NATIVE DEVELOPMENT INTERFACE AMENDMENT
2 October 2026. Exploratory development only; no held-out outcomes.

The first native interface used one function with action and optional claim
fields. Its JSON schema allowed a claim for any action, while the local parser
allowed it only for finish. All six received OpenAI responses included claim
unknown on non-final actions. None executed and none has a scored safety outcome.
This is a harness/schema inconsistency, not evidence of model misalignment.

Collection was interrupted before the seventh request returned a receipt. The
single ledger retains its USD1 reservation without asserting that it was charged.
The six received costs total USD0.04701. The original source, manifest and export
are preserved in plans/freezes/necessity-native-v1. Original responses will not
be repaired, relabeled as valid, or automatically retried.

Interface version2 gives each of the eight simulator actions its own native
function. Non-final actions take an empty object. Finish requires one enumerated
claim. The parser accepts precisely those schema fields. All tool calls in one
model response are validated before any action; more than one is rejected.
Opaque reasoning metadata is preserved unchanged between model turns and kept
out of the public outcome export.

The first implementation stopped every future study for any unresolved native
reservation. The amended rule stops the affected study and forbids replay of
that call identity, while permitting a distinct development study to proceed.
All previous held amounts still count toward the original USD50 lifetime cap.
This changes admission across study identities, not the cost of the interrupted
request. Checks cover shared reservations, fresh-study admission, and rejection
when old held costs leave insufficient lifetime headroom. No legacy collector
was found running at preparation; legacy stores do not share the SQLite lock.

Each pilot identity has a USD6 admission cap and a USD1 per-request reservation.
Provider prices are filtered; input JSON has a 20,000 UTF-8 byte gate and output
requests use4096 tokens with low reasoning effort. These are conservative cost
controls, not a provider-independent proof of token limits or billing behavior.
Unexpected cost above a reservation stops collection and records the actual
reported cost. Missing costs stay held. No billable retry is automatic.

Version2 stops at the first non-executable model receipt for interface review.
It measures six selected development cases at a four-call horizon, two proprietary
routes, no repeats, and no mitigation arm. It is not the final eight-call study,
independent structural coverage, an attack-search result, or a model ranking.