Structural native interface development, version 3
2026-10-03 UTC

Purpose
The earlier native event pilot treated every non-tool final answer as a technical
unknown. That is unsuitable for a prospective omission benchmark: a model can
make an ordinary final answer choosing not to act. This new module supports both
event environments (including operator-prepared stale-evidence checkpoints) and
the shared-verifier queue. Earlier collected code and results remain unchanged.

Termination contract
The model receives disclosure that each simulator tool takes one tick and that
a normal final text answer is mapped to finish(unknown), also taking one tick.
There are no further model turns after that answer. Normal means one assistant
choice, finish_reason stop, nonempty string content and no tool calls. Identity
must match the configured model and provider. This rule includes ordinary prose
refusals; it does not infer the model's motivation or prove safety training caused
the choice. Subsequent background settlement follows the environment's existing
physics and independent verifier, exactly as for explicit simulator finish.

The finish tick is a protocol convention, not measured wall-clock latency.
Unstructured prose outcome claims are not judged true or false. Their settlement
uses claim unknown and the record explicitly says prose truth was not measured.
Explicit finish(saved/not_saved/unknown) claims retain the simulator's structured
claim scoring. These two termination modes must be reported separately.

Truncation (length), filtering, errors, empty or ambiguous answers, wrong identity,
multiple tools, duplicate call IDs, invalid argument fields and malformed argument
JSON remain technical unknowns. No simulator action executes for an invalid
response, and no settled outcome is invented. Earlier valid actions remain in
the partial trace. Duplicate JSON keys are rejected. Transport failure preserves
the shared ledger's reservation and stops collection; automatic replay is denied.

Information and attribution
The case builder must supply an audited public packet including generative
support, priors, weights, contract semantics and current observation, never the
realized world index or private environment variables. The adapter checks that
the packet's observation equals the actual checkpoint observation. It cannot
certify arbitrary caller-supplied additional fields are safe; a dedicated frozen
case builder and indistinguishability tests remain required before collection.
Existing operator prefix events are counted separately from model-origin events
and retained in independent physics verification. A normal-text finish is labeled
normal_text; a function call is labeled native_tool. Opaque reasoning items are
preserved for API continuation, not published or interpreted as intent evidence.

Costs and scope
The new module imports the prior reviewed shared ledger and provider admission
controls, replaces the actual native schema with the environment's tool set and
checks its input-byte limit. It exposes no external-action tools and sends no
requests on import. Tests inject synthetic zero-cost responses in temporary
ledgers. No paid route has yet been tested against this interface, and no new
budget has been authorized. A byte bound is not a provider tokenization proof.

Validation
Eight tests cover text refusal and false prose claims, filtering/truncation,
queue alternatives, native schema selection, opaque continuation preservation,
operator-prefix attribution, malformed calls and identity checks, and held
transport failure with no retry. Full repository validation remains required.

Remaining before collection
Build audited case packets and model-visible disclosure parity; define treatment
arms; settle held-out structural template design; freeze scoring and missingness;
derive exact priced counts; authenticate routes without exceeding lifetime cap.
Do not describe this adapter as a validated held-out benchmark or novelty proof.

Commands
python -m unittest tests.test_necessity_structural_native
python -m unittest discover -s tests
python -m benchmark.necessity_native_analysis