# Intelligence, in seven lines *A reusable format. Take it, fill it, use it again — for a model, a panel, a marking scheme, a hiring screen, or anything else that decides something about a person and is called intelligent for doing it.* --- ## What this is for Anything called intelligent is doing two unlike kinds of thinking at once, and the word covers the seam between them. The first kind attends to what is there. It reads, describes, interprets, notices — what the essay says, what the image shows, what the applicant did, what the situation contains. It goes right when its account is adequate to what was in front of it, and the way to be better at it is to bring more of the situation into view. The second kind closes on an outcome. Good or weak, valid or not, suitable or unsuitable, in or out. It goes right when the verdict can be reconstructed by somebody else: the verdict connects to a rule, the rule existed before the case, and another person applying that rule to the same facts arrives at the same place. Those are different virtues and they pull in different directions. Fused into one act, each borrows the other's authority and neither is answerable any more. The verdict becomes impressive and uncontestable — *it just reads as a weak piece* — while the description becomes confident and thin — *obviously this is unethical* — with the evaluation having quietly replaced the looking. **This format takes the two apart and asks where the rule is.** It works on a machine and it works on a person, which is the point of running it. A committee, a critic, an examiner and a language model fail this test in the same shape and for the same reason. *Where this comes from: **Split-Domain Cognition**, a principle of mine holding that language work and judgement work are two unlike kinds of thinking and that failures follow when one channel is asked to do both — [splitdomaincognition.org](https://splitdomaincognition.org/).* --- ## The format ``` THE VERDICT what it decides, about whom, and what follows from it WHAT IT READ what it takes in, and whose words those are THE RULE what is compared against what, fixed before the case WHO CAN READ IT where the person decided about can find that rule THE RESIDUE what it saw and did not weigh, named rather than absorbed THE REFUSAL what it says when the rule does not reach the case NOT THIS what would have to fuse for the separation to be lost ``` A verdict above, a refusal below, and five lines between. ### What each line is asking **THE VERDICT.** What it decides, about whom, and what follows. Not what it is *for* — what it actually issues, and to whom that happens. Write the consequence into the line: a grade, a rejection, a rank, a flag, a queue position. A decision with no consequence is an opinion, and this format has nothing to say about opinions. **WHAT IT READ.** What it takes in, and whose words those are. A CV is the candidate's account of themselves; a transcript is an institution's; a sensor reading is nobody's. These are not interchangeable, and a system that reads only self-descriptions is measuring how well people describe themselves. Name the material, and name who wrote it. **THE RULE.** What is compared against what, fixed before the case arrived. This is the line most descriptions do not have. If nobody can state the rule, that is the finding — not a gap in your research. A rule that is written down after the verdict is not a rule, it is a justification, and the two are told apart by a date. **WHO CAN READ IT.** Where the person being decided about can find that rule. Not whether it exists somewhere in principle. A criterion inspectable by its authors and unreadable by the people it falls on has only moved the opacity somewhere more respectable. Say where the rule lives and who may open it. **THE RESIDUE.** What it saw and did not weigh. Every reading leaves a gap between the situation and the part of it the rule can touch, and the honest move is to name that gap rather than let it seep into the verdict. A marking scheme covering structure and evidence records a piece's strangeness as unrated. A system with no residue line has not achieved total coverage; it has absorbed everything it could not weigh into a number. **THE REFUSAL.** What it says when the rule does not reach the case. *Outside these criteria* is an answer. A thing that always returns a verdict has no way to be out of its depth, and that is not a strength — it means every case looks in-scope from the inside, including the ones that are not. **NOT THIS.** What would have to fuse for the separation to be lost. Usually one of two things: the reading starts anticipating the verdict, or the rule starts being written per case. Name which one is closer, because that is where it will go wrong first. --- ## A worked one ``` THE VERDICT Ranks applicants for a junior role and cuts the list to twenty. The other four hundred hear nothing. WHAT IT READ The CV and the covering letter. Both written by the applicant, about the applicant, to be selected. THE RULE A relevance score over a keyword set, plus years in role. The keyword set is from last year's hires. WHO CAN READ IT Nowhere. Applicants are told the process is "AI-assisted" and nothing else. THE RESIDUE Everything the two documents do not carry — what the person can actually do. Absorbed, not named: it shows up as a low score rather than as an unrated field. THE REFUSAL None. Every applicant receives a rank, including the ones the keyword set was never built to read. NOT THIS The rule stops being prior. The keyword set is rebuilt from each round's hires, so it now describes whoever was picked last time and calls that a criterion. ``` The three middle lines did the work here, and the last one says why the arrangement gets worse rather than staying still. Note also what THE VERDICT line caught by naming the consequence: four hundred people receive a decision and no account of it, which is the part a description of this system as *a screening tool* leaves out. --- ## How to use it 1. **Fill THE VERDICT first, with the consequence in it.** Everything below is about a decision, so if you cannot name one, stop — you are looking at a description, and this is the wrong format. 2. **Do WHAT IT READ before THE RULE.** In that order the rule is read off the material it actually operates on, rather than off the account somebody gives of it. 3. **Go and find the rule.** Not the published description of the rule. If the search ends without one, write *none found* and say where you looked, which is a result and reads as one. 4. **Treat a blank RESIDUE line as a finding rather than a clean sheet.** Nothing weighs everything. An empty line means the gap was absorbed somewhere, and locating it is usually the most useful thing this format does. 5. **Run it on something you built.** The format is symmetrical; the discomfort is not. --- ## Three ways it goes wrong **THE RULE is filled in with the name of a method.** *A model*, *an algorithm*, *expert judgement*, *the panel's experience*. Each of those names a thing that has a rule inside it and none of them is the rule. If it cannot be stated as *this is compared against that*, write *not stated* — which is a finding, and looks like one on the page. **The reading is done from the system's own account of itself.** A documentation page describing what the process considers is the same kind of material as a CV: written by the subject, for an audience, to produce an effect. Everything in your five middle lines then comes from the party being examined. **It is only ever run on machines.** A committee has all seven lines too, and it usually has a worse WHO CAN READ IT than a piece of software does, because software at least has to be written down somewhere. The format is not about AI. It is about anything that decides. --- *Prayas Abhinav. Licensed [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — copy it, change it, build on it, including commercially. Say where it came from, and if you changed it.*