WHAT THIS RUNS ON
Eleven sponsor surfaces, and the line of code that calls each one.
Five Google Cloud services and all six of Parallel's product surfaces. Nothing here is a logo on a slide: each entry names the module and function, says what job it does, and reports whether that path is reachable right now.
A service marked required is one the workflow cannot run without. Remove Gemini and nothing can read the fragment. Remove Parallel Search and there is no open web at all, and this product would rather stop than answer a historical question from model memory.
Live status by service
THE DIVISION OF LABOUR
Models research. Code verifies. A human decides.
Three Parallel surfaces are exposed to the agents as tools, because choosing when to search, when to research holdings and when to enumerate catalogues is a judgement call a reasoning agent should make. Search, Task and FindAll sit in app/partners/parallel_research.py.
The other three are called by deterministic code, never by a model, because each one exists to check or outlast the model's own work. A system that both gathers the evidence and rules on whether the evidence holds up can talk itself into either answer. Extract, Task Group and Monitor sit in app/partners/parallel_verify.py.
- Extract re-opens each cited page. A quotation that is in a search snippet but not in the live document is refused by the gate: it can only ever weaken a verdict, never support one.
- Task Group gives every surviving candidate its own researcher that has never seen the others, and asks it only to disprove that candidate. The Skeptic agent, working in one context window, inherits its own earlier conclusions; these runs cannot.
- Monitor handles the case the product is most likely to hit: no identification possible today. Archives digitise continuously, so an abstention becomes a standing weekly query on the strings visible in the frame.
TESTED AGAINST THE STANDARDS ARCHIVISTS USE
No archivist has reviewed this. So we tested it against their rulebook instead.
The strongest evidence for a tool like this is a practitioner using it and saying what is wrong. We do not have that, and this is not a substitute for it. What we could do is check conformance against the published standards archivists actually work to: the FIAF Moving Image Cataloguing Manual and EN 15907, with every requirement quoted, page-cited, and backed by a test that passes or fails against real output.
It did what an external standard is supposed to do: it found something we had got wrong. FIAF A.2.5 requires that an unidentifiable fragment still receive a supplied/devised title, so an archivist can file it and find it again. An abstention used to return no title at all. It now returns one, built by code from observed clues and marked as cataloguer-supplied.
Full quotations, page numbers and reasoning: STANDARDS-CONFORMANCE.md · machine-readable at /v1/standards · enforced by tests/test_standards_conformance.py.
TRACK REQUIREMENTS
What the Parallel track asks for, and where to check it.
The requirement is that the project actively uses Parallel's Search API at runtime, via the official SDK, and that the integration is present in the code rather than only in a README. The official parallel-web Python SDK is pinned in requirements.resolved.txt; search_archival_evidence is a registered tool on two of the five agents; and GET /health/integrations above performs a genuine Search call against the live credential rather than reporting a cached opinion.
The Google Cloud requirement is met by google-adk and google-genai, both imported and called on the request path. The workflow in app/adk_app.py is executed by a real ADK Runner in app/adk_runtime.py, against Gemini on Vertex AI, hosted on Cloud Run.
Every AI call in this system is a Google Cloud model or a Parallel built-in feature. There is no other model provider anywhere in the dependency lock.