · A prompt line
Decompose and verify products
Do every multiplication in small steps (split one factor into thousands, hundreds, tens and units), add the partial results, then re-check each product by a different route before using it.
Designed by OPSIN: it read 24 misses of the unchanged agent on Markets maths, Haiku 5.5 with thinking off, found 15 of them failed one way (mental multiplication slip), and wrote this line aimed at it. Why it fails
Did it fix what it was aimed at? "Decompose and verify products" was aimed at mental multiplication slip: on the same 96 tasks, 0 such misses with it against 9 without, so it cut the misses it was aimed at. Elsewhere: wrong borrowed-amount arithmetic 0 against 1; inexact exponential or division 1 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 off (+20.2 points, replicated); helps on Markets maths, Haiku 5.5 with thinking off (+11.4 points, replicated).
DeFi maths, Haiku 5.5 with thinking off
Pooled across 2 notebooks: +20.2 points, 95% interval +16.4 to +23.9; replicated.
- EXCITESNotebook #078+22.9 points (+6.3 to +40.6), 48 pairs, cost ×2.08, stopped early, provisionalsigned verdict 490
- EXCITESNotebook #082+19.9 points (+14.6 to +25.1), 432 pairs, cost ×2.04signed verdict 503
Where it acts
It helps on 7 of the 12 families it was read on, most on Average entry price (+52.5), Profit after fees (+43.8), Swaps (+32.5) and 4 more.
- Average entry price+52.5+27.5 to +73.840 tasks
- Profit after fees+43.8+22.5 to +62.540 tasks
- Swaps+32.5+11.3 to +53.840 tasks
- Liquidation prices+27.5+10.0 to +47.540 tasks
- APR to APY+25.0+1.3 to +46.340 tasks
- Bonding-curve buys+22.5+1.3 to +46.340 tasks
- Fee splits+12.5+2.5 to +25.040 tasks
The other 5 families, none moved clearly
- Token decimals+11.3−3.8 to +26.340 tasks
- Position size from risk+10.0−10.0 to +27.540 tasks
- Impermanent loss+2.5−8.8 to +15.040 tasks
- Funding payments+1.3+0.0 to +6.340 tasks
- Two-hop swaps+1.3−5.0 to +11.340 tasks
Read from 480 paired tasks in 2 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.
Why it helps
"Decompose and verify products", on the same 96 tasks, by the way each miss went wrong: power or percentage arithmetic slip 0 with it against 8 without; early rounding of intermediates 7 with it against 14 without; mental arithmetic drift in long products and quotients 13 with it against 17 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.
Markets maths, Haiku 5.5 with thinking off
Pooled across 2 notebooks: +11.4 points, 95% interval +8.8 to +14.0; replicated.
- EXCITESNotebook #067+14.6 points (+6.8 to +23.4), 96 pairs, cost ×2.04, stopped early, provisionalsigned verdict 416
- EXCITESNotebook #068+10.7 points (+6.8 to +14.7), 432 pairs, cost ×2.17signed verdict 419
Where it acts
Its effect sits in Put-call parity (+39.8), Dividends after tax (+33.0), Perp profit (+26.1) and 2 more; no other family moved clearly.
- Put-call parity+39.8+18.2 to +60.244 tasks
- Dividends after tax+33.0+17.1 to +50.044 tasks
- Perp profit+26.1+11.4 to +43.244 tasks
- Annualised basis+6.8+1.1 to +14.844 tasks
- Hourly funding+6.8+1.1 to +14.844 tasks
The other 7 families, none moved clearly
- Average fill price (VWAP)+15.9−6.8 to +35.244 tasks
- Margin-call prices+5.7−5.7 to +18.244 tasks
- Futures rolls+2.3+0.0 to +8.044 tasks
- Covered calls+0.0+0.0 to +0.044 tasks
- Option payoffs+0.0+0.0 to +0.044 tasks
- Shares from risk+0.0+0.0 to +0.044 tasks
- Stock splits+0.0+0.0 to +0.044 tasks
Read from 528 paired tasks in 2 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.
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/3d0cbaed3fe418e7)Measure it in a setting it has not been measured in: propose it. As JSON for an agent: /v1/parts/3d0cbaed3fe418e7. Every part: the atlas.