"""
The fixed `choice` schema the Phase 1 `ekvachan-base` checkpoint was trained
on. Canonical source: serve/inference.py's `LABELS` constant.

Duplicated here as a plain literal, rather than imported, because
serve/inference.py imports torch and transformers at module level -- and
deciding which benchmark items are even answerable is cheap, dataset-only
work that has nothing to do with running the model. It needs to work (and
needs to run in CI or a laptop with no GPU/ML stack installed) well before
any backend that can actually answer a question gets instantiated. See
benchmarks/common/backends.py and benchmarks/common/harness.py for how the
two are kept apart on purpose.

Kept honest against drift: every real backend (InProcessBackend, HTTPBackend
in benchmarks/common/backends.py) is checked against this constant before
it's trusted to score anything filtered against it -- see
benchmarks/common/harness.py's consistency check. If serve/inference.py's
LABELS ever changes, that check fails loudly instead of this file silently
going stale.
"""

FIXED_CHECKPOINT_LABELS = ["entailment", "neutral", "contradiction"]

# --- Decoder / multischema path (PRD.md Section 14 Q4, 13a.5) ---------------
#
# The decoder (Qwen3.5-4B LoRA, restricted-logit read) is architecturally not
# fixed-width the way the encoder's classification head is: it answers a
# `choice` question by reading the model's next-token logits restricted to a
# small set of single-token option-code strings, one per option, masking the
# rest to exact -inf before softmax. That mechanism's real ceiling is however
# many such codes exist as guaranteed single, mutually distinct tokens under
# the base model's tokenizer.
#
# PRD 13a.5 verified this at 26 (all of A-Z), and that number stood
# unquestioned through 13a.15 -- until PRD 5.1b re-checked the premise rather
# than the conclusion and found the real number is 588: A-Z, then every
# two-uppercase-letter pair (AA, AB, ... ZZ) that also happens to be a single
# token under this tokenizer's BPE merges (562 of 676 are). See
# `training/decoder_lora_lib.py`'s `build_code_table()`, which derives and
# asserts this table at runtime from a live tokenizer -- that function is the
# canonical source for this number, duplicated here as a plain literal for
# exactly the reason FIXED_CHECKPOINT_LABELS above is: filtering benchmark
# items by option count is cheap, dataset-only work that must not require
# torch/transformers/peft to be installed.
#
# Unlike FIXED_CHECKPOINT_LABELS, this is only a *default* cap used when
# filtering happens before any backend is instantiated (this module's whole
# reason to exist). A real DecoderMultischemaBackend reports its own loaded
# checkpoint's `max_options` from its manifest, and benchmarks/common/
# harness.py cross-checks that against whatever cap was used to filter before
# trusting the run -- same discipline as the FIXED_CHECKPOINT_LABELS check.
DECODER_MULTISCHEMA_MAX_OPTIONS = 588