Ch. 9: Discretion Points — Human Judgment in Automated Governance

“Discretion is a science of the understanding, to discern between falsity and truth, between wrong and right, between shadows and substance, between equity and colourable glosses and pretences.” Sir Edward Coke, The Institutes of the Laws of England, 1628

Separation Without Isolation

The OLRF’s separation principle, with AI as the fact-finder and the normative infrastructure as the evaluator, is among the architecture’s most important constitutional commitments. But what if the norm itself requires judgment? When the legislature has deliberately left the outcome open and has instructed the administering authority to weigh competing considerations, to assess proportionality in the individual case, to exercise discretion, does the separation principle require that AI withdraw entirely, leaving the official unsupported?

The answer is no. The separation principle does not require the isolation of human judgment from AI assistance. It requires the isolation of normative authority from probabilistic inference. These are different requirements, and the Discretion Point architecture is designed to satisfy both simultaneously: structured AI support for human discretionary decisions, without transferring normative authority to the AI system.¹

Forms of Human Judgment

Human judgment in administrative decision-making is not a single, undifferentiated activity. It takes distinct forms, each with different legal characteristics, different requirements for the quality of reasoning, and different implications for the role that AI can legitimately play. Four categories can be distinguished. They are not specific to any single legal tradition. They appear, under varying terminology, in every mature administrative law system.²

Bound decisions

are cases in which the law fully determines the outcome once the relevant facts are established. If the applicant’s income is below the threshold and the residency requirement is met, the benefit is granted. The official has no choice to make. The law prescribes the result. These decisions are the natural domain of fully automated evaluation: a Decision Tree can model the conditions, receive validated facts, and produce a deterministic outcome without any human intervention. This is precisely the class of decisions that §35a VwVfG permits for full automation.³

Discretionary decisions

are cases in which the law deliberately grants the administering authority a range of legally permissible outcomes and leaves the choice between them to the official’s judgment. The legislature has decided that the correct outcome cannot be specified in advance for all cases. It depends on the circumstances: the severity of the situation, the individual’s particular hardship, the proportionality of the measure, or the balance between competing public interests. The official must weigh these factors and choose. A different official, weighing the same factors, might legitimately arrive at a different result. What the law requires is not a specific outcome but a defensible process of reasoning: that the official considered the relevant factors, stayed within the legal limits, and did not act arbitrarily.⁴

Assessment margins

arise where the law uses terms whose application requires expert evaluation. Whether a building is “structurally sound,” whether a substance poses a “significant risk to public health,” or whether an applicant is “professionally reliable” cannot be resolved by applying a fixed rule to fixed facts. The law delegates the evaluation to an authority with the requisite expertise and, within limits, shields that evaluation from full judicial substitution. The boundary between what falls inside the assessment margin and what can be judicially reviewed is itself a subject of ongoing legal development in most jurisdictions.

Proportionality assessments

overlap with discretion but deserve separate mention because they involve a structured form of reasoning that most legal systems recognise as distinct. The official must assess whether the measure under consideration is suitable to achieve its purpose, whether it is necessary (or whether a less intrusive alternative would suffice), and whether the burden it imposes on the individual stands in reasonable proportion to the benefit it achieves for the public interest. This is not a binary test. It is a structured weighing exercise in which the same facts can support different conclusions depending on how the competing considerations are valued.

These four categories are not mutually exclusive. A single decision may involve bound elements (the threshold conditions that must be met before discretion arises), a discretionary element (the choice of measure), an assessment margin (the evaluation of a technical term on which the discretion depends), and a proportionality assessment (the justification that the chosen measure is proportionate). What matters for the OLRF architecture is that each category imposes different requirements on the decision-maker and therefore defines a different boundary for legitimate AI involvement.

Mapping to Legal Frameworks

These categories find an archetypal statutory expression in the German Administrative Procedure Act (VwVfG), which is why the OLRF uses it as a primary reference. The mapping is instructive because it demonstrates how abstract decision-making categories translate into concrete legal requirements, but it is not imparative for this Framework to reference it to German law.

§35a VwVfG, introduced in 2017, permits fully automated administrative acts under two cumulative conditions: the applicable norm must not confer discretion (Ermessen) on the authority, and it must not involve an assessment margin (Beurteilungsspielraum). The legislative history is explicit: automation is permitted only for bound decisions whose outcome follows deterministically from the application of legal conditions to established facts.

§40 VwVfG provides the framework for how discretion must be exercised. Where a public authority acts at its discretion (nach Ermessen), it must do so in accordance with the purpose of the authorising provision and within the legal limits (pflichtgemäß). The Federal Administrative Court has elaborated this into a detailed doctrine of discretionary errors (Ermessensfehlerlehre) that defines four ways in which an otherwise lawful exercise of discretion becomes judicially reviewable:

Ermessensausfall (failure to exercise discretion): treating a discretionary decision as if it were bound. An official who allows an AI system to make the discretionary determination and merely ratifies it commits Ermessensausfall, regardless of formal human sign-off. This is the rubber-stamping problem, and it is the primary constitutional risk of poorly designed AI assistance at Discretion Points.

A borderline case, analysed by Braun Binder, arises where administrative guidelines (ermessenslenkende Verwaltungsvorschriften) narrow the discretionary margin to zero through self-binding: the parameters can in principle be encoded in an algorithmic decision tree, but the constitutional legitimacy of treating the result as fully bound remains contested⁵.

Ermessensüberschreitung (exceeding discretion): choosing an outcome beyond the legal limits of the authorising provision.

Ermessensunterschreitung (under-exercise of discretion): applying a blanket rule without genuine consideration of the individual case.

Ermessensfehlgebrauch (misuse of discretion): exercising discretion for purposes other than those intended by the authorising provision, or without adequate consideration of relevant factors.

§39 VwVfG adds a procedural requirement: discretionary administrative acts must be accompanied by a written statement of reasons (Begründungspflicht) that discloses the essential legal and factual grounds of the decision. For discretionary decisions, courts interpret this as requiring a genuine account of the considerations that led to the specific exercise of discretion, not a generic recitation of the authorising provision.

Other legal systems⁶ express similar structural distinctions in different terminology. French administrative law distinguishes compétence liée from pouvoir discrétionnaire. English administrative law works with Wednesbury reasonableness and proportionality review. EU law applies the proportionality principle under Article 5 TEU with a three-part test. The OLRF’s Discretion Point architecture is designed to be jurisdiction-neutral in its structure while being mappable to any of these frameworks through the sub-normative linkage system⁷.

Discretion Points Across Three Models

The function of Discretion Points shifts across the three models, though their constitutional purpose remains constant: wherever the legislature has reserved a decision for human judgment, the automated system must enforce that reservation.

In Model A, the Discretion Point is reached when the deterministic evaluation engine encounters a node classified as discretionary in the Decision Tree. The engine pauses. The AI agent layer has assembled the facts. The tree has evaluated all deterministic conditions. What remains is a judgment that only a human can exercise.

In Model B, the Discretion Point is reached either by the Decision Tree’s validation framework (when it encounters a discretionary node during validation of the Legal Agent’s determination) or by the Legal Agent itself (when the agent recognises that the statutory text requires human judgment). In Model B, the agent may have already formed a preliminary view of the case, but the Discretion Point architecture requires that this view not be presented as a recommendation to the official. The agent’s analysis feeds into the AI-Assistance-Package as factual and legal context, not as a proposed outcome.

In Model C, the Legal Agent identifies Discretion Points during its autonomous reasoning and escalates them for human decision. Because the agent is applying the statutory text directly (without the Decision Tree’s structure to identify discretionary nodes), the agent’s ability to recognise where human judgment is constitutionally required is itself a critical capability that the Agent Capability Attestation (Control 5) must certify.

An agent operating under Model C must be certified not only for autonomous legal reasoning but also for correct escalation behaviour. The distinction is critical. The hardest test of an autonomous agent is not whether it can reason correctly in clear cases. It is whether it can recognise when a case exceeds the boundary of autonomous reasoning and must be escalated to a human decision-maker. An agent that produces a confident-seeming determination in a case that should have been escalated is more dangerous than an agent that produces an incorrect determination in a case within its competence, because the former error is invisible: the system appears to function correctly while the constitutional safeguard has been silently bypassed. For this reason, the certification test suite for Model C (Chapter 10) includes scenarios in which the correct response is to escalate rather than to decide. An agent that fails to escalate in these scenarios fails the certification, regardless of the quality of its reasoning in other scenarios. Escalation competence is not an optional supplement to reasoning competence. It is its constitutional complement.

In all three models, a workflow that attempts to skip a Discretion Point cannot produce a valid, signed Decision Package. This enforcement is constant across the entire architecture.

The AI-Assistance-Package: Architecture and Constraints

When the evaluation reaches a Discretion Point and pauses, the official who must exercise judgment does not receive a blank screen. The OLRF proposes a structured information environment called the AI-Assistance-Package: a set of legally relevant materials, assembled by AI tools and organised for efficient human review.⁸

The AI-Assistance-Package is not a recommendation. It does not suggest an outcome, rank alternatives, or indicate a preferred direction. It provides the official with the legal and factual context needed to exercise well-informed discretion, while preserving the constitutional requirement that the discretionary choice itself remains entirely human. Its design is governed by five constraints that translate constitutional principles into concrete architectural requirements.

Constraint 1: No recommendation, no weighting. The Package shall not contain a recommended outcome, a preferred option, or any weighting of alternatives. An AI system that presents a “recommended” outcome to a human who then ratifies it has, in functional terms, made the decision. The Package presents the range of legally permissible outcomes. It does not indicate a destination within that range.

Constraint 2: Source attribution for every claim. Every legal proposition in the Package must be attributed to a specific, verifiable source. Every assertion about what courts have recognised as proportionate, every identification of an analogous precedent, every summary of relevant administrative guidelines must be traceable to its origin. An AI-generated synthesis of case law that cannot be traced to specific judgments is epistemically untrustworthy for the purposes of proper discretionary reasoning.

Constraint 3: Separation of law from fact. The Package must present the established facts and the applicable legal framework as distinct, clearly delineated components. The official must be able to see, without ambiguity, what the AI system established as fact and what the legal framework requires as a matter of law. The subsumption (the application of law to fact) is the official’s task, not the Package’s.

Constraint 4: Transparency of limitations. The Package must disclose what it does not know. Where the case law corpus is incomplete, where the precedent population is small, where the proportionality corridor is based on limited data, the Package must say so explicitly. An official who relies on an AI-generated analysis without being informed of its limitations cannot exercise informed judgment.

Constraint 5: Timeliness of legal sources. The Package must disclose the date of the most recent legal source included and the date of the last corpus update. An official must be able to assess whether the Package reflects the current state of the law or whether relevant recent developments may have been missed.

Tools for the Different Forms of Judgment

The AI-Assistance-Package is assembled by four tools, each addressing a different dimension of the discretionary decision.

1: retrieve_case_law.

This tool queries the relevant case law databases for judicial decisions applying the same provision in factually similar circumstances. It returns structured summaries with full citation, identifying the legal principles applied by the court, the factual circumstances considered determinative, and the outcome. The output is organised to display the spectrum of judicial responses to comparable situations, enabling the official to understand the legal corridor within which their discretion operates.

Tool 2: analyse_precedents.

This tool queries the authority’s own anonymised decision record: the population of Discretion Point decisions previously made by the same authority under the same provision. It identifies factually similar cases and the range of discretionary outcomes that were chosen. Its purpose is twofold: to support consistency in the exercise of discretion across similar cases, as required by the principle of equal treatment;⁹ and to provide the official with an empirical picture of how discretion has been exercised in practice. The output distinguishes between decisions that were subsequently challenged and upheld, decisions that were challenged and quashed, and decisions that were not challenged.

Tool 3: map_proportionality_corridor.

This tool builds a structured representation of the proportionality corridor applicable to the discretionary decision[^118]: the range of outcomes that courts have recognised as proportionate under the applicable legal framework, and the factors identified as relevant to assessing proportionality. The output is a structured spectrum. At one end stand the outcomes that courts have recognised as clearly proportionate even in the most demanding circumstances. At the other end stand the outcomes that courts have recognised as disproportionate even in the least demanding circumstances. In the middle lies the range whose proportionality depends on the specific weight of the competing considerations in the individual case. The official’s task is to place the present case within this spectrum. This requires genuine judgment.¹⁰ The spectrum supports it. It does not pre-empt it.

Tool 4: check_consistency. This tool performs a statistical analysis of the Discretion Point decision population to identify whether the outcome being contemplated is within the normal range for factually similar cases, or whether it represents a significant departure. Its purpose is not to prevent departures. Justified departures are a legitimate and legally required exercise of discretion in cases with distinctive circumstances. The purpose is to ensure that the official is aware of the departure and has consciously decided that the present case warrants it. Where the tool identifies a significant departure, it flags this explicitly and requests that the official’s reasoning address the distinction¹¹.

The four tools correspond to the four forms of human judgment. Tools 1 and 3 directly support proportionality assessments by mapping the legal boundaries within which the weighing exercise must take place. Tool 2 supports the consistency dimension of discretionary decisions by making the authority’s own practice visible. Tool 4 operates as a safeguard against both Ermessensunterschreitung (blanket application without individual consideration) and unintentional departure from the equal treatment principle.

The Discretion Point Lifecycle

The complete lifecycle proceeds through five stages. Each generates a structured artefact that becomes part of the Decision Package.

Stage 1: Pause and notification. The evaluation reaches a node classified as a Discretion Point. Evaluation is suspended. A structured notification is generated to the responsible official, including the norm anchor, the established facts, and an estimated complexity classification.

Stage 2: AI-Assistance-Package generation. The four tools are invoked. The Package is assembled and delivered to the official in a structured interface that presents the outputs clearly labelled, source-attributed, and explicitly marked as informational rather than recommendatory.

Stage 3: Official review and decision. The official reviews the Package and exercises judgment. The interface requires completion of a structured reasoning record: the legal framework applied, the factual circumstances considered, the considerations weighed, the assessment of proportionality, the discretionary choice made, and the reasoning for that choice (with a minimum length requirement as a structural obstacle to rubber-stamping). The time taken is logged as part of the audit trail.

Stage 4: Discretion Point output. The completed reasoning record, together with the Package that informed it, is assembled into a typed Discretion Point Output carrying the official’s identity, authentication credentials, timestamp, and cryptographic signature. This is attached to the Decision Package as a distinct component alongside the evaluation engine’s deterministic outputs. The distinction between the two is preserved: deterministic results are signed by the evaluation engine, discretionary outputs are signed by the human official.

Stage 5: Evaluation resumes. The Discretion Point Output is submitted to the evaluation engine as a fact input: typed, schema-conformant, and carrying the official’s signature as provenance. The engine treats it as it treats any other fact input. It does not re-evaluate it. It uses it as input to the next step in the Decision Tree. The human decided. The engine continues.

Discretion Point Types

The AI-Assisted Discretion Point (Type B) is one of three types that the OLRF specification recognises. The taxonomy establishes that AI assistance is an option, not a requirement.

Type A: Unassisted.

The official receives the established facts but no AI-generated assistance. Appropriate for highly context-specific discretionary authorities where structured case law retrieval would provide limited value, or where the case law corpus is too sparse to support reliable analysis.

Type B: AI-Assisted.

The full AI-Assistance-Package is generated. This is the primary type for discretionary authorities with established case law, identifiable precedent populations, and assessable proportionality corridors. It covers the large majority of recurring administrative discretionary decisions.

Type C: Escalated.

The decision is escalated beyond the processing official to a supervising official, a collegial body, an independent review panel, or (where constitutionally required) a court. Appropriate for decisions with precedent-setting implications, genuine legal uncertainty requiring authoritative resolution, or fundamental rights restrictions requiring heightened procedural protection.

The classification of a Discretion Point as Type A, B, or C is not fixed at the time the Decision Tree is published¹². It can evolve as the legal landscape develops. A Discretion Point that begins as Type A (because no case law exists) may be reclassified to Type B once a sufficient body of precedent has accumulated. A Discretion Point that begins as Type B may be reclassified to Type C if judicial decisions reveal that the discretionary question is more legally contested than initially assumed. These reclassifications are documented in the Registry’s version history and reflected in the Coverage Map.

The Rubber-Stamping Safeguard

The rubber-stamping problem is the primary constitutional risk of the AI-Assisted Discretion Point architecture. It arises when AI assistance becomes, in practice, AI decision-making: officials ratify AI-generated outputs without exercising genuine independent judgment. If this happens systematically, the formal presence of a human in the loop does not prevent Ermessensausfall. The decision is functionally automated regardless of the signature at the bottom.¹³

The OLRF draft addresses this risk through two complementary mechanisms operating at different scales.

Structural prevention

Operates at the individual decision level through the design constraints on the Package and the reasoning record interface. Four features work in combination. The prohibition on recommendations ensures that the Package does not present a default outcome that the official can simply accept. The source attribution requirement ensures that the official receives identifiable legal materials rather than an opaque AI synthesis, making passive reliance harder to sustain. The structured reasoning record requires documentation of which considerations were weighed and why the chosen outcome was selected, with a minimum length requirement as a structural obstacle. And the enforced engagement time is logged, providing an auditable data point.

These features do not guarantee genuine engagement. No architectural mechanism can guarantee the quality of human judgment. What they can do is ensure that rubber-stamping requires active effort rather than passive acquiescence. The path of least resistance through the Discretion Point interface requires the official to engage with the legal framework, address the relevant considerations, and document a reasoned choice.

Audit detection

Operates at the population level through the Coverage Map’s analysis of Discretion Point outputs across all cases processed under a given Decision Tree. Three signals are monitored:

Outcome clustering: where the distribution of discretionary outcomes is significantly more concentrated than the legal framework would predict, given the diversity of the underlying factual circumstances.

Reasoning similarity: where the free-text reasoning records for a population of decisions exhibit high textual similarity, suggesting that officials are copying or minimally adapting a standard text rather than documenting individual judgment.

Completion time anomalies: where a significant proportion of decisions are completed in times inconsistent with genuine engagement with the Package and the reasoning record requirements.

None of these signals is conclusive in isolation. They become meaningful when they occur in combination, or when their magnitude exceeds what the legal and factual context can plausibly explain. Where audit detection identifies these signals, the Coverage Map flags the relevant Discretion Point type for quality review, triggering an audit of a sample of individual decisions. Where the pattern is sufficiently concerning, it triggers escalation to the relevant parliamentary oversight body.

The combination of structural prevention and audit detection reflects a realistic assessment of what architectural design can and cannot achieve. It cannot force officials to think carefully. It can make carelessness visible, traceable, and consequential.¹⁴

Martini, M. and Nink, D., “Wenn Maschinen entscheiden: Vollautomatisierte Verwaltungsverfahren und der Persönlichkeitsschutz,” NVwZ-Extra 10/2017, pp. 1—14 (analysing the conditions under which AI assistance at discretion points is compatible with both § 40 VwVfG and Art. 22 GDPR). ↩
Davis, K.C., Discretionary Justice: A Preliminary Inquiry, Louisiana State University Press 1969 (the foundational treatment arguing that the central problem of administrative law is not the elimination of discretion but its structuring through visible standards and reviewable procedures) ↩
Deutscher Bundestag, Entwurf eines Gesetzes zur Modernisierung des Besteuerungsverfahrens, BT-Drucksache 18/7457 vom 3. Februar 2016, S. 48 ff. See also: Prell, L., “§35a VwVfG: Vollständig automatisierter Erlass eines Verwaltungsaktes,” PUBLICUS 2017.3; Rath, C. and Breidenbach, S., “Einsatz von KI in der Verwaltung: Bremst das VwVfG?,” Legal Tribune Online (LTO), 2025. ↩
Stelkens, P., Bonk, H.J., and Sachs, M. (eds.), Verwaltungsverfahrensgesetz: Kommentar, 10. Aufl., C.H. Beck 2023, § 40 Rn. 1—120. ↩
Braun Binder, N., DÖV 2016, pp. 891 (894); see also Martini, M. and Nink, D., “Wenn Maschinen entscheiden … vollautomatisierte Verwaltungsverfahren und der Persönlichkeitsschutz”, NVwZ-Extra 10/2017, pp. 1 (10 ff.), who use the term “algorithmischer Entscheidungsbaum” for precisely the structure that the OLRF’s Decision Tree formalises. The OLRF treats this borderline case through the Discretion Point architecture: a norm element that is formally discretionary but practically narrowed to zero by administrative guidelines is classified as a Type A Discretion Point (parametrised) rather than as a bound element, preserving the constitutional distinction while permitting deterministic evaluation where the parameters are stable. ↩
For the comparative analysis of how different legal systems structure and review administrative discretion: Ranchordás, S. and de Waard, B. (eds.), The Judge and the Proportionate Use of Discretion: A Comparative Administrative Law Study, Routledge 2016 (with dedicated chapters on German administrative law (Marsch/Tünsmeyer), French administrative law (Sanchez), English law (Davies/Williams), Dutch administrative law (de Waard), EU proportionality (Haguenau-Moizard/Sanchez), and US administrative law (Mathews), demonstrating that all systems recognise a structural distinction between bound and discretionary decisions, while differing substantially in the intensity and methodology of judicial review) ↩
Mathews, J., “Proportionality Review in Administrative Law,” in: Cane, P. et al. (eds.), Oxford Handbook of Comparative Administrative Law, 2nd ed., Oxford University Press 2020 (identifying three principal axes of variation across jurisdictions: how extensively proportionality review is employed, how intensively it is conducted, and how discursively courts present their analysis). Mathews shows that the German Ermessensfehlerlehre, the French erreur manifeste d’appréciation, the English Wednesbury unreasonableness (progressively converging with proportionality since the Human Rights Act 1998), and the EU three-part proportionality test under Art. 5 TEU all serve the same structural function (constraining discretion while preserving the decision-maker’s legitimate range of choice) through different doctrinal instruments. The OLRF Discretion Point architecture is designed to be mappable to all of these frameworks through the sub-normative linkage system, because the underlying structural distinction (between bound and discretionary elements of a norm) is constant across all of them, even where the terminology and the intensity of judicial review differ. ↩
The concept of structured AI assistance for discretionary decisions, as distinct from AI recommendation or AI decision-making, builds on the distinction between decision support and decision automation established in Parasuraman, R. et al., “A Model for Types and Levels of Human Interaction with Automation,” IEEE Transactions on Systems, Man, and Cybernetics, Vol. 30, No. 3, 2000, pp. 286—297. Parasuraman’s taxonomy identifies ten levels of automation, ranging from “the computer offers no assistance” through “the computer suggests an alternative” to “the computer decides everything.” The OLRF’s AI-Assistance-Package is designed to operate at Parasuraman’s Levels 2—4 (the computer offers a complete set of alternatives, or narrows the selection down to a few), explicitly excluding Levels 5 and above (the computer suggests one alternative, or executes automatically). This positioning is a constitutional requirement, not a design preference. ↩
The constitutional principle of equal treatment in the exercise of discretion (Selbstbindung der Verwaltung) derives from Art. 3 Abs. 1 GG: an authority that has exercised its discretion in a certain way in a population of cases creates a legitimate expectation that it will exercise it in the same way in comparable future cases, unless there is a justifiable reason for departure. See: BVerwGE 34, 278 (280) (1970); BVerwGE 92, 153 (155) (1993). ↩
Barak, A., Proportionality: Constitutional Rights and Their Limitations, Cambridge University Press 2012, especially Chapters 7—10 (developing a four-stage framework: proper purpose, rational connection, necessity, and proportionality stricto sensu, each stage narrowing the range of permissible outcomes). For the specific connection between proportionality analysis and administrative discretion in German law: Alexy, R., “Ermessensfehler,” JuristenZeitung 41 (15/16), 1986, pp. 701—716 (the foundational analysis demonstrating that proportionality and the doctrine of discretionary errors are structurally related: a disproportionate exercise of discretion is always also an Ermessensfehlgebrauch, which is why the proportionality corridor and the permissible range of discretionary outcomes are, in constitutional terms, the same space). ↩
Maurer, H. and Waldhoff, C., Allgemeines Verwaltungsrecht, 20. Aufl., C.H. Beck 2020, § 24 Rn. 21—26. The OLRF’s check_consistency tool operationalises this principle computationally: by comparing a contemplated discretionary outcome against the statistical distribution of prior outcomes for factually similar cases, it makes the self-binding effect of the authority’s own practice visible to the official at the moment of decision rather than reconstructable only in retrospect during judicial review. ↩
The German Bundesverfassungsgericht has articulated this as the Wesentlichkeitstheorie applied to procedural design: the more significantly a decision affects fundamental rights, the more demanding the procedural requirements that the constitution imposes on the decision-making process. See: BVerfGE 49, 89 (126—127) (Kalkar I, 1978); BVerfGE 98, 218 (251—252) (1998). In English administrative law, the same principle appears as the doctrine of variable intensity of review: the more significant the interference with individual rights, the more the court will require of the decision-maker by way of justification. See: R (Daly) v Secretary of State for the Home Department [2001] UKHL 26, per Lord Steyn at [28] (establishing that the intensity of review is context-dependent and must be calibrated to the nature and gravity of what is at stake). The OLRF’s three-type taxonomy translates this graduated principle into architectural form: Type A (unassisted) is appropriate where the discretionary question is context-specific and the risk of AI-induced bias outweighs the benefit of structured support. Type B (AI-assisted) is the default for recurring decisions where established case law, precedent populations, and proportionality corridors make structured support both feasible and constitutionally valuable. Type C (escalated) is reserved for decisions whose precedent-setting implications, legal uncertainty, or fundamental rights significance require the authority and institutional legitimacy of a senior official, collegial body, or court. The dynamic reclassification of Discretion Points over time (from Type A to B as case law accumulates, from B to C as legal uncertainty deepens) reflects the insight that the appropriate procedural form is not fixed by the norm but evolves with the legal landscape in which the norm is applied. For the theoretical foundations of matching procedural form to decision weight: Mathews, J. and Sweet, A.S., “All Things in Proportion: American Rights Doctrine and the Problem of Balancing,” Emory Law Journal, Vol. 60, 2011, pp. 797—875 (arguing that the intensity of procedural and substantive scrutiny should be calibrated to a sliding scale rather than assigned to categorical tiers, and that proportionality provides the analytical framework for this calibration). ↩
Zweig, K.A., Ein Algorithmus hat kein Taktgefühl: Wo künstliche Intelligenz sich irrt, warum uns das betrifft und was wir dagegen tun können, Heyne 2019, Kapitel 8—9 (arguing that whether a human official exercises genuine judgment or merely ratifies a machine-generated recommendation is architecturally, not behaviourally, determined). ↩
For the broader problem of “automation bias” (the tendency of human operators to defer to automated suggestions even when those suggestions are incorrect): Skitka, L.J. et al., “Does Automation Bias Decision-Making?,” International Journal of Human-Computer Studies, Vol. 51, No. 5, 1999, pp. 991—1006. Skitka’s findings (that automation bias occurs even among experienced professionals and even when the automated suggestion is clearly erroneous) are the empirical foundation for the OLRF’s design choice to prohibit recommendations in the AI-Assistance-Package (Constraint 1) and to use population-level audit detection as a second line of defence against the individual-level effects that structural prevention alone cannot eliminate. ↩