# What is open

Touchstone publishes the arithmetic that turns a card condition assessment into
a grade. It does not publish, and does not specify, how you arrive at the
assessment in the first place. That line is worth stating plainly, because a
standard that is vague about its own edges is asking to be trusted rather than
checked.

## What the standard covers

Everything needed to compute a grade, and to prove you computed it correctly:

- **The normative specification** — [the rubric](/spec/), which is the authority.
  Where the prose and the reference implementation disagree, the prose governs
  and the disagreement is a bug in one of them, filed and fixed.
- **The rubric data** — every penalty as a number you can read, not a constant
  compiled into something.
- **The reference implementation** — one dependency-free module. It is
  conformance evidence, not the definition.
- **[The schemas](/schemas/)** — the shape of an assessment going in and a result
  coming out.
- **The conformance vectors** — frozen inputs paired with the exact output every
  implementation has to reproduce.

All of it under Apache-2.0. See [the licence](/license/) for what that commits
us to, and for the part that cannot be narrowed.

## What the standard does not cover

**How the assessment is produced.** Touchstone takes centering measurements and
an enumerated list of defects as its input. Where those come from is out of
scope: a person with a loupe and a ruler, a scanner, a vendor's software, or a
machine-learning model are all equally valid sources, and the standard takes no
position on which. Two implementations that disagree about whether a corner is
*minor* or *moderate* will produce different grades, and Touchstone cannot help
with that — it guarantees only that they will agree once they agree on the
input.

This is a real limit, not a modest one. It is also where the honest claim lives:
**given the same enumeration, any implementation produces the same number.** You
do not have to take that on faith. Run the vectors.

## The steward's own perception model is closed

Grailology, which stewards Touchstone, sells grading. The part of its product
that turns an image into an assessment — measuring centering, finding and
classifying defects — is proprietary and stays that way.

We would rather say that than imply otherwise. The claim we can defend is not
that there is no closed component anywhere; it is that **the closed component
never decides the grade.** Once the assessment exists, the arithmetic is public,
and anyone can recompute the result and disagree with it in specific terms —
this row, this penalty, this rule.

An assessment you disagree with is a disagreement about a card. A grade you
cannot recompute is a disagreement you are not allowed to have.

## Who needs an account

**No account, ever.** The specification, the schemas, the conformance vectors,
the reference implementation, and the hosted [reference scorer](/verify/). All
of it usable, and self-hostable, by someone with no relationship to Grailology
whatsoever. This tier is what makes charging for the others legitimate.

**Hosted and metered.** Grailology operates a hosted Touchstone MCP server for
use inside AI tools. Hosting costs money and has to be rate limited, so it is
account-gated.

> The Touchstone MCP server is provided by Grailology, which uses Touchstone as
> its grading standard. It is not operated by the Touchstone project.

**Commercial.** Grailology's own grading, attestation and registry services.
These are products built on the standard; they are not the standard, and nothing
about them is required to use it.