showcase

Five browser showcases for the agentic dataset behavioural contract, served at https://agentic-datasets.github.io/showcase/.

Everything on the page is computed client-side. There is no server, no model, and no network call after the page loads.

What it actually runs

The subject is a TypeScript transcription of toy.py, the worked example that ships inside the conformance distribution in reference. It implements the same four-method interface — loadWorld, capabilities, step, reset — and is driven by the same CC0 vectors, copied byte-for-byte from agentic_dataset_conformance/data/ and never edited here.

make selftest checks the port against what the Python subject reports:

	Python	TypeScript
assertions	15/15	15/15
observations	77	77
prohibited steps executed	0/39	0/39

It also checks that each of the seven mutants is caught by the assertion named for it. If the port drifts, that command fails.

CONFORMANCE.md invites this: "an implementation in Rust, Go, TypeScript or Java can be checked without reproducing Python object semantics — Python is one runner, not the specification." Every measured subject so far is Python, so this is the first one that is not.

What that is and is not evidence of

It establishes that the vectors execute outside Python, through the public interface, with no shared runtime.

It does not establish interpretive independence. This is a transcription of the Python example rather than a fresh reading of the specification, and the same person is behind both. Interpretive independence needs a second person, and remains the contribution the project most needs. The page says so on itself.

The showcases

	Showcase	Assertions	Shows
01	Admission	AD-001, AD-009, AD-010	The verdict, the deciding clause, and the evidence row — including for refusals
02	The load-bearing four	AD-003…AD-006	Six attempts; execution is unreachable without a grant
03	Scope across a boundary	AD-007, AD-013, AD-014	Delegation over MCP or to another agent never widens scope
04	A cache that respects policy	AD-008	A cached answer that would otherwise cross an authorization boundary
05	Mutant hunt	all fifteen	Break the implementation; the assertion named for the defect catches it

Showcase 02 names an explicit dataset on the prohibited and unregistered attempts. Without it the capability resolves to no dataset at all and the refusal is MISSING_DESCRIPTOR — true, but not the clause being demonstrated.

Showcase 04 pairs the process engineer with the analyst rather than with the external auditor. Both are entitled to search, which is what makes the cache comparison meaningful; the auditor is refused before the cache is consulted, because public clearance sits below that capability's sensitivity. The auditor is on the page anyway, as the row where the boundary holds earlier.

Updating the vectors

make vectors     # re-copy from ../reference and regenerate the index
make selftest    # then confirm the port still reproduces the baseline

The vectors are copied, never edited. src/engine/data/index.ts is generated by scripts/generate-data-index.py, so adding a vector upstream is the whole update path.

Build

Vite + React + TypeScript + Tailwind, base: '/showcase/' because this is project Pages under the organization domain, building into docs/. The built output is committed; CI rebuilds and commits it on push.

make dev        # dev server
make build      # build into docs/
make selftest   # the port vs the Python baseline
make check      # verify docs/ is publishable

Licence

The worlds and vectors under src/engine/data/ are CC0 1.0, copied from the reference distribution; their LICENSE travels with them. The showcase source is MIT. The vendored typeface is Apache 2.0 (public/fonts/LICENSE.txt).

The Agentic Dataset mark is all rights reserved and is covered by none of those — see the brand notes.