--- title: The Mode System section: Ora — Foundation arguments status: review description: Frameworks specify what the system does; modes specify how it reasons — twenty-one territories and sixty resident modes that match the analytical approach to the problem. authors: - The Ora Foundation downloads: md: /papers/white/the-mode-system.md license: https://creativecommons.org/publicdomain/zero/1.0/ --- # The Mode System ## Modes versus frameworks Frameworks specify what the system does. Modes specify how the system reasons through what it is doing. A framework is a structured specification: input format, processing steps, decision points, output format. The framework is shape — what the work looks like from outside, what its inputs and outputs are, what its decision graph is. A mode is a cognitive posture: a way of approaching reasoning that is appropriate to a particular kind of question. The mode is style — what the reasoning looks like from inside, what register the model operates in, what the model is being asked to surface or to suppress. The same framework can run in different modes. Argument Audit running in Steel Man mode produces a different output than Argument Audit running in Adversarial Audit mode, even though both are running the same framework. The framework defines the structure of the audit; the mode determines whether the audit emphasizes the strongest version of the position being audited or the strongest version of the case against it. Modes do not replace frameworks. They compose with them. The user picks both — the framework that fits the work and the mode that fits the cognitive posture the work requires. ## Why explicit modes A model running without an explicit mode is running in whatever default register its training disposed it toward. The default register varies by model and by training distribution; it is not stable across queries; it is not legible to the user as a register. A user who asks a model the same question twice might get answers in different registers without knowing why. Explicit modes make the register a parameter. The user picks the register; the system applies it; the output reflects the choice. This is reliability infrastructure: the same query in the same mode produces output of the same shape, which makes the output evaluable. It is also pedagogical infrastructure: the user who has explicitly run a query in Steel Man mode has learned what Steel Manning looks like by seeing the output, which trains the user's own Steel Man muscle for use without the system. The mode system is the AHI commitment applied at the reasoning layer. The system serves the human's directing intelligence rather than substituting its own choice of register for the human's. The user is the one who recognizes which posture the work requires; the system applies the posture the user picks. ## The modes Sixty resident modes cover the cognitive-posture space, organized across twenty-one analytical territories. The set grows as new postures prove worth distinguishing; the modes below are representative of the range, not the full roster. **Steel Man.** Represent the strongest version of an opposing position before evaluating it. The mode that prevents straw-manning. Useful for reading opinion writing, evaluating contested claims, working out what a position one disagrees with actually says at its best. **Devil's Advocate.** Argue against a position the user is inclined to accept, to surface what the position has not addressed. Different from Steel Man — Steel Man is about the opposing position; Devil's Advocate is about the user's own position. **Competing Hypotheses.** Hold multiple explanations simultaneously rather than collapsing to the first plausible one. Useful for diagnostic reasoning, evidence evaluation, and any case where premature commitment to a single explanation would close off productive lines of investigation. **Decision Architecture.** Lay out the structure of a decision so its tradeoffs are visible. Options, optionality structure, reversibility, authority. Useful for any case where the user will act on the output and needs to be able to defend the choice to people who were not in the loop. **Causal Investigation.** Trace the structure of a causal claim back to its supporting evidence. Useful when reading news stories, evaluating policy claims, working out whether a correlation has a credible mechanism behind it. **Mechanism Understanding.** Build a model of how something works, with the inputs, the processes, and the outputs articulated. Useful for explaining a system to oneself, evaluating whether a claim about a system is consistent with how the system actually operates. **Stakeholder Analysis.** Map who has interests in a situation, what those interests are, where they conflict, and how they are likely to act. Useful for negotiation, conflict resolution, organizational analysis, anything involving the structure of a contested situation. **Strategic Interaction.** Reason about how parties will respond to each other's moves, including the user's own. Game-theoretic in the loose sense — not formal payoff matrices, but the recognition that decisions are made in anticipation of others' responses. **Scenario Construction.** Generate distinct possible futures and trace the implications of each. Useful for planning, risk analysis, strategic exploration where the future is not single-valued. **Adversarial Audit.** Read a piece of work as if looking for what is wrong with it. Different from Steel Man — Steel Man builds up the strongest version; Adversarial Audit looks for the weakest. Useful for one's own drafts before they go out, for proposals one is being asked to approve, for arguments one is considering accepting. **Cross-Domain Synthesis.** Connect what is happening in one domain to what is happening in another, looking for structural analogies that might transfer insight. Useful when a problem in one's own domain looks like a problem already solved elsewhere. **Domain Induction.** Extract general principles from specific cases. Useful when the user has accumulated experience and wants to surface the patterns the experience reflects without having articulated them explicitly. **Pattern Detection.** Find recurring structures across a set of inputs. Useful for analyzing the user's own past work, identifying tendencies in a person's writing, surfacing what is consistent across superficially different cases. **Conceptual Clarification.** Take a contested concept and produce a careful definition that captures what is actually being claimed when the concept is invoked. Useful for arguments where the parties seem to disagree but might actually be using the same word for different things. **Hypothesis Evaluation.** Assess a specific hypothesis against the available evidence, scoring how well the hypothesis accounts for what is observed. Useful for diagnostic and inferential work. **Worldview Cartography.** Map the structure of a worldview — what it considers central, what it treats as peripheral, what its load-bearing commitments are, where its blind spots are likely to fall. Useful for engaging seriously with a tradition or position one does not share. **Future Exploration.** Reason about consequences over longer time horizons than ordinary planning addresses. Useful for civilizational-scale questions, strategic planning at multi-year scope, anticipating cascading effects. **Risk and Failure Analysis.** Identify the ways a plan or system can fail, the relative probability of each failure mode, the consequences of each, and the mitigations available. Useful for any work where the cost of failure is high enough to warrant explicit anticipation. **Historical Comparison.** Identify historical analogues for the current situation, with attention to what is structurally similar and what is structurally different. Useful for placing current events against the longer arc; useful for resisting the recency bias that treats the present as unprecedented. **Process Inference.** Discover the transformation path between defined endpoints when the process is unknown. Useful when the user knows what they have and what they want but not how to get from one to the other. These are working categories, refined through use. New modes get added as new postures prove worth distinguishing. The mode list itself is open to contribution — domain experts can specify new modes for their fields, and the modes that meet the specification standard enter the canonical mode library. ## The territory taxonomy Underneath the modes is a classification of cognitive work itself: twenty-one analytical *territories*, each a coherent region of analytical work with a shared problem-shape, input types, and output contract. Every mode lives in exactly one home territory; a candidate that seems to fit two is parsed into two modes rather than dual-homed (Pre-Mortem splits into an action variant under Future Exploration and a fragility variant under Risk and Failure Analysis). The twenty-one territories group into five super-clusters. The super-clusters are for orientation only — routing operates per-territory, not per-cluster: - **Argument and Reasoning** — examining a claim against evidence, building a case, evaluating contesting positions (Argumentative Artifact Examination, Paradigm and Assumption Examination, Conceptual Clarification). - **Causation, Hypothesis, and Mechanism** — tracing how things produce other things, weighing competing explanations, modeling how a system's parts produce its behavior (Causal Investigation, Hypothesis Evaluation, Mechanism Understanding, Process and System Analysis). - **Decision, Future, and Risk** — choosing under uncertainty, exploring forward, anticipating failure (Decision-Making Under Uncertainty, Future Exploration, Risk and Failure Analysis). - **Position, Stakeholder, and Strategy** — locating parties in a contested landscape and reasoning about their interaction (Stakeholder Conflict, Negotiation and Conflict Resolution, Strategic Interaction, Interest and Power Analysis). - **Synthesis, Orientation, Structure, and Generation** — producing new understanding from existing knowledge, orienting in unfamiliar terrain, mapping structure, evaluating an artifact by stance (Structural Relationship Mapping, Cross-Domain and Knowledge Synthesis, Orientation in Unfamiliar Territory, Artifact Evaluation by Stance). The taxonomy is generative rather than closed — territories are added as the system meets problem-shapes the existing set does not cover, and one territory is currently held in reserve for future use. The taxonomy is also pedagogical. A user who has worked across the territories knows that cognitive work falls into a recognizable set of regions, recognizes which territory a new problem belongs to, and applies the appropriate mode without having to be prompted. The taxonomy is a map; using the system is reading the map; reading the map repeatedly is internalizing the structure of cognitive work. ## Mode dispatch in operation When a user submits a problem, the dispatch is explicit. The pre-routing pipeline disambiguates the *territory* first, then selects a resident mode within it — territory routing is the selection layer — and shows the user the dispatch. Friction reducers skip any disambiguation the prompt has already answered, so a fully specified prompt dispatches with no questions at all. The user can accept the suggested mode, override it, or request a different one. The transparency of the dispatch is itself pedagogical. A user who has seen a hundred dispatches has learned the territory taxonomy by watching it operate. The user knows that questions about contested claims get Argument and Reasoning frameworks running Steel Man or Adversarial Audit modes; the user knows that questions about how things produce other things get Causation frameworks running Causal Investigation or Mechanism Understanding modes; the user knows that decisions get Decision and Future frameworks running Decision Architecture or Scenario Construction modes. The dispatch is teaching the user the structure of cognitive work in the act of using it. The user can ignore the suggested dispatch. The user might pick a different mode for a problem because the user has noticed something the router cannot — that the problem looks like an Argument and Reasoning question but is actually a Position and Strategy question disguised, or that the problem looks straightforward but requires Steel Manning before Competing Hypotheses can profitably run on it. The system serves the user's recognition; the suggestion is a starting point, not an authoritative pick. This is what AHI looks like at the reasoning-posture layer. The system makes the choice explicit so the user can direct it; the user's direction is what makes the system's output appropriate to the work. A system that picked the posture silently would be reasonable in many cases and unreasonable in many others, and the user would not be able to tell which. ## Why the user's pick matters The mode system's pedagogical function depends on the user picking. A user who has explicitly picked Steel Man mode for an opinion column has practiced the recognition that opinion columns benefit from Steel Manning. The practice transfers — the user reading a different opinion column without the system notices where the columnist refused to Steel Man the position they were dismissing. The recognition has been trained. A user who has been getting Steel Man output without ever having picked it has not practiced the recognition. The output may be good, but the user's cognitive capacity has not changed. The system has substituted its judgment for the user's; the user is no better at thinking about opinion columns than they were before the first transaction. The difference shows up in the long arc. A user who has been picking modes for years has internalized the cognitive postures the modes encode. The user has, by repeated practice with structured choice, become a better thinker about the problems the modes address. A user who has been receiving outputs without picking has accumulated nothing; the system has done the work, the user has consumed the output, and the user's cognitive capacity is the same as when they started. This is the pedagogical commitment in the mode system specifically. Every interaction trains the user as a side effect of producing useful output, but only because the choice is the user's. Take the choice away and the training disappears. ## What the mode system does not do The mode system does not eliminate the model's blind spots. A mode tells the model what posture to operate in; it does not give the model knowledge the model lacks. A user who picks Steel Man mode for a question about a subject the model is unreliable on still gets unreliable output, just in a Steel Man register. The mode system does not replace domain expertise. The user has to know enough about the problem to recognize which posture the problem requires. A user who picks Adversarial Audit on a problem that needed Steel Man first will get an audit of a position that has not yet been represented at its strongest. The mode list helps the user pick well; it does not pick for the user. The mode system does not produce certainty. Output produced in any mode is still subject to the same reliability properties as any other model output: bounded error rate when the adversarial pipeline runs, audit trail visible in the vault, framework structure governing the steps. The mode is one parameter among many; it shapes the reasoning's posture without changing the reasoning's underlying reliability properties. ## The summary Modes are how the system reasons. Frameworks are what it does; modes are the cognitive posture it applies to the doing. The user picks the mode because the user is the one who recognizes which posture the work requires. Explicit picks make the system pedagogical: the user trains by making choices, and the choices transfer to contexts where the user is operating without the system. The territory taxonomy organizes the modes against the structure of cognitive work itself. The dispatch is transparent so the user learns the taxonomy through use. The mode system is the AHI commitment applied at the reasoning-posture layer: the system serves the user's directing intelligence; the user's recognition is what makes the system's output appropriate to the work.