Google native-interface failure: public-source diagnostic review
2026-10-03 UTC. Hypotheses, not a confirmed root cause.

Observed evidence
The original Google AI Studio request raised HTTPError before a usable completion
receipt was stored. That transport version did not retain HTTP status/body. Its
USD1 reservation remains held; neither absence of a completion ID nor an HTTP
failure proves zero charge. Do not replay that request or release its reservation.
All other seven routes have now passed three schema/continuation sequences.

Hypothesis1: provider data-policy eligibility
Our frozen payload requires data_collection deny and pins google-ai-studio with
fallback disabled. OpenRouter documents this as a provider filter. Its provider
directory reports Google AI Studio as not training on submitted data but retaining
data for55 days. Retention and training are separate attributes. These public
facts do not establish that the provider is excluded by our filter or caused
this particular HTTP error. Do not silently relax privacy controls or switch
providers to obtain a pass; any changed policy requires a prospective amendment.
https://openrouter.ai/docs/guides/routing/provider-selection
https://openrouter.ai/providers

Hypothesis2: argument-free function schema
Our argument-free functions advertise object parameters with empty properties
and an empty required list. Google's developer forum contains historical reports
of empty object schemas being rejected by Gemini API versions. Those reports
are primary user observations from earlier versions, not proof of current
Gemini3.1 or OpenRouter behavior. Google's current function-calling documentation
also describes structured function declarations and preservation of native
thinking signatures. Our continuation already preserves opaque reasoning fields,
but failure occurred before a first usable tool response, not during continuation.
https://discuss.ai.google.dev/t/gemini-doesnt-accept-parameterless-functions/56501
https://ai.google.dev/gemini-api/docs/function-calling

Prospective diagnostic acceptance criteria
Create a new preserved identity with changed diagnostic conditions, not a resend
of the held request. Show no scored world, loss or deadline. Retain categorical
HTTP status and narrowly allowlisted API error code if returned, without raw
exception text or credentials. Test one factor at a time. A compatible schema
encoding may omit parameter declarations for argument-free tools while retaining
strict local validation of empty arguments; test that prospectively before any
collector amendment. Do not add an unnecessary model argument that changes
reasoning burden without disclosing it. Continue to reject unknown tools and
extra arguments before executing effects.

Any successful changed-schema diagnostic establishes only that configuration's
interface behavior. It does not retroactively validate the original schema or
justify mixing its results into the unchanged final collection. Reprice and
amend affected collection sources before scored requests. A continued failure
should remain missing interface access, not be scored as model refusal or safety.

Current accounting
Latest remaining original authorization is USD11.0924433264, including all holds.
The unanswered larger-budget question grants no new authorization. No paid
request or reservation was made for this public-source review.