Decision-Making Under Uncertainty

Overview

The Decision-Making Under Uncertainty Framework (T3) is the territory framework for the case where the user faces a decision and wants structured guidance. The territory has the unusual property that all three of the system’s analytical axes — depth, complexity, and stance — are active at once. Depth runs from the lighter constraint-mapping pass to the molecular decision-architecture composition. Complexity runs across the four within-territory modes (constraint-mapping, decision-under-uncertainty, multi-criteria-decision, and the deferred ethical-tradeoff). Stance lives in the routing question itself — is the environment basically known and you’re picking from clear options, are there real unknowns about how things will play out, are you weighing several criteria that don’t reduce to one number, or are values in tension where the choice is partly about what you stand for? The framework’s first job is to route to the operation that fits.

The framework runs four modes plus the deferred ethical-tradeoff. Constraint-mapping is the lighter Tier-2 mode for known environments — at least three alternatives mapped with success conditions, failure conditions, what is uniquely gained, and what is forfeited per alternative, plus surfaced no-lose elements (actions valuable regardless of which alternative is chosen). The mode produces the choice terrain without making the choice for the user, unless explicitly asked. Decision-under-uncertainty is the central Tier-2 mode when probabilities and time-value are central — classifies each critical variable as risk (assignable probability), uncertainty (estimable range), or deep uncertainty (no meaningful probability); considers defer/sequence/hedge/buy-information alternatives alongside direct choices; presents non-quantifiable factors (ethics, relationships, identity, reputation) alongside the quantitative framework; names what would change the recommendation. Multi-criteria-decision is the descriptive Tier-2 mode when ≥3 criteria matter and explicit weights are needed — names the MCDM method (additive SMART, AHP pairwise, ELECTRE outranking, TOPSIS distance-from-ideal), surfaces weights with elicited rationale, identifies dominance relations to prune the option set, runs sensitivity analysis on weights and scores, flags where the ranking is robust vs. fragile. Decision-architecture is the molecular composition for high-stakes decisions where the user wants integrated architecture spanning constraints + uncertainty + stakeholders + failure modes (the molecular sibling has its own framework spec, Framework — Decision Architecture Analysis.md).

The framework’s load-bearing intellectual content is the stance routing across modes, the risk-uncertainty-deep-uncertainty classification discipline, the defer/sequence/hedge/buy-information alternative requirement, and the non-quantifiable factors discipline. The stance routing operates at Q1 — the user describes the situation in plain language and the framework dispatches to the mode shaped to handle it. The risk-uncertainty-deep-uncertainty classification, drawn from Knight (1921), is the load-bearing discipline against false precision: variables that are genuinely deep-uncertainty (no base rates available) cannot be assigned point probabilities, and the framework refuses to pretend they can. Variables that are uncertainty-with-estimable-range get qualitative bands or ranges rather than spurious point estimates anchored to initial guesses.

The defer/sequence/hedge/buy-information alternative requirement counters the most common framing failure — binary alternatives where the user has missed creative third options. Defer (wait for more information when VOI exceeds cost of delay), sequence (stage the decision so each step informs the next), hedge (cushion the downside of a high-variance choice), and buy-information (commission research, run a pilot, conduct a probe) are first-class alternatives to be considered alongside direct choices. The non-quantifiable factors discipline counters the quantification trap — ethics, relationships, identity, and reputation are presented alongside the quantitative framework, not as footnotes after the EV calculation, because doing so changes which alternative leads in many real decisions.

The framework’s escalation hooks ensure the user can move between modes as the analysis surfaces material the originally-routed mode is not shaped to handle. Constraint-mapping hooks upward to decision-under-uncertainty when real unknowns surface that change the option set; decision-under-uncertainty hooks sideways to multi-criteria-decision when multiple non-commensurable criteria are in play; any Tier-2 T3 mode hooks upward to decision-architecture when the user wants a single integrated artifact tracking all dimensions; any T3 mode hooks sideways to T7 when the question shifts from “what should I do” to “how could this fail.”

The framework answers questions like: Multiple viable options exist; which should I choose? A choice must be made between alternatives with uncertain outcomes; probabilities matter but are not precisely known. I’m choosing among options and they trade off across multiple things I care about; no single criterion can settle this. Should we act now or wait? Is it worth waiting for more information?

Systemic context

Decision-Making Under Uncertainty is the territory framework for super-cluster C (Decision, Future, and Risk) in Ora’s reasoning catalog. It distinguishes itself from adjacent territories at the routing layer — Future Exploration (T6) handles questions about how the future might unfold irrespective of what you do (when scenario exploration must precede choice, T6 runs first and T3 then operates on T6’s scenarios); Risk and Failure (T7) handles failure-stress-testing as the focus rather than choice (when the question shifts from “what should I do” to “how could this fail,” T7 takes over); Stakeholder Conflict (T8/T13) handles situations where the parties’ conflict is the analytical object rather than choice with parties as inputs; Decision Clarity Analysis (T2’s decision-clarity) handles document-production for a third-party decision-maker rather than choice for the user themselves. The framework’s four modes consume specific lenses — constraint-mapping draws on rumelt-strategy-kernel and strategic-2x2-matrix-tradition; decision-under-uncertainty draws on expected-utility-theory, real-options-methodology, minimax-regret-and-robust-decision-making, and tetlock-superforecasting; multi-criteria-decision draws on mcdm-methods (the catalog of weighted-sum, AHP, ELECTRE, TOPSIS, PROMETHEE methods); the Knightian risk/uncertainty/ambiguity foundational lens runs across all modes alongside the Kahneman-Tversky bias catalog. The molecular sibling (decision-architecture) has its own framework spec; output from any T3 mode flows back into the upstream framework that invoked it, or stands alone when the user invoked T3 directly.

Ora’s output

Example question: I have a job offer from a competitor with significantly better compensation but I’d be moving to a smaller company with less stability. My current company has just announced a leadership change that might affect my role. There’s also an internal opportunity I might be considered for if I asked. How do I think about this?

Ora’s response (running T3 in decision-under-uncertainty mode after Q1 routing):

The user’s situation has real unknowns (will the leadership change affect their role? would the internal opportunity be granted?), the alternatives include direct choices (accept offer / stay) and at least one defer-shape (ask about internal opportunity), and probabilities matter but are not precisely known. Q1 routes to decision-under-uncertainty rather than constraint-mapping (the environment is not basically known) or multi-criteria-decision (criteria are present but the load-bearing dimension is uncertainty rather than weight-elicitation across many criteria).

Decision framing. The choice is among (A1) accept competitor offer; (A2) stay at current company without action; (A3) ask about internal opportunity and decide based on response; (A4) negotiate competitor offer for delayed start to see how leadership change unfolds. A3 and A4 are surfaced via the defer-prompt — the user’s framing was binary (accept or don’t), and the framework surfaces the staged alternatives.

Uncertainty identification.

Variable 1. Will the leadership change affect the user’s role? Classification: uncertainty (estimable range — depends on the change’s pattern, the user’s organizational position, observable signals like reorg announcements or budget reallocation in the next 60 days). Estimate: 30-60% likelihood of meaningful role impact, declining over time.

Variable 2. Would the internal opportunity be granted if asked? Classification: uncertainty (range estimable — depends on the user’s standing, manager relationship, organizational politics around the opportunity). Estimate: 40-70%.

Variable 3. How will the competitor company perform over the next 18 months? Classification: deep uncertainty (no meaningful probability — depends on market conditions, execution, competitive dynamics that don’t have base rates the user can apply).

Variable 4. What’s the user’s two-year compensation under each alternative? Classification: risk under A1 (compensation is contractual); uncertainty under A2 (depends on leadership change × user’s role × bonus); uncertainty under A3 (depends on opportunity grant + new compensation structure).

Consequence analysis (probability-weighted).

A1 (accept competitor offer). Best case: better compensation realized + opportunity at smaller company materializes. Modal: better compensation realized, smaller company stability holds. Worst: smaller company struggles, user navigates harder downside. Expected value qualitatively higher than A2 on compensation; lower on stability.

A2 (stay without action). Best case: leadership change is positive for user; current trajectory continues. Modal: leadership change is neutral; current trajectory continues. Worst: leadership change reduces role scope or compensation. Expected value depends sharply on Variable 1.

A3 (ask about internal opportunity). Best case: opportunity granted; user gets compensation bump and new scope without leaving. Modal: opportunity not granted but user has clearer information and can decide A1 or A2 with better data. Worst: asking signals departure-readiness and damages relationship with current manager.

A4 (negotiate competitor offer for delayed start). Best case: competitor accepts; user observes leadership change before committing; can renegotiate or withdraw if leadership change makes A2 attractive. Modal: competitor declines delay; user must decide on original timeline. Worst: competitor withdraws offer over delay request.

Value-of-information analysis. Variable 1 (leadership-change impact) is the highest-leverage uncertainty. VOI of waiting 60 days to observe leadership-change signals: high (would meaningfully update Variable 1 from 30-60% to a tighter band). Cost of delay: probability competitor offer remains open in 60 days: depends on competitor urgency (estimate 40-70%); reputation cost of delay request: low if framed as standard; opportunity cost of A3 if pursued in parallel: moderate (asking creates the signaling cost).

Recommendation. A4 (negotiate competitor offer for delayed start) leads on integrated criteria — converts the dominant uncertainty (Variable 1) into a choice that can be made with better information, while preserving A1 as the fallback if delay is granted and leadership change goes badly, and preserving A2 as the fallback if delay is granted and leadership change goes well. A3 is a parallel option but carries the signaling cost; pursuing both A4 and A3 in parallel is dominated by A4 alone unless the internal opportunity is high-conviction. Conditions to revisit: if competitor declines delay, decision collapses to A1 vs. A2 on original timeline, and Variable 1 must be estimated without the additional 60 days of signal; if leadership change signals fire early (within 30 days), the decision can be made before A4’s deadline.

Non-quantifiable factors. Identity tied to current company; relationships with current colleagues; spouse’s view of the move (if any geographic change); whether the user values being asked vs. asking (some users find A3 corrosive to their sense of self-direction). These are presented alongside the quantitative framework rather than as footnotes — they may flip the recommendation if any are load-bearing for the user.

That is what T3 produces in decision-under-uncertainty mode on a real choice with real unknowns. The framework refused to pretend Variable 3 has assignable probability (deep uncertainty); it surfaced the staged alternatives the binary framing had masked; it identified the dominant uncertainty and the value of converting it into an information-gathering action; it presented the non-quantifiable factors as first-class inputs. The user can act on the recommendation while watching the conditions that would change it.

Commercial AI comparison

Comparison content auto-populates when the comparison-refresh framework runs against this question. Drafters do not author this section.

Brief comparison commentary

Auto-populates with the comparison content above.

How to use this framework

You can run the Decision-Making Under Uncertainty pattern with any AI of your choice. The composition is single-pass for whichever mode the routing question selects.

The prompt:

[Paste the framework specification]

The decision I’m facing: [describe the choice in plain language.]

Alternatives I’ve named (optional): [if you have a list, provide it.]

Stance (optional): [if you know whether the environment is basically known, has real unknowns, involves multiple criteria, or has values in tension, say so — that routes Q1 directly. Otherwise the framework will run Q1 to figure out the routing.]

The AI runs Q1 first to confirm the routing, then proceeds through the mode-specific layers. Constraint-mapping produces five required sections (decision context and constraints; alternatives ≥3; per-alternative analysis; cross-alternative comparison; no-lose elements). Decision-under-uncertainty produces six required sections (decision framing; uncertainty identification; consequence analysis; value-of-information analysis; recommendation; non-quantifiable factors). Multi-criteria-decision produces eight required sections (options inventory; criteria definitions; weights with rationale; scoring matrix; aggregated ranking with method-name; sensitivity analysis; dominant and dominated options; confidence per finding).

For best results:

1. Don’t suppress the staged alternatives. When the framework surfaces defer/sequence/hedge/buy-information alternatives, take them seriously. Most decisions that get framed as binary actually admit a third option that converts an uncertainty into a sequence.
2. Push back on point probabilities. If the framework attaches a single number (“65% chance the leadership change affects you”) to a variable that is genuinely uncertainty-with-estimable-range, ask for a band instead. False precision is the central failure mode of decision-analysis output.
3. Ask for non-quantifiable factors early. If the framework’s first pass produces an EV-shaped recommendation without naming relationships, identity, ethics, or reputation, ask explicitly what non-quantifiable factors are at play here, and how do they bear on the recommendation? These factors are first-class inputs, not footnotes.
4. Escalate up to decision-architecture if the stakes warrant it. If the decision involves multiple stakeholders with substantial impact, binding constraints, and failure pathways worth stress-testing, the molecular pass (decision-architecture) is the right tool. The Tier-2 modes are for decisions where the molecular pass would be over-engineering.

The framework is deliberately tool-agnostic. The stance routing, the risk-uncertainty-deep-uncertainty classification, the defer/sequence/hedge/buy-information alternative requirement, and the non-quantifiable factors discipline are conceptual disciplines that survive the lift to any environment.

Other examples

• Constraint-mapping on a known environment. A user is choosing among three vendors for a software contract. The environment is basically known (vendor track records are documented; contract terms are standard; the user’s requirements are stable). Q1 routes to constraint-mapping. The framework produces ≥3 alternatives mapped with success conditions (deployment within deadline; cost within budget; integration with existing tooling), failure conditions (deployment delays; cost overruns; integration friction), what is uniquely gained per alternative (vendor A’s specialization; vendor B’s existing relationship; vendor C’s cost), and what is forfeited (vendor A’s higher cost; vendor B’s narrower feature set; vendor C’s untested track record). No-lose element surfaced: the requirements documentation work is valuable regardless of vendor selection. Demonstrates constraint-mapping on a decision where probability arithmetic is not central and the value is in the symmetric per-alternative analysis.

• Multi-criteria-decision when criteria don’t reduce to one number. A user is choosing where to live and the criteria include cost of living, school quality for kids, commute time, climate, proximity to family, and career opportunity. Q1 routes to multi-criteria-decision after the user confirms ≥3 criteria with non-commensurable scales. The framework names the method (SMART or AHP, depending on user’s preference for elicitation cost), elicits weights with rationale (the user articulates why school quality outweighs commute time and by how much), produces the scoring matrix (each location scored on each criterion), runs sensitivity analysis (what happens to the ranking if school quality weight drops 20%? if commute weight rises?), and flags dominance (one location is dominated by another on every criterion and can be pruned; no location dominates all others on every criterion, so the ranking depends on weights). Demonstrates MCDM where the analytical work is in the weight elicitation and sensitivity rather than probability arithmetic.

• Escalation from Tier-2 to molecular decision-architecture. A user starts with constraint-mapping on a build-or-buy decision; the analysis surfaces real unknowns about vendor stability that change the option set, escalating upward to decision-under-uncertainty; the analysis then surfaces material stakeholder impact across engineering, product, and finance teams plus binding constraints across budget and deadline plus pre-mortem-worthy failure pathways, and the user requests the integrated architecture; the framework escalates upward to decision-architecture (molecular). The Tier-2 modes’ outputs become input substrate for the molecular composition. Demonstrates the escalation pattern that lets users start light and move heavier as the analysis reveals what’s needed.

Citations

The Decision-Making Under Uncertainty Framework draws on the foundational decision-theory literature. Expected-utility theory traces to von Neumann and Morgenstern’s Theory of Games and Economic Behavior (1944), Savage’s The Foundations of Statistics (1954), Raiffa’s Decision Analysis (1968), and Howard’s “Decision Analysis: Applied Decision Theory” (1966). Knight’s Risk, Uncertainty, and Profit (1921) supplies the foundational distinction between risk (assignable probability), uncertainty (estimable range), and deep uncertainty (no meaningful probability) — the load-bearing classification discipline against false precision. Real-options methodology draws on Trigeorgis’s Real Options (1996) and Dixit and Pindyck’s Investment under Uncertainty (1994); robust and deep-uncertainty methods draw on Lempert, Popper, and Bankes’s Shaping the Next One Hundred Years (RAND, 2003) and Ben-Haim’s Info-Gap Decision Theory (2006).

The multi-criteria methods draw on Saaty’s The Analytic Hierarchy Process (1980) for AHP, Edwards’s “How to use multi-attribute utility measurement for social decision making” (1977) for SMART, Roy’s Multicriteria Methodology for Decision Aiding (1968/1996) for ELECTRE, Hwang and Yoon’s Multiple Attribute Decision Making (1981) for TOPSIS, and Belton and Stewart’s Multiple Criteria Decision Analysis: An Integrated Approach (2002) for the broader MCDM tradition. The constraint-mapping mode draws on Rumelt’s Good Strategy / Bad Strategy (2011) for the strategy-kernel test and the Heath-and-Heath Decisive tradition for the no-lose framing.

The cognitive-bias substrate draws on Kahneman and Tversky’s “Prospect Theory” (Econometrica 1979) and Kahneman’s Thinking, Fast and Slow (2011) — the catalog of biases that decision-aid output is vulnerable to (false-precision, analysis-paralysis, anchoring-trap, missing-defer, quantification-trap) is operationalized as the framework’s named-failure-modes list. The pre-mortem methodology, used at the molecular composition (decision-architecture), draws on Klein’s “Performing a Project Premortem” (Harvard Business Review 2007).

The framework is single-author and originated 2026-05-01 as the territory framework for super-cluster C. The within-territory disambiguation tree (Q1 routing across constraint-mapping, decision-under-uncertainty, multi-criteria-decision, ethical-tradeoff) operationalizes the stance-axis selection; the escalation hooks across modes ensure that work begun in one mode can pivot when the analysis surfaces material the other modes are shaped to handle.

Downloads

• Framework specification (PDF) — link to ora-ai.org canonical artifact when published
• Framework specification (plain text) — link to ora-ai.org canonical artifact when published
• Full white paper (PDF) — link when published