[Superwheat]

Problem and resulting behavior

Superwheat/Gimbal-27B is a typed decision model served by its
own released helper (helper/gimbal_serve, vLLM pooling runner) at POST /v1/systemone. The shared
llama-cpp adapter does not fit that server: the server's native Noul type rejects outcome definitions, a
request holds at most 64 questions, and the full-precision probabilities are only in raw_probabilities.

This PR adds --adapter gimbal, a client for a running server that never starts, stops or configures it:

• Noul questions are sent as an explicit two-option Choice over the authored true/false definitions, in
  declared order, and mapped back to false/true (noul_transport: explicit-choice-v1).
• Choice keys are anonymous; Score levels are sent in ascending numeric order and mapped back to the
  declared option IDs; structured instructions and descriptions keep their JSON values.
• States other than strings or objects, and JSON scalar values, are sent as their JSON text.
• Cases above 64 questions are sent in consecutive chunks of 64 with the same state. The server scores
  each question independently (premise = that question's instructions plus the state, one hypothesis per
  candidate), so no question's candidates are split.
• Probabilities are the server's unrounded calibrated raw_probabilities from --full-precision, validated
  without renormalization. A server without --full-precision is a visible failure.
• Overflow is rejected by the server against its --max-tokens limit (HTTP 400) and kept as an unretried
  failure; transient statuses and transport errors get a bounded retry. Response bodies are never recorded.
• Kwargs through --adapter-kwargs: base_url, served_model, timeout, case_batch_size (1..32),
  retries (0..5) and runtime_settings (recorded metadata only).

OPEN_MODELS.md gains a Gimbal-27B section with the server command, transport rules and the input limit
(the longest english-v1 branch is 14,713 tokens with the model's tokenizer, so --max-tokens 16384 covers
the scope). evaluator/README.md links it.

Validation

• tests/test_gimbal.py (39 synthetic tests with httpx.MockTransport): transport of all three tasks with
  nontrivial option order, structured values and authored Noul definitions; numeric Score level order;
  JSON-text transport of array and scalar values; chunking of a 65-question case; malformed responses
  (missing full-precision maps, wrong keys, sums, booleans, tasks, IDs, usage, non-object payloads); model
  drift; inputs rejected before any call (over 256 options, over 10 levels, images, Box, duplicate IDs);
  unretried 400/413 without echoing bodies; bounded retries for 5xx and transport errors; batch alignment
  with individual failures; invalid settings; CLI kwargs passing and rejection of foreign kwargs.
• uv run tox (py311, lint, type): 725 passed, 7 skipped; ruff check . and ty check pass.
• CPU end-to-end check against the released helper (gimbal_serve at the pinned checkpoint, its real
  request schema, serialization, tokenizer and calibration, with a fake scorer in place of the GPU): all 137
  english-v1 benchmarks at dataset revision 10bcb8c6916b98b966cc1f03aab1781a7c354b4c completed with zero
  failed decisions and validated; 18,702 requests, 169,395 branches.
• Real-model GPU smoke and full run: see the linked Dataset PR (smoke smoke-v1 --limit 2, then the full
  scope on one B200).

Related result submission

Dataset PR for Gimbal-27B english-v1: https://huggingface.co/datasets/hotchpotch/s1mb-result/discussions/114

Language scopes and training exposure

English (english-v1) only. This PR changes code and documentation only; the result PR's metadata.json
carries the training-exposure declarations.