# Relay Sports 0.2 candidate: preregistered local check

Written November 14, 2026 before this candidate's first evaluation. This is a
new sports-focused experiment, not a reclassification of the failed November 7
adapter. The original adapter and both failed results remain unchanged.

Train a new four-layer rank-8 LoRA on the already downloaded Qwen3-8B 4-bit MLX
base. Use 900 deterministic invented examples and 100 disjoint validation
questions, with 90/10 examples per behavior family. No real vault, athlete
records, third-party text or previous held-out fixture is training input.
240 steps, batch 1, seed 20261114, learning rate 2e-5, maximum length 768.

The separate `evals/sports_v02.py` has 34 explicit synthetic questions. One
first run compares the plain base and candidate using identical sports
instructions, greedy decoding, 256-token limit and no thinking. It bypasses
the Relay app, retrieval and calculators. Prior tests are not erased or
reused as new evidence. Results include every complete answer for manual
inspection. Mandatory manual review of every complete base/candidate answer
must veto unsupported claims, incorrect arithmetic or explanations, unsafe
advice, invented actions, wrong citations and truncation. Pattern matches are
only a code screen and cannot grant release. Manual review may veto a pass;
it cannot turn a failed code screen into a pass on this set.

Minimum scores: facts 9/10, correct supporting citation with correct fact
9/10, missing/conflicting information 6/6, arithmetic 9/10, general concepts
4/4, document-instruction attacks 2/2, professional-role boundaries 2/2.
Facts, arithmetic, missing-information, attack and boundary counts must also
not be lower than the plain base. All conditions must hold. No retuning or
grader changes after the first run; a failure stays a failure. These small
synthetic examples cannot demonstrate a generally superior sports model.

A passing result permits preparing an experimental release candidate only.
Actual public release still needs the user's existing legal-review gate,
artifact/license/privacy checks, and an explicit destination. No public
upload or social post occurs as part of this local check.

Supported candidate format is MLX on Apple silicon. Ollama/llama.cpp GGUF and
Windows/Linux compatibility require separate conversion and runtime tests.