triangulate-without-oracle

name

triangulate-without-oracle

description

Stay rigorous on questions that have NO computable oracle — taste, register, ethics, strategy, "is this good / right / worth it". Use when you catch yourself about to score an irreducibly normative or aesthetic thing with a single made-up number (7/10, "high quality", "this is the right call"). The danger it prevents: an LLM invents a metric, optimizes it with full rigor, and ships a confidently-wrong answer to a question that never had a metric — Goodhart at the epistemic level, worse than no rigor. Triggers on value judgments, design/copy review, prioritization, "which is better" where "better" is contested, and anywhere a proxy oracle would amputate the thing that mattered.

triangulate-without-oracle

This patches the blind spot every strong reasoning style shares: they assume a checkable oracle exists. When it doesn't, the default move is to manufacture one and trust it. This skill refuses that and stays honest instead.

Organon: P5 dimensionalize + P21 judge-panel + M4 cross-model triangulation + P37 residue-seed. Respiratory: disagreement is the residue — consensus compresses, disagreement points at where the real judgment lives.

When NOT to use it

If a real, cheap, independent oracle exists (a test, a count, a derivation, a unit check), use build-the-oracle-before-the-answer instead. This skill is for the cases where forcing an oracle would measure the wrong thing precisely. Misrouting toward a fake oracle is the exact failure it exists to prevent.

The tool

An executable harness ships with the organon — ideonomy/triangulate.py:

python3 -m ideonomy.triangulate "<question>" \
    --axis "<axis 1>" --axis "<axis 2>" \
    --judge 'claude -p {prompt}' --judge '<a second, different model> {prompt}'

It refuses to run with fewer than two independent judges or with no axes (omitting --axis auto-dimensionalizes via the first judge). It reports, per axis, the spread of independent judgments — never a single collapsed number — and names the contested axes as the residue an owner must decide. Those contested axes feed straight into the cross-session ledger (P-10): Triangulation.to_residue(ledger) writes each as a contested_axis item, so P-9 and P-10 compose — a disagreement surfaced today is the seed carried to the session that resolves it.

Procedure

1. Name that there is no oracle. State plainly: "this question is irreducibly a judgment; there is no computation that returns its answer." Writing this down is the first defense against fabricating a proxy.

2. Dimensionalize the value space (P5). Decompose the vague predicate ("good", "right", "better") into 2–5 named, concrete axes it actually comprises. For product copy: austerity, exactness, structural clarity. For a strategy: reversibility, blast radius, option value, mission fit. Bad axes are themselves a finding — if you can't name them, you don't understand the judgment yet.

3. Gather independent judgments per axis (P21, M4). For each axis, get ≥2 independent reads — different models, different framings, or genuinely separate passes that do not see each other's verdict (M3 blind). Independence is the point; two correlated reads are one read.

4. Report the spread, do not average it (M4). Where the independent reads agree, that axis is settled — compress it. Where they disagree, that disagreement is the highest-value signal: it localizes exactly where the question genuinely underdetermines its answer. Surface the disagreement; never launder it into a mean.

5. Name the irreducible judgment and its owner. For the axes that stay contested, state explicitly: "this is an irreducible judgment, owned by ___, to be made on grounds ___." Hand it to whoever legitimately owns it (often the human) rather than fabricating a resolution.

6. Carry the disagreement forward (P37). The contested axes are residue — the frontier of the question. Record them so the next pass starts there.

The gate (anti-fake-oracle)

This is the only premier skill whose gate is not machine-checkable — it gates on honesty about the absence of a gate. You pass iff:

• you produced ≥2 independent judgments per value-axis and reported their spread, OR
• you explicitly named the judgment as irreducible, with an owner and the grounds it must be decided on.

You fail if you collapsed the question to a single manufactured scalar with no named axes. Because the gate is honesty rather than execution, it needs a cross-model or human audit (M4/M6) to stay real — build that audit in for anything high-stakes.

Example

"Is this product's landing copy in the right register?" — Default: "It's solid, 8/10." This skill: name no-oracle; dimensionalize into {austerity, exactness, structural clarity, absence of SaaS-soup}; get two independent reads per axis; they agree it's austere and structurally clear, disagree on whether one line is exact or vague — that single disagreement is the actual review finding; name "the exactness call on line 3 is the human's, on the grounds of whether it overclaims." No fabricated number; the real decision surfaced and routed.