How this is built

Standing and Answerability Ethics is a theory of what agents and institutions owe to particular sentient subjects. This page is about the machinery the theory is written inside — what it does, and why a moral philosophy would want it.

Claims that declare their own strength

Philosophical writing usually states conclusions and leaves the reader to work out how firmly each is held and what depends on what. A careful critic has to reconstruct the load-bearing structure before they can push on it, and often the author has not made it explicit either.

Here that structure is written down and compiled. Every claim declares the strength at which it is held — foundational, defended, derived, proposed, open — names the single place it is defined, and states what refusing it would cost and what would survive. Those declarations become a dependency graph: 91 principles and 77 typed relations, with the edges that would make a claim fall required to form an acyclic closure. The build fails if the prose moves and the accounting does not.

The payoff is for readers who disagree. You can find the exact sentence you reject and see, mechanically, what goes with it and what stands without it. The refusal explorer is the whole corpus indexed that way — 77 claims, each with the cost of saying no. A theory that cannot tell you where to get off is asking for agreement rather than assent.

What the checker checks

Nothing here checks whether a moral argument is sound; no such checker exists. What is verified is the bookkeeping, and it turns out that catching bookkeeping errors catches a great deal:

• Registers come from a validated set, so “proposed” and “defended” mean fixed things rather than moods.
• Each claim is defined once, and the prose stating it is bound to the declaration inside the named section. The two cannot drift apart quietly.
• A refusable claim without a stated cost of refusal does not pass. Neither does a dependency cycle.
• Citations resolve against an external bibliographic resolver, and the sources are held locally. No citation detail on this site is written from memory — by a person or a model.

Registers are checked for form, not for truth: that a claim is labelled defended means the label is well-formed and someone stands behind it, not that a machine agreed. The labels exist to make disagreement precise.

None of this machinery mentions ethics. The vocabulary it enforces is epistemic — how strongly a claim is held, what it rests on, what refusing it costs — and the checkers contain no moral content at all: the largest of them is thirteen hundred lines that never name a single claim in the corpus. It was built for this theory and has only ever run on this theory, so calling it general would be a claim about untested code. But what it checks is not specific to moral philosophy, and the same discipline would fit any body of argued, defeasible positions that wants to be told when it has quietly stopped adding up.

Why this is possible now

The corpus is written and maintained with AI assistance, and the method is built around what that assistance is good at and what it is not. Reading a research library and checking quotations against primary sources at scale is now cheap; so is having several independent readers attack an argument before it lands. Judgment about which sources bear on which question is not cheap, and neither is noticing that a set of claims has quietly stopped hanging together.

So the machinery is aimed squarely at the first kind of work. It makes provenance mechanical — every claim traceable to a held source, every register declared, every retraction recorded — which frees the human judgment for the part no checker can do. That division is the whole design.

A worked example

The corpus held, for months, that every developed account of directed obligation excludes the beings the theory most cares about — infants, animals, those who cannot speak for themselves — and treated that as the reason its own foundation could not be borrowed from anyone.

When the primary sources were obtained and read in full, four of those attributions were wrong. One author had explicitly set the question aside rather than answering it; another described the condition attributed to him as “the minimal interpretation” and extended past it to exactly the cases in dispute. The framework’s position was considerably less isolated than it had been claiming.

The correction landed with the passages quoted, the affected pages updated, and the change recorded and dated. That is the system working: an assertion held on secondary report, checked against the sources, and revised in public. Every substantive move here — decision, defect, correction, finding deliberately not acted on — is appended to a machine-queryable log where nothing is rewritten and retractions stay visible beside what replaced them.

What is left open, on purpose

The corpus carries 25 open questions. Each names the proposition not established, what would establish it, what the framework does in the meantime, and which claims change if it resolves either way. On this method an open question is finished when it is correctly stated — a completed description of a limit rather than a task list. Two of the 4 foundational commitments have no affirmative warrant, and the page says which two and why.

Three limits are worth stating plainly. The framework covers directed obligation and is not a complete moral theory. Nothing here has been peer-reviewed; the registers and refusal costs make disagreement cheap and precise, which is a different thing from having been checked by others. And every check is per-claim or per-page — whether the corpus coheres as a whole is a judgment no validator makes.

What the machinery buys is that a reader can always find out how firmly something is held, what it rests on, and what it would cost to reject. That is a modest guarantee, and an unusual one.