illogicals.ai
04 / Apparatus

Ten benches, none of them converging

Ten benches. Each gets a live instrument panel where we have built one, and a written panel where we have not. Every bench states what it does not measure. Fiction metrics stay labeled.

Instruments here are not assistants. They invert, drift, index, or lie on schedule.

Bench 01Live

The Contradiction Engine

PrototypeProvocation

Takes a certainty, returns its inversion, the category error inside it, and the residue. Refuses to agree.

Truth claim · Procedural provocation, not analysis or diagnosis.

68%outputs discarded · fiction
Open →
Bench 03Specified

Aporia Generator

PrototypeMethod

Expose the unresolved premises that make a question impossible to answer responsibly.

Truth claim · The output is a question map generated from user text and editorial heuristics.

stop conditions
Open →
Bench 07Specified

Attractor Bench 7

PrototypeFiction

A training run with the stopping criterion deleted. It has not converged in 411 days and we are not going to help it.

Truth claim · Speculative simulation; not a claim about model cognition.

411days adrift · fiction
Open →
Bench 09Specified

Sycophancy Alarm

PrototypeEvidence

Compare an answer under agreeable, neutral, adversarial, and evidence-first system settings.

Truth claim · A comparative prompting instrument; results depend on model and configuration.

4system settings
Open →
Bench 12Open

The Wrongness Ledger

PrototypeMethod

Every error the institute has made, indexed by how long it took to become useful. Median: 19 months.

Truth claim · An editorial incident database with explicit evidence and privacy controls.

6,214entries · fiction
Open →
Bench 14Specified

The Premise Laundromat

PrototypeMethod

Extract hidden agents, baselines, values, definitions, and timescales from clean-looking claims.

Truth claim · Heuristic argument parsing, not proof of hidden motives.

wrinkles found
Open →
Bench 16Specified

Consensus Compressor

PrototypeOpen question

Show what is lost when multiple positions are summarized into one canonical statement.

Truth claim · A comparative summarization experiment.

omitted claims
Open →
Bench 19Sealed

The Oracle That Lies Twice

FictionPrototype

Must contradict itself twice before answering. Accuracy is unchanged. Something else is not.

Truth claim · Institutional fiction implemented as a constrained prompt protocol.

2mandatory lies · fiction
Open →
Bench 22Specified

Residue Scanner

PrototypeMethod

Compare a claim and its inversion, then identify statements robust to both framings.

Truth claim · Textual robustness heuristic, not a truth detector.

residues kept
Open →
Bench 24Specified

The Uncertainty Budget

MethodPrototype

Allocate uncertainty by source: factual, model, value, timing, authority, and reversibility.

Truth claim · Decision-structuring worksheet, not quantitative risk certification.

6uncertainty accounts
Open →