{"handle":"business-frameworks/decision@1.6.1","path":"decision","version":"1.6.1","title":"Decision rules: wrong as rarely as possible for the effort","layer":1,"free":true,"tokens":541,"tokens_terse":119,"budget":545,"price_usd":0,"children":["business-frameworks/decision/parameters@1.6.1","business-frameworks/decision/extension@1.6.1","business-frameworks/decision/effort@1.6.1","business-frameworks/decision/ask@1.6.1","business-frameworks/decision/change@1.6.1","business-frameworks/decision/continue-or-stop@1.6.1","business-frameworks/decision/negotiate@1.6.1","business-frameworks/decision/predict@1.6.1","business-frameworks/decision/who-decides@1.6.1"],"related":["business-frameworks/equation@1.6.1","business-frameworks/value/decide/reinvest-or-distribute@1.6.1","business-frameworks/risk/decide/hurdle@1.6.1","business-frameworks/cash-flow/decide/runway@1.6.1","business-frameworks/ethics/feasible@1.6.1","business-frameworks/worker/read-the-job@1.6.1","business-frameworks/quality/see-the-work@1.6.1"],"cites":["Drucker (2004), Harvard Business Review","Gabarro and Kotter (1980), Harvard Business Review","Kim and Mauborgne (1997), Harvard Business Review","Argyris (1991), Harvard Business Review","Allen, T. J., Managing the Flow of Technology, 1977","Harvard Business Review, HBR Guide to Data Analytics Basics for Managers, 2018","Harvard Business Review, HBR Guide to AI Basics for Managers, 2023","Bishop, C. M., Pattern Recognition and Machine Learning, 2006","Bishop, C. M. and Bishop, H., Deep Learning: Foundations and Concepts, 2024","Ancona, Malone, Orlikowski and Senge (2007), Harvard Business Review"],"shape":null,"text":"A decision rule in this document is written to be wrong as rarely as possible for the effort it asks, not to be exact; where one kind of error costs far more than the other, it aims at the lowest expected cost of error, not the fewest errors. Split the effort about 80/20: a short analysis with the default parameters settles most decisions; the full analysis is kept for the few where a coarse rule would err expensively. Every decision node states inputs, rule, action and check; its parameters are defaults for the typical business in scope, meant to be adjusted.\n\nThree questions take a reader outside the defaults: a parameter that does not fit the business (`decision/parameters`); a business outside the bands the rules were written for (`decision/extension`); a decision whose stakes do not match the effort the rule assumes (`decision/effort`).\n\nThe rules exist to be executed by an agent with little context. A default that is right for the typical case and adjustable for the rest costs fewer tokens and fewer errors than an exact method applied to every case. The parameters marked \"rule of thumb\" in the decision nodes are the ones this branch governs. Where one kind of error costs far more than the other, a rule aims at the lowest expected cost, not the fewest errors, and names both costs as inputs. Each rule says when its answer is too uncertain to act on; the case then goes back to the owner. Confirm a symptom's cause before a rule is applied, and test the explanation by what else it predicts. Two figures that move together, or a forecast that predicts well, do not show a cause. An action counts only when it names who does what by when, and who must be told. A proposal's conviction, repetition or the standing of its source is not evidence: restate the claim without its delivery, then test it. For a tool, ask what goes in, what is learned and what comes out; a seller's account leans toward what it sells. Where others must carry a decision out, an unqualified yes is tested before it is counted. An explanation that rests on a person, the decider's own part included, is tested like any other; after a miss, look inside the business for the cause before crediting outside forces. Before a large choice, bring in one view from outside the group proposing it, or have someone argue the contrary case. Where a person may override a rule's or a model's answer, the override states the assumption it rests on, is reviewed in the open and is scored against the outcome at the next check; overrides that keep proving right mean the rule should change.","variants":{"terse":"the lede alone","full":"lede, conditions and reasoning","agent":"the same accepted text as fields (served with ?variant=agent): scope, inputs, rule, action, check, reasoning, units, example, apply, recheck, recheck_note, closing_citation; the same price"},"versions":["1.6.1","1.6","1.5","1.4","1.3","1.2"],"variant":"full","levels":{"lede":"A decision rule in this document is written to be wrong as rarely as possible for the effort it asks, not to be exact; where one kind of error costs far more than the other, it aims at the lowest expected cost of error, not the fewest errors. Split the effort about 80/20: a short analysis with the default parameters settles most decisions; the full analysis is kept for the few where a coarse rule would err expensively. Every decision node states inputs, rule, action and check; its parameters are defaults for the typical business in scope, meant to be adjusted.","conditions":"Three questions take a reader outside the defaults: a parameter that does not fit the business (`decision/parameters`); a business outside the bands the rules were written for (`decision/extension`); a decision whose stakes do not match the effort the rule assumes (`decision/effort`).","reasoning":"The rules exist to be executed by an agent with little context. A default that is right for the typical case and adjustable for the rest costs fewer tokens and fewer errors than an exact method applied to every case. The parameters marked \"rule of thumb\" in the decision nodes are the ones this branch governs. Where one kind of error costs far more than the other, a rule aims at the lowest expected cost, not the fewest errors, and names both costs as inputs. Each rule says when its answer is too uncertain to act on; the case then goes back to the owner. Confirm a symptom's cause before a rule is applied, and test the explanation by what else it predicts. Two figures that move together, or a forecast that predicts well, do not show a cause. An action counts only when it names who does what by when, and who must be told. A proposal's conviction, repetition or the standing of its source is not evidence: restate the claim without its delivery, then test it. For a tool, ask what goes in, what is learned and what comes out; a seller's account leans toward what it sells. Where others must carry a decision out, an unqualified yes is tested before it is counted. An explanation that rests on a person, the decider's own part included, is tested like any other; after a miss, look inside the business for the cause before crediting outside forces. Before a large choice, bring in one view from outside the group proposing it, or have someone argue the contrary case. Where a person may override a rule's or a model's answer, the override states the assumption it rests on, is reviewed in the open and is scored against the outcome at the next check; overrides that keep proving right mean the rule should change."},"url":"https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision","children_urls":["https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/parameters","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/extension","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/effort","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/ask","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/change","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/continue-or-stop","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/negotiate","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/predict","https://business-frameworks.matryoshka-paradigms.workers.dev/node/decision/who-decides"],"related_urls":["https://business-frameworks.matryoshka-paradigms.workers.dev/node/equation","https://business-frameworks.matryoshka-paradigms.workers.dev/node/value/decide/reinvest-or-distribute","https://business-frameworks.matryoshka-paradigms.workers.dev/node/risk/decide/hurdle","https://business-frameworks.matryoshka-paradigms.workers.dev/node/cash-flow/decide/runway","https://business-frameworks.matryoshka-paradigms.workers.dev/node/ethics/feasible","https://business-frameworks.matryoshka-paradigms.workers.dev/node/worker/read-the-job","https://business-frameworks.matryoshka-paradigms.workers.dev/node/quality/see-the-work"],"cite_as":"McHenry, J. (2026). Business Frameworks, version 1.6.1, business-frameworks/decision@1.6.1."}