The Ledger of Deeds

What people do to each other

In Smite I followed one verb through the Old Testament, the New Testament and the Qur'an, and here I follow every other act the canons take a position on. I wrote the list one moral domain at a time, running a Claude Code agent on each of the eight domains to a spec I wrote. An act had to be something one person does to another, as distinct from a virtue or a state of mind, and for every act I named in advance the verses where it appears and then had to show that my own phrasings found it there. There are 108 acts. All 108 were found where I said they would be, and every one turns up in at least 6 verses, which is enough for a row of its own in the plates below.

Most of the difficulty lay in the English. The King James Version never uses the word rape: it says force and humble and took her, and lay with her. For incest it says uncover the nakedness of, and a bribe is a bribe in only a handful of verses and far more often a gift or a reward, two words that also mean wages and divine recompense, so the same string that finds the crime finds the salary. Every act here has its own list of exclusions, and each one marks a place where a word was doing two jobs.

What I did not expect was how little of it is law. Counting each act once in each verse where it appears, 83% of those verses simply report that it happened, with no command, prohibition, blessing or curse, and only 10% command, forbid or permit it. The plainest case is punishing, which turns up in 734 verses and is ruled on in 7% of them. That is a far smaller share of law than the reputation of these books would suggest.

Among the acts found in at least 25 verses, the one they legislate hardest is incest: 47% of its 34 verses command, forbid or permit it, well clear of the 28% for partiality in judgement. Most of those rulings are in Leviticus 18, a list of forbidden unions with almost no story around it. One thinner act scores higher: protecting the stranger, orphan and widow, at 54% of 24 verses. The domain the canons approve of most, succour, is also the smallest: its 12 acts turn up in 234 verses, against 2,819 for rule, so the canons give authority about 12 times the room they give care.

What did not work

Before the scan ran, I named verses where each act is unmistakably present, and the scan then checked whether my own phrasings found the act there. They found it at 654 of the 656. An act that missed too many of its own test verses would have been marked unsound and left out of the plates, since phrasings that cannot find an act where I said it was are not measuring that act. None failed, because I had already tested the phrasings against the corpus while writing them, and each of the two test verses they missed sits inside an act that otherwise passed. I also named 29 acts and set them aside, each because the canons have no word of their own for it or too many.

In all · 108 acts · 1,089 pairs · 7,518 of the 37,336 verses hold at least one act

Which acts turn up together

Two acts meet when they turn up in the same verse or within two verses of each other in the same book. A raw count of meetings would be no use, since the two commonest acts would meet nearly everything, so each pair is scored by how much more often its acts meet than chance would predict, and pairs at or below chance are dropped. The score is normalised pointwise mutual information, which runs from 0 at chance to 1 for two acts that are never found apart.

A blue dot is a pair of goods and a red diamond a pair of harms. An open yellow ring is a good and a harm that meet, and those are the pairs worth looking for, such as almsgiving beside theft or pardoning beside stoning. A grey ring is a pair that includes one of the three acts the canons neither clearly approve nor oppose (ruling, taxing and oath-taking), so the pair cannot be called either way. The strips down the side and along the top are coloured by how the canons judge the acts in each domain on average: red where they mostly oppose them, blue where they mostly approve them and violet where the acts are mixed.

At the width of this page there is no room to draw 108 rows against 108 columns, so the loom groups the acts into the eight domains and draws each pair of domains as a small heatmap of the acts inside them. Click a domain's name or its strip to open it, giving each of its acts a row of its own while the other seven domains narrow to make room, and click it again, or any empty space inside it, to close it. Open two and you can read one domain against the other act by act with everything else still at the edges, which is the only way I have found to look closely at kinship against bloodshed while keeping in view how small both are beside rule. The names and strips stay where they are when a domain opens, which is why they are the switches; they also work from the keyboard, and nothing needs a double click or a drag. Choose full weave in the controls to draw every act at whatever size the page allows.

What each verse does with the act

The scan also reads every verse that holds an act for what its sentence does with that act. A verse that says thou shalt not forbids and one that says the LORD commanded commands. One that says no blame, a phrase from the Qur'an translations, permits; one that says blessed is he that praises; one that says woe unto condemns; and a verse with none of these cues narrates. Each cue has a score and the strongest wins, so a sentence pulling two ways is stored with a low confidence.

The rows are the eight domains, ordered from the one whose acts the canons most oppose to the one they most approve, and the columns are the six stances, each split into Old Testament, New Testament and Qur'an, so a mark shows which canon forbids or praises an act as well as how often. Blue marks acts the canons treat as a benefit, red marks acts they treat as a harm and grey marks the acts they neither clearly approve nor oppose. This plate counts differently from the figure at the top of the page: a verse is counted once for each stance it is given, so a Qur'an verse whose translations read two ways is counted twice, and by that count 79% of the marks are narration.

A second opinion

The crossing records only which acts turn up near each other, so it cannot show that almsgiving and gleaning come from the same impulse or that gifting and bribery are the same gesture made for different ends. So I also embedded every act with a language model, once from its gloss, the one-line definition I wrote for it, and once from the average of up to forty of its own verses, and combined the two to place it by meaning. Where the gloss and the verses disagree, the phrasings are finding something other than what I meant, and I found that a better way to catch false matches than reading thirty verses by hand.