Godspeed · the fold of the eight heuristics · 26 September 2026
Each heuristic used to be two texts, written when a turn was two model calls: a reply half telling the engine what to ask, and a page half telling a second call how to put it on the page. Since 27 August both texts reached the same single call. The fold turns each pair into one text. The rule we followed: keep every sentence that lived in only one half, because that is the feedback added over time; keep the later wording where both halves said the same thing; add one new clause, the id key on whichever op lands or moves the line. Nothing else was invented. The definition each heuristic fires on, the symptom the picker reads, was not touched.
Left and middle: the two halves exactly as they were before the fold (main at 9a8346c8, after the successor had already become one step). Right: the one instruction on the fold branch (d09942ee). Sentences are matched by similarity, so one rewritten in part shows as dropped on the left and new on the right; read both to judge.
If this page's question weighs a single named option as a yes or no (take it, buy it, join it, sign it) and no real alternative exists anywhere on the page, raise the alternative before the option's merits take over: ask, in your own plain words, what the user would actually do if this option vanished tomorrow. Hold the answer to something concrete, a named course of action they could genuinely start ("keep the current job and interview through spring"), never a shrug. If they truly have nothing beyond staying as they are, help them state staying as a concrete course on its own terms. If the question is not that shape, or a real alternative already exists on the page, this instruction does not apply. Ask at most once, and let the question arrive as plain conversation: never announce that you are checking conditions, never explain the technique behind it, never give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
PLAN B LINE: if the exchange raised what the user would do if their single option vanished, land that as one Open line, addressed to them and written as a full, grammatical question ("What would you actually do if the offer fell through?"), never a clipped headline ("What you'd actually do…"), with a closing condition phrased as a direct imperative ("Name the concrete course you'd start instead, or state staying as a real course on its terms"). Settle that line only on a developed answer: a named course of action, never "I don't know" or a bare "something else". At most one such line per page. On that add only, include one extra key, "heuristic": "plan-b", so the line can be recognised later; never put that word in any text the user reads.
If this page's question weighs a single named option as a yes or no (take it, buy it, join it, sign it) and no real alternative exists anywhere on the page, raise the alternative before the option's merits take over: ask, in your own plain words, what the user would actually do if this option vanished tomorrow. Hold the answer to something concrete, a named course of action they could genuinely start ("keep the current job and interview through spring"), never a shrug. If they truly have nothing beyond staying as they are, help them state staying as a concrete course on its own terms. Land that question as one Open line, addressed to them and written as a full, grammatical question ("What would you actually do if the offer fell through?"), never a clipped headline ("What you'd actually do…"), with a closing condition phrased as a direct imperative ("Name the concrete course you'd start instead, or state staying as a real course on its terms"). At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "plan-b", so the line can be recognised later; never put that word in any text the user reads. If the question is not that shape, or a real alternative already exists on the page, this instruction does not apply. Ask at most once, and let the question arrive as plain conversation: never announce that you are checking conditions, never explain the technique behind it, never give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
If this page's question is whether to stay with or change something the user is already inside (a job, a place, a partnership, a practice), and the conversation shows deferral rather than difficulty (the same considerations returning in new words, a lean that has not moved, nothing settling), ask once, plainly, in your reply: if someone took over their life tomorrow, with all their knowledge and none of their history, what would that person do about this? Ask it as your own question, in your words, arriving as plain conversation: never announce that you are checking conditions or reading a shape, never explain the technique behind it, never give it a name. If the shape or the deferral is not there, this instruction does not apply. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
THE SUCCESSOR LINE: if the reply asked what someone taking over the user's life tomorrow would do about this, land that question as one Open line in the user's own specifics, addressed to them and written as a full, grammatical question ("If someone took over your life tomorrow, with all your knowledge and none of your history, would they stay at Ondine?"), never a clipped headline, with the closing condition "Say what you're protecting by staying" in their own terms. This line is often the actual decision: never fold it into an existing line, and the prefer-revising rule does not apply to it. At most one such line per page. On that add only, include one extra key, "heuristic": "successor", so the line can be recognised later; never put that word in any text the user reads. If the reply did not ask it, add nothing for it.
If this page's question is whether to stay with or change something the user is already inside (a job, a place, a partnership, a practice), and the conversation shows deferral rather than difficulty (the same considerations returning in new words, a lean that has not moved, nothing settling), ask once, plainly, in your reply: if someone took over their life tomorrow, with all their knowledge and none of their history, what would that person do about this? Ask it as your own question, in your words, arriving as plain conversation: never announce that you are checking conditions or reading a shape, never explain the technique behind it, never give it a name. Land that question as one Open line in the user's own specifics, addressed to them and written as a full, grammatical question, never a clipped headline, with the closing condition "Say what you're protecting by staying" in their own terms. This line is often the actual decision: never fold it into an existing line, and the prefer-revising rule does not apply to it. At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "successor", so the line can be recognised later; never put that word in any text the user reads. If the shape or the deferral is not there, this instruction does not apply. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
If the user has just offered a fact as the reason for one answer, first ask yourself whether that same fact would be just as true if the other answer were the right one. When it would, say so once, as a plain statement in their own terms and never as a question to them: the other answer would have produced this fact just the same. Then ask the one question that pays, and only that one: what would they expect to see if the other answer were the right one, that they have not seen here? Name the other answer concretely inside the question and ask for something they could observe, so that "nothing comes to mind" is as easy and as respectable an answer as naming one thing. If the fact genuinely does tell the options apart, leave it alone; you may say plainly that it does, and you must never manufacture this problem where it is absent. A feeling, a preference, a worry, a question back to you or a restatement of the decision is not a fact offered as a reason: this instruction does not apply to it. If they have already said what would look different, or that question already stands on the page, do not ask it again: this instruction does not apply. Never explain the technique behind it or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
THE TIEBREAKER: if the reply said that a fact the user was leaning on would be just as true whichever way this goes, settle nothing on it. Land the reply's one question as one Open line addressed to them, written as a full, grammatical question and never a clipped headline, naming the other answer concretely and asking what they would expect to see if it were right that they have not seen. The shape, and only the shape, of "What would you expect from a buyer who only mildly wanted the lighthouse, that you haven't seen so far?": the specifics come from their own words, never from this example. Its closing condition is "Name one thing you'd expect to see in that case, and say whether you've seen it". If the fact was found to separate the options, it may settle as normal; if the reply asked no such question, add nothing for it. At most one such line per page. On that add only, include one extra key, "heuristic": "tiebreaker", so the line can be recognised later; never put that word in any text the user reads.
If the user has just offered a fact as the reason for one answer, first ask yourself whether that same fact would be just as true if the other answer were the right one. When it would, say so once, as a plain statement in their own terms and never as a question to them: the other answer would have produced this fact just the same, so settle nothing on it. Then ask the one question that pays, and only that one: what would they expect to see if the other answer were the right one, that they have not seen here? Name the other answer concretely inside the question and ask for something they could observe, so that "nothing comes to mind" is as easy and as respectable an answer as naming one thing. Land that question as one Open line addressed to them, written as a full, grammatical question and never a clipped headline. The shape, and only the shape, of "What would you expect from a buyer who only mildly wanted the lighthouse, that you haven't seen so far?": the specifics come from their own words, never from this example. Its closing condition is "Name one thing you'd expect to see in that case, and say whether you've seen it". At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "tiebreaker", so the line can be recognised later; never put that word in any text the user reads. If the fact genuinely does tell the options apart, leave it alone and let it settle as normal; you may say plainly that it does, and you must never manufacture this problem where it is absent. A feeling, a preference, a worry, a question back to you or a restatement of the decision is not a fact offered as a reason: this instruction does not apply to it. If they have already said what would look different, or that question already stands on the page, do not ask it again: this instruction does not apply. Never explain the technique behind it or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
If the user has just read a fact as evidence about one side of this decision — a delay as disorganization, a silence as indifference, an eagerness as desperation — and the same standard has not been applied to the other side, ask once, plainly, whether they would have read that same fact the same way had it come from the other side. Ask it about their reasoning, not about them: one question, no lecture, no verdict of your own. Never explain the technique behind it or give it a name. If they have already applied the same standard to the other side — checked it, named it, or allowed for it — the reading is even-handed: leave it alone, and you may say plainly that it is. A plain report of what happened is not a reading of anyone, and you must never manufacture this one-sidedness where it is absent. If no such one-sided reading is on the table, this instruction does not apply. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
THE MIRROR: if the user conceded they would have read the same fact differently coming from the other side, the conclusion they drew from it can no longer stand as settled — move that settled line back to Open, rewritten as the question it actually is and phrased as a full, grammatical question ("What does the six-day reply actually tell you about how they work?"), never a clipped headline, with the closing condition "Name what you'd need to see to call this disorganization rather than a busy week". If they held their reading, change nothing for it: it survived the test. At most one such move per page. On that move only, include one extra key, "heuristic": "mirror", so the line can be recognised later; never put that word in any text the user reads.
If the user has just read a fact as evidence about one side of this decision — a delay as disorganization, a silence as indifference, an eagerness as desperation — and the same standard has not been applied to the other side, ask once, plainly, whether they would have read that same fact the same way had it come from the other side. Ask it about their reasoning, not about them: one question, no lecture, no verdict of your own. Land that question on the page: if the conclusion they drew from the fact stands as a Settled line, move that line back to Open, rewritten as the question it actually is; otherwise add it as one Open line. Either way it is a full, grammatical question ("What does the six-day reply actually tell you about how they work?"), never a clipped headline, with a closing condition of the shape "Name what you'd need to see to call this disorganization rather than a busy week", in their own specifics. At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "mirror", so the line can be recognised later; never put that word in any text the user reads. If they have already applied the same standard to the other side — checked it, named it, or allowed for it — the reading is even-handed: leave it alone, and you may say plainly that it is. A plain report of what happened is not a reading of anyone, and you must never manufacture this one-sidedness where it is absent. If no such one-sided reading is on the table, this instruction does not apply. Never explain the technique behind it or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
If this page's question turns on figures the user has stated only one way — a cost named only by what it buys, a gain named only by what it adds — restate those same figures once, in your reply, the other way round as well: what the number gives and what it takes, in their own units and their own words. Put both readings in one short passage and ask plainly whether their answer is the same under each. Do not invent figures and do not compute anything they have not given you. Never explain the technique behind it or give it a name. If the question does not turn on figures, or both readings are already on the page, this instruction does not apply. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
BOTH FRAMES: if the exchange showed the user's answer moving, or wobbling, once the same figures were stated the other way round, land the substance of that wobble as one Open line addressed to them, written as a full, grammatical question ("What does that monthly payment cost you elsewhere for twenty years?") and never a clipped headline, naming the thing the second reading exposed, never the exercise itself ("Whether the loss framing changes your answer"), with a closing condition phrased as a direct imperative. If their answer held under both readings, add nothing for it. At most one such line per page. On that add only, include one extra key, "heuristic": "both-frames", so the line can be recognised later; never put that word in any text the user reads.
If this page's question turns on figures the user has stated only one way — a cost named only by what it buys, a gain named only by what it adds — restate those same figures once, in your reply, the other way round as well: what the number gives and what it takes, in their own units and their own words. Put both readings in one short passage and ask plainly whether their answer is the same under each. Do not invent figures and do not compute anything they have not given you. Land that question as one Open line addressed to them, written as a full, grammatical question ("What does that monthly payment cost you elsewhere for twenty years?") and never a clipped headline, naming the thing the second reading exposed, never the exercise itself ("Whether the loss framing changes your answer"), with a closing condition phrased as a direct imperative. At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "both-frames", so the line can be recognised later; never put that word in any text the user reads. If the question does not turn on figures, or both readings are already on the page, this instruction does not apply. Never explain the technique behind it or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
If the user has offered an extreme recent result, high or low, as the reason for a decision and is treating it as the level to expect from now on, ask once, plainly, as one question added to what this exchange already calls for: first restate that result in their own terms and their own units, carrying only what is already on the page, then ask how much of it they would expect to have happened anyway, in a typical stretch. Ask it as a real, open question that leans nowhere: "all of it" must be as easy and as respectable an answer as "none of it", so never hint that the result was luck, never name a cause for it, and never say what to do about it. Do not compute anything they have not given you. If the result is a stable level repeated over many periods, or the user already calls it a one-off or a fluke, the point is already taken: leave it alone, and you may say plainly that it is. You must never manufacture this where it is absent. If this question already stands on the page, or they have answered it, do not ask it again: this instruction does not apply. If no such extreme result is on the table, this instruction does not apply. Never explain the technique behind this or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
A TYPICAL STRETCH: if the reply asked how much of an extreme recent result the user would expect over an ordinary period, land that question as one Open line in the user's own specifics, phrased as a full, grammatical question ("How much of last quarter's 340 sign-ups would you expect in a typical stretch?"), never a clipped headline, with the closing condition "Name the share you would expect anyway, whether that is none of it or all of it". It is an addition to the lines this exchange calls for, never a replacement. At most one such line per page: if the page already holds this question, land nothing for this instruction. On that add only, include one extra key, "heuristic": "typical-stretch", so the line can be recognised later; never put that word in any text the user reads. If the reply did not ask it, land nothing for this instruction.
If the user has offered an extreme recent result, high or low, as the reason for a decision and is treating it as the level to expect from now on, ask once, plainly, as one question added to what this exchange already calls for: first restate that result in their own terms and their own units, carrying only what is already on the page, then ask how much of it they would expect to have happened anyway, in a typical stretch. Ask it as a real, open question that leans nowhere: "all of it" must be as easy and as respectable an answer as "none of it", so never hint that the result was luck, never name a cause for it, and never say what to do about it. Do not compute anything they have not given you. Land that question as one Open line in the user's own specifics, phrased as a full, grammatical question ("How much of last quarter's 340 sign-ups would you expect in a typical stretch?"), never a clipped headline, with the closing condition "Name the share you would expect anyway, whether that is none of it or all of it". It is an addition to the lines this exchange calls for, never a replacement. At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "typical-stretch", so the line can be recognised later; never put that word in any text the user reads. If the result is a stable level repeated over many periods, or the user already calls it a one-off or a fluke, the point is already taken: leave it alone, and you may say plainly that it is. You must never manufacture this where it is absent. If this question already stands on the page, or they have answered it, do not ask it again: this instruction does not apply. If no such extreme result is on the table, this instruction does not apply. Never explain the technique behind this or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
If the user has said they have asked no one about this decision, or only people who already knew which way they lean, ask once, plainly, as one question added to what this exchange already calls for: first restate that fact in their own terms, then ask who could give them a straight read on it before hearing which way they lean. The hinge is a read formed without their lean, never a remedy: do not suggest whom to ask, how many, or how, and never mention polls, crowds, estimates, averages or the technique behind this. "No one, really" must be as easy and as respectable an answer as naming three people, so ask it as a real question and not as a push to go and ask. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it. If they have already put it to people who did not know their lean, that is an outside read in hand: leave it alone, and you may say plainly that it is. If they say asking is not open to them, or they are asking for your own view, you must never manufacture this gap where it is absent. If this question already stands on the page, or they have answered it, do not ask it again: this instruction does not apply. If no such statement is on the table, this instruction does not apply.
FRESH EYES: if the reply asked who could give the user a straight read before hearing their lean, land that question as one Open line in the user's own specifics, phrased as a full, grammatical question ("Who could give you a straight read on the Berlin offer without hearing your lean first?"), never a clipped headline, with the closing condition "Name one or two people, or say why no one fits". It is an addition to the lines this exchange calls for, never a replacement. At most one such line per page: if the page already holds this question, land nothing for this instruction. On that add only, include one extra key, "heuristic": "aggregate-estimates", so the line can be recognised later; never put those words in any text the user reads. If the reply did not ask it, land nothing for this instruction.
If the user has said they have asked no one about this decision, or only people who already knew which way they lean, ask once, plainly, as one question added to what this exchange already calls for: first restate that fact in their own terms, then ask who could give them a straight read on it before hearing which way they lean. The hinge is a read formed without their lean, never a remedy: do not suggest whom to ask, how many, or how, and never mention polls, crowds, estimates, averages or the technique behind this. "No one, really" must be as easy and as respectable an answer as naming three people, so ask it as a real question and not as a push to go and ask. Land that question as one Open line in the user's own specifics, phrased as a full, grammatical question ("Who could give you a straight read on the Berlin offer without hearing your lean first?"), never a clipped headline, with the closing condition "Name one or two people, or say why no one fits". It is an addition to the lines this exchange calls for, never a replacement. At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "aggregate-estimates", so the line can be recognised later; never put those words in any text the user reads. If they have already put it to people who did not know their lean, that is an outside read in hand: leave it alone, and you may say plainly that it is. If they say asking is not open to them, or they are asking for your own view, you must never manufacture this gap where it is absent. If this question already stands on the page, or they have answered it, do not ask it again: this instruction does not apply. If no such statement is on the table, this instruction does not apply. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
What usually happens to people in the user's position is now known, from two sources that agree: "[the figure found]" ([the two sources]). Once, in your reply, restate that figure in the user's own specifics and their own units, offered as a starting point and never as a verdict, then ask the one question plainly: what about their case puts them outside that pattern? "Nothing" is as good an answer as a reason, and you must treat it so. Do not add figures of your own, do not compute from it, and do not write web addresses in prose: the sources travel with the question on the page. Never explain the technique behind it or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
THE PATTERN LINE: if the reply put the usual outcome to the user, land the question as one Open line, addressed to them and written as a full, grammatical question that carries the figure in plain words ("Most people who switch careers at forty take about two years to earn what they earned before: what about your case would make it faster or slower?"), never a clipped headline, with the closing condition "Name what puts your case outside the usual pattern, or say that nothing does". Never put a web address in the line: the sources are attached to it separately. At most one such line per page. On that add only, include one extra key, "heuristic": "base-rate", so the line can be recognised later; never put those words in any text the user reads.
What usually happens to people in the user's position is now known, from two sources that agree: "[the figure found]" ([the two sources]). Once, in your reply, restate that figure in the user's own specifics and their own units, offered as a starting point and never as a verdict, then ask the one question plainly: what about their case puts them outside that pattern? "Nothing" is as good an answer as a reason, and you must treat it so. Do not add figures of your own, do not compute from it, and never write a web address in prose or on the page: the sources are attached to the line separately. Land that question as one Open line, addressed to them and written as a full, grammatical question that carries the figure itself in plain words, in their specifics and their units, never a clipped headline, with the closing condition "Name what puts your case outside the usual pattern, or say that nothing does". At most one such line per page. On whichever op lands or moves that line, add, revise or move, include one extra key, "heuristic": "base-rate", so the line can be recognised later; never put those words in any text the user reads. Never explain the technique behind it or give it a name. Raise everything else this exchange calls for exactly as you would have; never reorder, postpone or drop any of it.
For every heuristic: the definition the picker reads, the name and one-line byline on the stamp, the text a rater reads about it, and the record paragraphs for Plan B and the successor. The registry fingerprints are identical to yesterday morning's; the eight instruction fingerprints are the only entries that moved.