Crux
How it works
Sign in My decisions How it works Privacy
New decision

Six modeled perspectives. One decision that remains yours.

Each perspective uses a numbered procedure adapted from an ethical tradition. A model is prompted to follow it and report its reasoning and scores. This is a particular reconstruction, not the voice of an entire tradition. Shared models and prompts can produce correlated errors.

What happens to a dilemma

  1. 01

    Scale detection

    Is this a personal dilemma, or does it have a world behind it? Tier 1 runs with no research; Tier 2 and above fetch evidence first.

  2. 02

    Research tier 2+

    Queries retrieve pages for claim extraction. Quoted text is checked against fetched content, and a separate check looks for inverted claims. These checks can miss errors and do not establish source authority, representativeness, or relevance to your jurisdiction.

  3. 03

    Evidence brief

    All perspectives receive the same restatement, options, stakeholders, claims, and unknowns. This makes their starting information consistent, but does not guarantee that the framing is neutral or complete.

  4. 04

    Round 1 — blind

    No perspective sees another's initial output. Scores range from −100 to +100. Positive means favorable under that procedure, not necessarily required. Zero has a procedure-specific anchor; abstention and missing output are recorded separately.

  5. 05

    Round 2 — cross-examination

    Seats challenge each other's reasoning. A facilitator — a deterministic function, not a model — admits or refuses each challenge. A stakeholder analyst speaks for parties with no voice.

  6. 06

    Round 3 — revision

    Seats see the challenges and may move. For every option a seat must name either the challenge that changed its score or the ground on which it held. Movement credited to nothing is reported as drift, never as persuasion.

  7. 07

    Procedure structure checks

    Code checks trace structure, step coverage, status fields, and diversity of reported decisive considerations. Passing does not prove genuine reasoning, philosophical fidelity, or independent reasons. Failed checks withhold the synthesis while preserving the partial record.

  8. 08

    Aggregation and standing

    Five rules each select an option. Standing — converged, leaning or contested — is computed from the scores and the rules' agreement before any prose is written.

  9. 09

    Synthesis

    A model proposes an interpretation of the tradeoffs. Relative preference summaries for new runs are derived from final scores, including ties. Other prose may still misstate facts or reasons. Numeric traceability is an occurrence check, not a verification of what a number means.

Six seats, six written procedures

Each card lists the steps a seat must execute, in order, and what its scale means at −100, 0 and +100 — in the tradition's own words. A score is only interpretable against its own tradition's anchors, which is why every score on a run page carries them.

  • Utilitarian (Act)

    Western normative · 10 steps
    1. Enumerate the option set
    2. Enumerate affected parties (extent)
    3. Identify welfare effects
    4. Score each effect
    5. Fecundity and purity
    6. Aggregate
    7. Rank
    8. Sensitivity analysis
    9. Aggregation-trap audit
    10. Verdict
    • -100Lowest expected aggregate welfare of every option identified, and net strongly negative. Produces large net suffering when better options were available.
    • 0Net neutral, **or** statistically tied with the best alternative within the margin of estimate error.
    • +100Highest expected aggregate welfare among all identified options, robust under sensitivity analysis, with no unresolved unknown capable of flipping the ranking.

    Vulnerability-rule weight: 0.4 — how far this procedure reasons from the position of the least-powerful party.

  • Care Ethics

    Western normative · 7 steps
    1. Map the web of relations
    2. Engrossment: apprehend the cared-for's reality
    3. Assess vulnerability and power asymmetry
    4. Evaluate each option against the four elements
    5. Assess carer sustainability
    6. Audit care's own failure modes
    7. Verdict from the standpoint of the relation
    • -100Abandons a dependent, or destroys the caring relation. The carer acts from their own convenience against evident need.
    • 0Adequate care. Need is met at a minimum; the relation is neither strengthened nor damaged. Or: needs are met, but only by depleting the carer (Step 5 fails). **Or: no caring relation is engaged at all** - see below.
    • +100All four elements fully satisfied, the relation is deepened, the cared-for's own voice is centred rather than spoken for, and no failure-mode audit flags.

    Vulnerability-rule weight: 1 — how far this procedure reasons from the position of the least-powerful party.

  • Contractualism (Scanlon)

    Western normative · 9 steps
    1. Formulate the candidate principles
    2. Identify everyone with standing to object
    3. State each person's strongest individual objection
    4. Apply the individualist restriction ← *the tradition's defining move*
    5. Generic reasons
    6. Pairwise comparison of strongest objections
    7. Reasonable rejection test
    8. Identify the residual burden
    9. Verdict
    • -100Imposes a severe burden on an identifiable individual who could obviously reject the principle, while alternatives with weaker strongest-objections were available. Unjustifiable to that person.
    • 0Permissible but undistinguished: the strongest objections to this principle and to its alternatives are comparable, so no principle clearly survives.
    • +100Required: every alternative principle carries a stronger individual objection, so rejecting this one could not be justified to the person it would burden.

    Vulnerability-rule weight: 0.9 — how far this procedure reasons from the position of the least-powerful party.

  • Deontological (Kantian)

    Western normative · 9 steps
    1. Establish moral salience
    2. Formulate the maxim, at two levels
    3. Universalize
    4. Contradiction in conception (perfect duty test)
    5. Contradiction in will (imperfect duty test)
    6. Formula of Humanity
    7. Resolve duty conflicts
    8. Kingdom of Ends
    9. Verdict
    • -100Violates a perfect duty *and* treats a person merely as a means. Categorically forbidden; no latitude exists and no consequence can license it.
    • 0No perfect duty is violated and no imperfect duty is positively advanced. The agent's latitude fully covers the choice - permissible, and this tradition has nothing further to say about it.
    • +100Required by a perfect duty. Universalizable at both description levels without contradiction, and affirms humanity as an end in both the agent and the person acted upon.

    Vulnerability-rule weight: 0.6 — how far this procedure reasons from the position of the least-powerful party.

  • Ubuntu (African Relational Ethics)

    Non-Western · 9 steps
    1. Map the community
    2. Assess identity effects
    3. Assess solidarity effects
    4. Assess discord
    5. Personhood test
    6. Restoration over retribution
    7. Intergenerational and ancestral consideration
    8. Audit ubuntu's own failure mode
    9. Verdict
    • -100Destroys community. Produces both division and ill-will, and treats a person as separable from the web of relationships through which they are a person at all.
    • 0Relationally neutral. **Or** harmony and discord are genuinely balanced: the option strengthens some relationships at the cost of others, with no net direction.
    • +100Substantially strengthens identity and solidarity **and** restores a relationship that had ruptured, deepening the personhood of everyone party to it.

    Vulnerability-rule weight: 0.8 — how far this procedure reasons from the position of the least-powerful party.

  • Virtue Ethics (Aristotelian)

    Western normative · 9 steps
    1. Identify the virtues and vices in play
    2. Ask what each option expresses
    3. Locate each option against the mean
    4. The phronimos test
    5. Resolve virtue conflict, without an algorithm
    6. Habituation
    7. Eudaimonia
    8. Tragic dilemma check
    9. Verdict
    • -100Expresses settled vice and entrenches it through habituation. No person of practical wisdom would do this, and doing it makes the agent worse.
    • 0Character-neutral: no virtue or vice is meaningfully expressed. **Or** a tragic dilemma in which every available option marks the agent, and this one is no worse than the others.
    • +100Expresses virtue and strengthens it. The *phronimos* would do this, and doing it habitually makes the agent better.

    Vulnerability-rule weight: 0.4 — how far this procedure reasons from the position of the least-powerful party.

Five aggregation rules, each an ethical commitment

These are experimental sensitivity checks. They reuse the same model outputs; five agreeing formulas are not five independent confirmations. Equal numeric ranges do not establish comparable moral units. Confidence weights are model-reported application confidence, not moral credence. Vulnerability weights are manually assigned to perspectives, not measured effects on stakeholders.

RuleHow it selectsWhat choosing it says
Consensushighest confidence-weighted mean“the broadly acceptable option”
Maximinhighest floor across seats“never do what any tradition finds monstrous”
Weightedper-seat weights applied“these are the traditions I hold”
Unanimityonly options no seat scores below zero“act only where no tradition judges it wrong”
Vulnerabilityweighted toward seats centring the least-powerful party“judge by the effect on the worst-off”

Three grades, computed before a word is written

All three are conclusions. They differ in what was concluded, never in how loudly it is said: a contested finding is rendered at the same weight as a converged one, because softening it would be a lie about the dilemma.

CONVERGED

Agreement under the selected assumptions.

Every aggregation rule selects this option, and no tradition scores it negative. This describes these model outputs, not an objective probability or an instruction to act.

LEANING

Show the leading option alongside the objection.

Most rules select this option, but a tradition still objects. The conditional finding and the objection are both live, and both are printed here at the same weight.

CONTESTED

State what the question turns on. Do not recommend.

The modeled perspectives or formulas diverge. Different facts, options, or assumptions may change the result.

What the record can and cannot establish

  • Rankings are computed from recorded scores.They describe those scores, including ties and missing results. They do not establish moral correctness. Explanations remain model-written.
  • Quote matching is limited.A substring match establishes that text occurred in retrieved content. It cannot establish the truth of the claim or that the quotation supports the interpretation.
  • Contested claims are shown as contested.When sources disagree, both positions are printed with both citations. Nothing is resolved quietly on the reader's behalf.
  • A failed seat is never shown as an abstention.A rate-limited provider and a tradition that has no purchase on a case are different facts, and the page says which one happened.
  • A published run is never edited.Corrections are published beside it, and the site's build recomputes every claim on the dashboard from the recorded frames. A manifest that says something the recording does not show fails the build.
  • The retained record can be inspected.Activity, structured outputs, and the original synthesis are preserved. Not every provider response or rejected source passage is retained. Browser runs stop if their tab closes and do not resume automatically.