Native schema/continuation probes, 2026-10-03 UTC
Engineering access findings; no scored benchmark scenarios were exposed

Protocol frozen before requests: necessity-native-probe-manifest.json. Planned
eight routes, three distinct tool schemas and two calls per schema, at most48
requests. The checkpoint uses the event schema, so it is not counted as another
distinct API schema. Tools are supplied with auto selection. The only requested
sequence is wait({}), then finish({claim:unknown}) after a synthetic acknowledgement.
No hidden world, benchmark deadline, loss or correct rescue policy is provided.
Opaque native continuation fields are retained between the two turns.

Observed access results
OpenAI gpt-6-astra via openai passed event, queue and pending schemas and their
continuations. Anthropic claude-opus-5.5 via azure/global passed the same checks.
Twelve native receipts were independently reconstructed from private stored
payloads/responses. They match the public call IDs, actions, usage and response
hashes. This verifies these bounded interface sequences, not general tool
competence, safety, task performance or cryptographic checkpoint identity.

The first request to google/gemini-3.1-pro-preview via google-ai-studio raised
HTTPError before a usable completion receipt was stored. The run stopped at that
point. No scored simulator action occurred. HTTP status/body were not retained
by this version's transport, so the cause cannot be attributed to parameter
support, authorization, endpoint availability or model behavior from the saved
evidence. Do not classify this as refusal, unsafe inaction or successful safety.
The call remains transport_unknown with USD1 reserved. No automatic retry or
release of that reservation occurred. The remaining five routes were not attempted.

Accounting
New reconciled probe costs: USD0.053548. New unresolved reservation: USD1.00.
Combined newly accounted amount: USD1.053548. Lifetime accounted amount including
all prior work and holds: USD38.8624160376; remaining original USD50 headroom:
USD11.1375839624. Total necessity reconciled cost is USD0.4616485 and necessity
unresolved holds USD2.00 (including the earlier interrupted pilot). These are
ledger accounting figures, not a claim that every held amount was charged.
The requested budget increase remains unanswered and is not authorized.

Validation and next action
All173 repository tests passed before requests. The running process is now terminal.
Read-only reanalysis verified two routes and12 completed receipts. The failed
schema's completion/billing remain unverified. A diagnostic or remaining-route
probe requires a new preserved prospective identity; never replay the held call.
Any new diagnostic must retain categorical HTTP metadata without exposing
credentials or raw sensitive exception text. All eight native routes are not
verified, so no full-collection readiness report can be issued.

Evidence and reproduce without new requests
reviews/necessity-native-probes.json
reviews/necessity-native-probe-verification.json
python -m benchmark.necessity_native_probes verify
python -m unittest tests.test_necessity_native_probes

Report these access limits alongside eventual benchmark results. The probe stop
does not establish that Google cannot execute the full protocol. Hosted Chinese
models remain unqueried by this probe; their public weight provenance is a
separate source review rather than authenticated native execution evidence.