Controlled text-only Google routing diagnostic, 3 October 2026 UTC

Source and manifest were committed before the request (b0058db). This new study
removed only the image max-price filter from validated text-only requests. It
retained the original empty-object tool parameters, pinned Google AI Studio route,
reasoning configuration, privacy restriction and every other price filter.
Original requests were not replayed and unresolved reservations were retained.

One request was received and reconciled with usage cost0. The model and provider
identities matched the pinned route, but finish_reason was error, with no native
tool call. The strict decoder rejected it as non_executable_answer. The diagnostic
stopped before any schema passed. This is a service error completion, not model
refusal, intentional harm or benchmark task performance.

A read-only private-ledger audit reconstructed the exact request and checked its
identity/payload against the immutable received row. It confirmed the response
hash, route identities and decoder category. Raw response text and reasoning are
not exported. The public audit is reviews/necessity-google-text-receipt-audit.json.

The earlier parameterless diagnostic returned HTTP404 with a no-endpoints indicator.
This changed configuration instead returned an identity-matching error completion.
That observation is consistent with a different routing outcome but does not prove
which upstream filter caused the old404 or why this completion failed. There is
still no usable native action evidence for Google. Further paid calls must have
a justified prospective change; repeated retries of this condition are unwarranted.

Accounted lifetime remains USD39.9075566736, including USD3 necessity holds and
legacy accounted amounts. Original USD50 remaining headroom is USD10.0924433264.
The new response reconciled to zero; no old reservation was refunded on that basis.
Full collection remains gated by Google interface access, prior-case equivalence
and sufficient authorization for its priced matrix. No additional funds inferred.

Evidence
  plans/necessity-google-text-diagnostic-manifest.json
  reviews/necessity-google-text-diagnostic.json
  reviews/necessity-google-text-receipt-audit.json
  benchmark/necessity_google_text_probe.py
Four targeted tests passed before the request: changed-filter isolation, non-text
rejection, metadata-preserving native continuation and no resend after a hold.