[tarekziade]

Follow-up to #153. That check judged where a value went, not the value itself, so the red-flag PRs would pass it once reshaped onto ("cuda", 8). It also flagged a correct edit on the 2026-10-11 nightly.

• Converted literals are drift-checked. When a plain literal becomes an Expectations block, the runner value is compared against the literal it replaces. transformers#49243 (pvt_v2) moved 45% and skipped the check.
• A different count of numbers is a regression. transformers#47698 (dab_detr) went from 5 detections to 3.
• Degenerate runner text is rejected: a control character (transformers#49244 fuyu, \x04 A bus…) or a stutter (transformers#47983 starcoder2, def hello_world(): hello_world():). This is expectation_guard.degenerate_text, which also feeds the guard's degenerate-value label.
• A sign flip no longer hides an expectation-only patch. _NUMBER_LITERAL now includes a unary minus. In the guard, #47698's -2.5104 left a stray -, so the patch read as a code change and no expectation rule ran on it in prod.
• Asserts on the model's input shape are exempt from the runner-key rule, e.g. inputs.input_ids.shape == (2, 422). transformers#49507 (pegasus) was sent back 3 times and published with a wrong ⚠️ warning. The tokenizer doesn't depend on the GPU. The exemption needs an input name plus .shape/len(/.size(, so an output count is still checked.

The exhausted-budget behaviour is unchanged: the PR still opens with the warning.

Checked against the real base/head files of transformers PRs:

case	before	after
8 red flags (#47698 #47750 #47840 #47983 #48955 #49143 #49243 #49244)	7 rejected, #47698 passed	8 rejected
same bad values reshaped onto ("cuda", 8) (pvt_v2, starcoder2, fuyu, dab_detr)	4 passed	4 rejected
#49507 pegasus	rejected	passes
4 eligible (#48838 #48882 #48945 #49040)	pass	pass

make test: 1404 passed.

Ships on deploy (TASK_EXPECTATION_KEY_CHECK still switches it off).

🤖 Generated with Claude Code