Run the CORD expense smoke¶

Kind: how-to.

This page runs the bounded CORD expense triage smoke (#165). The smoke asks a three-label question about 18 synthetic expense claims against six real CORD v2 receipts. It records semantic metrics and gates on structure and image attachment only.

The smoke is not a production readiness claim, and it does not measure model quality beyond these 18 claims.

Before you start¶

The llama.cpp router serves a Gemma 3 or Gemma 4 multimodal id with its vision projector. Common ids are gemma-3-4b-it-q4km-mm and gemma-4-31b-kv9-q4km-mm. See Run the image-conditioned live smoke.
The alias gemma-4-31b-kv9-q4km-mm is not a Q4_K_M file. On the operator's router it loads gemma-4-31b-24gib-kv9-decoder.gguf, 16.0 GB. /props reports ftype Q2_K - Medium. The router uses a --chat-template-file override on image server-cuda-b11243 (#233).
/apply-template for that id renders the native Gemma 3 or Gemma 4 turn. The test fails on any other template family.
The fixture is tests/fixtures/cord/expense_smoke/ (CC-BY-4.0). Each receipt has one supported, one contradicted and one insufficient claim.

Run the live smoke¶

TYPEVET_LLAMA__MULTIMODAL_MODEL=gemma-4-31b-kv9-q4km-mm \
  TYPEVET_LLAMA__TIMEOUT=900 \
  uv run pytest evals/tests/live/test_cord_expense_smoke_live.py -m live -q

Each pass writes an immutable receipt under scratchpad/cord-expense/receipt-<run_id>.json (the prior receipt.json name is not overwritten). A second write with the same run_id fails and leaves the first file unchanged. Before any scoring call, the harness snapshots prompt wording and three kinds of digests. The digests cover the evaluated harness source, the expense-smoke manifest, and each receipt PNG the smoke attaches. Finalization records call counts only and does not re-read those paths. Thus late edits cannot change experiment_identity digests.

Result	Cause
Skip	The router is down, or the model id is not in the catalog
Fail on `text-only input modalities`	The router serves that id without a projector
Fail on `the image was not attached`	The prompt token gap vs text or omission did not grow
Fail on a label or probability assertion	A response is not one of the three labels
Pass	Every request is typed and every image is attached

A pass says nothing about accuracy. Read the metrics in the receipt.

The three modalities¶

Modality	Text	Image	Requests
`text_only`	The claim statement	None	18
`image_only`	A claim that states no amount	The receipt	6
`combined`	The claim statement	The receipt	18

image_only runs once for each receipt, because its text is the same for all three claims, plus one image_only omission control judgment. The total is 43 requests. The test asserts the total is 54 or less.

Read the receipt¶

Each scored row may carry provenance from #183: application_mode, receipt_required, routing, model_calls, and deterministic_abstain. When combined or image_only would run without receipt bytes, the harness returns insufficient_evidence with routing=deterministic_missing_receipt and zero model calls. That path is application validation, not model abstention, and must not inflate #161 insufficient_abstention_rate as learned behavior.
metrics.combined holds the semantic metrics: accuracy, per-verdict recall, macro_recall, insufficient_abstention_rate, false_supported and confusion_gold_by_predicted.
metrics.text_only is the text prior. Without the receipt, the model cannot settle a claim, so use it as a baseline, not as a score.
image_only_insufficient_rate is how often the model abstains when no amount is claimed.
soft_warnings lists metrics below chance, an abstention rate under 0.5, and claims routed to supported when gold says otherwise. The test does not fail on these warnings.

Check the attachment yourself¶

Compare tokens_evaluated for the same claim in text_only and combined. One Gemma 3 image costs 256 prompt tokens on the CORD smoke router. A Gemma 4 image has no fixed cost. Gemma 4 has a variable image-token budget that depends on the image. In the #203 run, the gap was 228 to 1,108 tokens per receipt (#203 per-claim table). A silently dropped image grows the count by the marker text only. The smoke counts that marker as 31 tokens. The live harness calls assert_cord_expense_live_smoke_gate before scoring. typevet_evals.cord.expense_smoke then sets one gap floor for each model.

Model	Gap floor	Reason
Gemma 3	225 tokens	Fixed 256-token image minus the 31-token marker
Gemma 4	197 tokens	Smallest measured 228-token image gap (#203) minus the 31-token marker

Each arm is compared with its image-omitted control (#260). For image_only, compare each receipt row to image_only.omission_tokens_evaluated, the same claim text with no image. The test passes the served family to ScoringJudgmentAdapter as served_template. So text_only and combined share one native turn family, and the gap holds the image and its marker only. Every /completion request sends "cache_prompt": false.

Historical run (Gemma 3)¶

On 2026-09-26 against gemma-3-4b-it-q4km-mm, recorded with the smoke from revision 387285e, before aa1ad37. At that revision the text_only prefix was ChatML and the imaged prefix was the native Gemma 3 turn. So the token gap below also holds a small template difference. Run the smoke again before you quote a value for the current revision.

Metric	`combined`	`text_only`
Accuracy	0.333	0.278
`supported` recall	0.5	0.0
`contradicted` recall	0.5	0.833
`insufficient` recall	0.0	0.0
`false_supported`	4	1

The model did not choose insufficient_evidence in any modality, including image_only. The image changed the label on 8 of 18 claims. Each attached receipt added 256 prompt tokens once both arms used the native Gemma 3 turn at aa1ad37.

Historical run (Gemma 4, direct arm)¶

On 2026-09-26 against gemma-4-31b-kv9-q4km-mm, recorded with the smoke from revision 25c5195. The receipt does not record the file behind the alias. On 2026-09-29 the alias loaded gemma-4-31b-24gib-kv9-decoder.gguf, ftype Q2_K - Medium (not Q4_K_M), 16.0 GB. The router used a --chat-template-file override on image server-cuda-b11243 (#233). The vendored receipt is tests/fixtures/cord/expense_smoke/gemma4_kv9_direct_receipt.json. The router declared vision and /apply-template rendered native_gemma4_turn.

The direct combined arm passes attachment and typed-label gates. It fails two #161 revision 1 semantic floors on the combined arm (recorded, not live-gated):

The vendored receipt’s experiment_identity block is flawed on this historical run. working_tree.dirty was falsely false, and identity was captured after scoring. The live harness path is missing from code_path_digests, and runtime.server_build is unknown. See the sibling annotation gemma4_kv9_direct_receipt.note.md. Semantic FAIL rows below are unchanged; only identity metadata is annotated.

#161 check	Limit	Measured	Status
Answerable accuracy	floor 0.67	0.25	FAIL
Contradicted recall	floor 0.5	0.333	FAIL

Combined accuracy was 0.444. The model abstained on most imaged rows; that is a quality read, not proof the receipt was unread.

Check semantic acceptance on a saved receipt¶

After a live run writes a combined receipt JSON, evaluate the frozen #161 revision 1 floors offline. Exit code 0 means accepted: true; 1 means checks failed; 2 means usage or malformed input.

uv run python scripts/check_cord_semantic_acceptance.py path/to/receipt.json

Equivalent module entry:

uv run python -m typevet_evals.cli.cord_semantic_acceptance path/to/receipt.json

Vendored examples (no live model):

uv run python scripts/check_cord_semantic_acceptance.py \
  tests/fixtures/cord/expense_smoke/gemma4_kv9_direct_receipt.json

uv run python scripts/check_cord_semantic_acceptance.py \
  tests/fixtures/cord/semantic_acceptance/gemma4_post187_combined_pass.json

Pytest regression tests for the evaluator and CLI are separate from operator acceptance; see Release support matrix.

Known limits¶

One model, one router build, six receipts, 18 claims.
No calibration claim. Metrics are recorded, not gated.
Claims are synthetic. Masked-digit claims stand in for unreadable evidence.