Writing Standards into Programs
写进程序
Once a threshold is written into a program, every time the program runs it carries out that line of ought once. This paper examines how writing standards into programs changes the cost of questioning them, the visibility of the normative trade-offs they embody, and the way those trade-offs are repeatedly carried out. The choices of target variable, proxy variable, threshold and fairness metric each record a line on the normative ledger; once written into a program, they appear in the shape of parameters (S35). In arrangements that execute automatically, without a corresponding stage of fresh judgment, every run of the program carries out that line once (S36). The presumption of conformity has legal authorization and legal effect, and misrouting is not found on that basis; to prove misrouting one must check which challenge was not handled by the relevant standard and the specific review, and whether the presumption was used to register that challenge as answered (S37). Mathematics can show that several fairness conditions generally cannot be satisfied at once; it shows that a choice must be made, not which one (S38). That someone bears the judgment of deployment does not prove that the premises in the parameters have been sufficiently questioned (S39). The paper ends in conversation with Passi and Barocas. Five propositions in all (S35 to S39), stated at the end of each section and listed together at the end.
Writing Declaration: This paper was co-drafted with Claude (Anthropic). All intellectual decisions, framework design, and final editorial judgments were made by the author.
Introduction
Paper 6 described how credit scores are adopted as a gate for lending: one party writes the score model, another adopts it as a gate, and changing it goes through a procedure of validation and approval. This paper looks at the same thing from another angle. A loan underwriting process is used below as an illustration; it does not refer to any particular lender, and the numbers are illustrative too.
Once a score model is adopted, it is written, together with a cut-off score, into a lender's underwriting program. An application comes in, the program takes the score, compares it with the cut-off, and gives a result. Where the cut-off is set is a trade-off: how many applications wrongly judged good to let in, how many wrongly judged bad to keep out. This trade-off is made once, on the day the cut-off is set; after that, every time the program runs, it is carried out once. Suppose ten thousand applications a day; it is carried out ten thousand times; while the rules and parameters stay unchanged, every time is the same; without an associated record of reasons, those repeated executions no longer reveal the trade-off they carry out.
Writing rules into a program is not new in itself. What is new is that the ways of questioning them may have changed: to question one, one must first find where it is; it looks like a number; the same trade-off can be carried out over and over in a short time. This paper speaks of these three changes, and of the conditions under which they occur.
The reading follows the Prequel. This paper concerns the second domain, standards that act on persons (Prequel §5). §1 speaks of the ought in parameters; §2, of automatic execution; §3, of the presumption of conformity; §4, of what mathematics can and cannot show; §5 says that bearing is not the same as having questioned; §6 is in conversation with Passi and Barocas.
§1 The Ought in Parameters
To write a judgment into a program, several choices must first be made, and each records a line on the normative ledger.
Target variable. What is the program to predict? "Will this person repay on time," "is this application worth approving," "does this patient need extra care" are different targets. Translating a vague purpose into a variable that can be measured and predicted is itself a trade-off.
Proxy variable. What is to be predicted often cannot be measured, so something measurable stands in for it: past spending for need, past default for risk. Between the proxy and what it stands for there is a gap, and where that gap is left is another trade-off.
Threshold. The prediction gives a score; where the cut-off is set decides how two kinds of error are distributed.
Fairness metric. If the program is required to be fair to different groups, by which definition of "fair": equal pass rates across groups, equal error rates across groups, or the same score meaning the same risk in every group? These definitions, as §4 will say, generally cannot be satisfied at once.
Passi and Barocas studied the actual work of data science and pointed out that problem formulation, that is, translating a business goal into a problem that can be predicted, is itself full of normative choices, and that questions of fairness already appear at this step, not only in the model (Passi & Barocas 2019). Selbst and colleagues placed such problems under several "abstraction traps": what is lost when a social problem is abstracted into a technical one (Selbst et al. 2019).
Once written into a program, these lines all appear in the shape of parameters: a variable name, a number, a switch. They are still on the normative ledger; if there is no associated record of reasons beside the parameters, they are not easily seen as trade-offs. This has the same form as the quantification of Paper 5; writing into a program is quantification taken one step further.
S35 The ought in parameters. The choices of target variable, proxy variable, threshold and fairness metric each record a line on the normative ledger; once written into a program, they appear in the shape of parameters.
§2 Automatic Execution
A parameter written into a program is not necessarily executed automatically. It depends on whether, between the program's output and the final decision, there is a stage of fresh judgment.
If there is not: the program gives a result, and the result is the decision, or is almost always adopted as the decision. Then every run of the program carries out that line of ought once. The cut-off is one trade-off; with automatic execution, that one trade-off is copied onto every application.
If there is: the result goes through a person's review; the one decided about can appeal; on appeal someone looks again. Such arrangements are checked separately: whether the reviewer can change the result, whether an appeal can reach the parameters, whether the one who looks again looks at the individual case or also at the cut-off itself.
Mitchell and colleagues surveyed choices and assumptions often overlooked in discussions of algorithmic fairness, one of which is treating a prediction as a decision: a prediction says "what will happen," a decision says "what should be done" (Mitchell et al. 2021). Where there is no stage of fresh judgment, the two are pressed into one.
What programming may change are three things about questioning; whether it changes them depends on the configuration.
Cost. To question a parameter, one must first know where it is and what it is. It is in code, in a configuration file, in a technical document. When the parameters and their reasons are not public, the questioner must first obtain them, and the cost of questioning may rise.
Visibility. When a person makes a judgment, the trade-off shows in his reasons, and one can ask him why. A parameter shows only a number; without an associated record of reasons, why it is this number is not beside it, and the normative trade-off is not easily seen.
Manner of repetition. A person making a judgment each time may differ slightly each time, and may stop and think on some occasion. While the rules and parameters stay unchanged, a program is the same every time, and the same trade-off is carried out repeatedly. If a trade-off has a remainder, that remainder is copied in the same way, on a large scale.
S36 Automatic execution. In arrangements that execute automatically, without a corresponding stage of fresh judgment, every run of the program carries out that line of ought once; arrangements that embed review, appeal and human judgment are checked separately.
§3 The Presumption of Conformity
Article 40 of the EU Artificial Intelligence Act provides that high-risk AI systems which conform to harmonised standards whose references have been published in the Official Journal of the European Union are presumed to conform to the requirements set out in Section 2 of Chapter III of the Act, to the extent that those standards cover them (Regulation (EU) 2024/1689, Article 40(1)). The object of the presumption, as the text is written, is the requirements set out in that Section, not every requirement.
In October 2026 the reference to the first harmonised standard was published: EN 18286:2026, on quality management systems for artificial intelligence (Commission Implementing Decision (EU) 2026/2274). The decision maps it to item 9 of Annex I of the European Commission's standardisation request, the provider's quality management system, including post-market monitoring; which articles of the Act it covers, the decision does not state article by article. The European Commission separately stated that providers implementing this standard obtain "a presumption of conformity with the Article 17 requirements covered by the standard"; the page also says: "It is mapped to Article 17 of the AI Act" (European Commission, 19 October 2026). The object of the presumption in Article 40(1) is the requirements set out in Section 2 of Chapter III; Article 17 is in Section 3 of Chapter III. A sentence added to paragraph 2 of the same Article in 2026 has the Commission request standardisation organisations to draw up standards facilitating joint compliance with, and presumption of conformity for, the requirements or obligations of Sections 2 and 3 of Chapter III (Regulation (EU) 2026/1744). Which provisions this standard covers exactly, and with what limits, depends on the standard's own correspondence annex, which this paper has not read.
The presumption itself has legal authorization and its own legal effect: a system that conforms to this standard is presumed to conform on the requirements the standard covers. On this alone, misrouting is not found. The presumption is an arrangement the legislator made deliberately: it gives a limited review a definite consequence.
To prove misrouting, two things must be checked. First, whether the item questioned falls within the scope of the standard used and of what this review actually handled. As for EN 18286 specifically, this paper has not fully checked the full text of the standard and its correspondence annex, and cannot first list what it handled and what it did not. Second, whether the presumption has been used to register this item as answered: for instance, a document saying "this system conforms to the harmonised standard, and the relevant requirements are met," without stating which provision is met. The first depends on the actual scope of the standard and the review, which this paper has not fully checked; the second depends on how the presumption is used, and within the documents and search scope listed in this paper no corresponding record was found; the existing material cannot determine why none was found.
The presumption also has a channel for challenge. A member state or the European Parliament that considers a harmonised standard does not entirely satisfy the requirements it aims to cover can raise a formal objection; the Commission can maintain, restrict or withdraw its reference (Regulation (EU) No 1025/2012, Article 11). The decision on the storage tank standard in Paper 1's Introduction is one time this channel was used (Commission Implementing Decision (EU) 2017/1996).
S37 The presumption of conformity and misrouting. The presumption of conformity for harmonised standards rests on the corresponding provisions of Article 40 of the Artificial Intelligence Act, and is limited to the requirements or obligations the standard actually covers. The presumption itself has legal authorization and its own legal effect, and misrouting is not found on that basis; to prove misrouting one must check which challenge was not handled by the relevant standard and the specific review, and whether the presumption was used to register that challenge as answered.
§4 What Mathematics Can Show
On algorithmic fairness there is a set of mathematical results, often cited to show that "fairness cannot be achieved." This section takes only the theorems, and says what they show.
Kleinberg, Mullainathan and Raghavan formalized three fairness conditions: scores are calibrated within each group; those truly in the positive class receive the same average score across groups; those truly in the negative class do too. They proved that "except in highly constrained special cases, there is no method that can satisfy these three conditions simultaneously" (Kleinberg et al. 2016); the special cases are perfect prediction, or equal base rates in the two groups. Chouldechova proved a result in the same direction: when the base rates of two groups differ and the instrument is not a perfect classifier, an instrument satisfying predictive parity cannot at the same time equalize the false positive rates and the false negative rates of the two groups (Chouldechova 2017).
What these results show is this: a choice must be made. Several kinds of fairness that one would like to have cannot, in general, be had together; choosing one gives up part of another. They do not show which one to choose. Which to choose depends on which error weighs more on whom, and which inequality is less acceptable; this is a matter for the normative ledger, and the theorems do not choose in anyone's place.
This joins Paper 3. If a technical review proves that "this model satisfies calibration," and this conclusion is taken to answer "is this model fair," the scope of the review and the matter asked do not match: calibration is one of several kinds of fairness, and choosing it rather than another is a trade-off the review did not handle.
S38 Mathematics shows they cannot all be had, and does not choose for anyone. Several fairness conditions cannot, in general, be satisfied at once; this result shows that a choice must be made, not which one.
§5 Bearing Is Not Having Questioned
Once a program is deployed, someone always bears it. Paper 7 of The Second Enlightenment established this: when rules are written into a program, the one who decides on deployment sets down in advance a set of conditional dispositions, and "this judgment of deployment is his to bear in full, and he cannot say 'the program decided it' or 'the standard told us to write it'" (The Second Enlightenment, Paper 7, W44). Theory of Government said the same thing on the side of the construct: "When judgment is handed to a machine, the bearing is not handed over with it." (Theory of Government G8)
These two sentences speak of bearing. That bearing has found its place does not mean that the premises in the parameters have been questioned. One who decides on deployment may make the judgment of deployment without having sufficiently questioned the premises of the cut-off; this shortfall in verification is not made good by "someone already bears it." He still answers for the judgment of adopting this set of parameters and letting it act on persons accordingly: the deployer need not be the original author of the parameters, nor give a full guarantee of their scientific correctness, but for adopting them, and for how he uses them, he must answer. Bearing and having questioned are two things, recorded separately.
So, though the deploying side bears it, the questioning position can still be empty. The question beforehand, that is, whether before deployment the premises of the parameters have anywhere to be questioned, is in Paper 9 (S46).
S39 Bearing is not having questioned. That someone bears the judgment of deployment does not prove that the premises in the parameters have been sufficiently questioned.
§6 With Passi and Barocas
Passi and Barocas's 2019 paper starts from the actual work of data scientists and traces how a project is formulated, step by step, from a business goal into a problem that can be predicted. Two of their findings are closest to this paper. First, problem formulation is not done once; it is negotiated and revised again and again, and each revision changes what is to be predicted. Second, questions of fairness already appear at this step: which target variable is chosen, which proxy is used, already decides who will be treated how, without waiting for the model to be trained (Passi & Barocas 2019).
This paper is close to theirs, but asks about the other side. What they ask is: how is a formulation made, who makes it, under what constraints. What this paper asks is: once the formulation has been made and written into a program, if someone wants to question it, where do they go; once they ask, who can require an answer; can a review be initiated. They write of the process of formulation; this paper writes of the addresses after formulation.
The two can be joined. Their research shows that those lines on the normative ledger are written in through a process that can be observed; this paper says that once written in, those lines appear in the shape of parameters, and the cost of questioning them, the visibility of the trade-offs they embody, and the way those trade-offs are repeatedly carried out all change. So if the process of formulation leaves records, of what each revision changed and why, these records are the addresses for later questioning. This paper does not say how records ought to be kept; that is a matter for the rule layer, and Paper 9 offers it only as suggestion.
Where This Paper Stops
This paper does not say that any particular algorithm is unfair; it does not say that any particular law is wrong.
What it hands to Paper 8: if there is a seat to do the questioning, how long can it stay seated.
The remainder it leaves: how the presumption of conformity is used, no record was found within the listed scope; which provisions of Article 17 EN 18286 covers exactly, and how it connects with "the requirements set out in Section 2" of Article 40(1), the decision itself does not say, and this paper has not read the standard's correspondence annex; whether underwriting programs have a stage of fresh judgment, this paper has described only the general structure and has not checked lender by lender.
The Five Propositions of This Paper
- S35 The ought in parameters. Target variable, proxy variable, threshold and fairness metric each record a line; once written into a program they appear in the shape of parameters. (Descriptive)
- S36 Automatic execution. In arrangements with no stage of fresh judgment, every run carries out that line once; those with review, appeal and human judgment are checked separately. (Descriptive)
- S37 The presumption of conformity and misrouting. The presumption rests on the corresponding provisions and is limited to the requirements or obligations the standard actually covers; the presumption has authorization and effect, and misrouting is not found on that basis; to prove misrouting, check which challenge was not handled by the standard and the review, and whether the presumption was used to register it as answered. (Descriptive)
- S38 Mathematics shows they cannot all be had, and does not choose for anyone. Several fairness conditions generally cannot be satisfied at once; this shows that a choice must be made, not which one. (Descriptive; Kleinberg et al. 2016, Chouldechova 2017, theorems only)
- S39 Bearing is not having questioned. That someone bears the judgment of deployment does not prove that the premises in the parameters have been sufficiently questioned. (Conceptual distinction, own undertaking; The Second Enlightenment, Paper 7, W44, and Theory of Government G8 are taken over only for "the bearing is not handed over with it"; the question beforehand is in Paper 9, S46)
References
All SAE works below are cited by concept DOI, which always resolves to the latest version.
I · This system
- Qin, H. (2026). At the Door (SAE The Second Enlightenment, Paper 7) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24030415
- Qin, H. (2026). Judgment (SAE Theory of Government, Paper 1) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23412130
- Qin, H. (2026). Misrouting (SAE Theory of Disciplines, Paper 3) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24177759
- Qin, H. (2026). Shelving: Treating Premises as Absent (SAE Theory of Disciplines, Paper 5) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24177775
- Qin, H. (2026). Withdrawal (SAE Theory of Disciplines, Paper 6) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.24177783
II · Rule texts
- Commission Implementing Decision (EU) 2017/1996. Official Journal of the European Union.
- Commission Implementing Decision (EU) 2026/2274 (publication of the reference to EN 18286:2026). Official Journal of the European Union.
- Regulation (EU) No 1025/2012 on European standardisation, Article 11.
- European Commission. (2026, October 19). First harmonised standard supporting the AI Act published in the Official Journal. digital-strategy.ec.europa.eu.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 17, 40 (consolidated text of 27 July 2026).
- Regulation (EU) 2026/1744 of 8 July 2026, amending Regulation (EU) 2024/1689 (the sentence added to Article 40(2)).
III · External literature
- Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163.
- Kleinberg, J., Mullainathan, S., & Raghavan, M. (2016). Inherent trade-offs in the fair determination of risk scores. arXiv:1609.05807 (published in ITCS 2017).
- Mitchell, S., Potash, E., Barocas, S., D'Amour, A., & Lum, K. (2021). Algorithmic fairness: Choices, assumptions, and definitions. Annual Review of Statistics and Its Application, 8, 141–163.
- Passi, S., & Barocas, S. (2019). Problem formulation and fairness. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 39–48). ACM.
- Selbst, A. D., Boyd, D., Friedler, S. A., Venkatasubramanian, S., & Vertesi, J. (2019). Fairness and abstraction in sociotechnical systems. In Proceedings of the Conference on Fairness, Accountability, and Transparency (pp. 59–68). ACM.
Interfaces
Main upstream: The Second Enlightenment, Paper 7, W44; Theory of Government G8; Paper 3, S16, S17; Paper 5, S29; Paper 6, S34.
Also used: Articles 17 and 40 of the EU Artificial Intelligence Act and the publication of EN 18286; Article 11 of Regulation 1025/2012; Kleinberg et al. 2016; Chouldechova 2017; Passi & Barocas 2019; Selbst et al. 2019; Mitchell et al. 2021.
This series: S35 to S39.
Handed to Paper 8: if there is a seat to do the questioning, how long can it stay seated.