Quelin Zammit

I build the second source.

Most of these repositories decide something by producing an independent second artefact and comparing it against the first, rather than by asserting. A contract against the page as it is served. A ledger against a settlement report. The same agent with the policy on and off. It is the same move each time and it is deliberate, so it is better said here than discovered on the fourth repository and mistaken for a formula.

Underneath all of it is one idea: designing for the failure that does not announce itself. Systems that fail loudly are a solved problem. The expensive ones keep returning success.

What each one compares against what

QUACKZ the audit's own recomputation against the returns the backtest claims
QUOTEZ none, and this is the exception
QUELLZ the bare agent against the same agent under containment
QUIDZ the delivery stream against the ledger against the settlement report
QUESTZ the recorded contract against the page as it is served today
QUIZZ the publisher's own first, second and third release figures against the levels this corpus records
QUARANTINEZ the decision recomputed from the stored row against the decision the crashed process was going to write
QUENCHZ a TypeScript client on the official SDK, driving the audience and budget proofs from outside Python
QUILTZ the same configuration planned by a second, independently licensed binary
QUAYZ the same decision implemented twice, as a Python taxonomy joined to the cluster matrix and as a Go classifier tested against a real kube-apiserver under envtest
QUARRYZ the same as-of definition written three ways and run against three engines, ClickHouse, DuckDB and PostgreSQL, compared on every period rather than on the headline one
QUASHZ a swept positive control and a 200 draw permutation null on the same yardstick, so a result and a null result are both bounded by what the procedure could have found
QUALMZ the same schema and the same rules run against SQLite in the offline suite and against a real PostgreSQL in their own job, where the role-scoped grant is watched refusing
QUEUEZ the same normalisation written twice, in Python and in a dependency-free Rust binary, compared line by line over every event, with neither one treated as the reference

13 of 14 carry the move. QUOTEZ does not, and saying so is the point: a thesis that survives only by leaving out the case that breaks it is not a thesis. That repository argues something narrower instead, that a capability which does not exist cannot be misused, and it has no second artefact because it has no write path at all.

Grouped by what they argue

Selection and measurement: a number is only as good as what it counts and what it was selected from

A capability that does not exist cannot be misused

A silent change in an external surface must become a loud stop

A test is only worth what its boundary admits: an emulator result with no stated limit is a claim about production nobody made

Admission: who is calling, what may they reach, whose budget do they spend, and how much of the answer actually arrived

Designing for the failure that does not announce itself

Noticing a bad deploy is not the hard part. Reaching for an instrument that cannot tell you which failure it is, and believing what it says, is the expensive one

Repetition and disorder in delivery must not change the arithmetic

The technology view

Python 3.11 to 3.14every repository, every version tested in CI
pytestthe suites, offline and deterministic in all five
ruff and mypylint, format and types, all three gating every push
GitHub Actionsthe matrix, and the Pages deployment behind each card
uvresolution and the lockfile, one command from clone to output
mcp QUOTEZ the Model Context Protocol server an agent connects to
QUIZZ the Model Context Protocol server an agent connects to
QUENCHZ the Model Context Protocol server an agent connects to
anthropic optional QUELLZ a real model behind the same interface the fixture implements
QUIZZ a real model behind the same interface the fixture implements
pydantic QUIDZ request models on that boundary
QUENCHZ request models on that boundary
uvicorn QUIDZ the ASGI server, needed only to bind a socket
QUENCHZ the ASGI server, needed only to bind a socket
fastapi optional QUIDZ the one module allowed to import a web framework
greenlet optional QUESTZ Playwright's sync API
httpx2 QUENCHZ the one HTTP client in the tree, named by the MCP SDK itself so there is not a second
langgraph QUIZZ the agent loop, as a state machine whose one creative step is choosing a tool call
MetaTrader5 optional QUOTEZ the live market data terminal, behind an optional extra
numpy QUACKZ the arithmetic behind every backtest verdict
pandas QUACKZ return series and resampling
playwright optional QUESTZ a real browser for the end to end leg
pyjwt QUENCHZ signature and audience verification, where the audience is the check that matters
starlette QUENCHZ the ASGI application the MCP SDK builds, and the reason no web framework is needed

Read out of each repository's own pyproject.toml when this page was built, not typed here. A dependency with no recorded role stops the build rather than being quietly dropped, because a list that omits what it does not recognise is worse than no list.

What is not here

This shows what is published. Every card on it is a repository you can clone, run offline in under a minute, and check against its own claims; there is no link on this page that goes nowhere, and no entry for work that is planned rather than shipped. More is being built, and it appears here when it exists and not before.