The run
One packing procedure changed one requirement:
label_position: LEFT → TOP_RIGHT. This is what happened.
Every screenshot and every number below is recorded evidence from a real run against real Google Cloud, with real photographs and real Gemma inference. It is not a live interaction, and this page does not call any model.
What the photographs looked like
The verification chronology
Both attempts are kept. A change that took two tries is a different operational fact from one that passed immediately.
| # | Expected | Observed by Gemma | Truth Engine verdict |
|---|---|---|---|
| 1 | TOP_RIGHT | LEFT | FAIL |
| 2 | TOP_RIGHT | TOP_RIGHT | PASS |
The model reported a position. The verdict is arithmetic performed
by deterministic code against the value recorded at remediation time —
DRIFTZERO TRUTH ENGINE deterministic comparator.
The model calls, as recorded
| Output | Model | Latency | Tokens in / out |
|---|---|---|---|
LEFT | google/gemma-4-26b-a4b-it-maas | 5.918 s | 341 / 2 |
TOP_RIGHT | google/gemma-4-26b-a4b-it-maas | 5.272 s | 341 / 4 |
Two inferences, both on Vertex AI MaaS, serverless and on-demand. The
model's entire contribution is a position from a closed set: LEFT,
TOP_RIGHT or INCONCLUSIVE. Anything else is rejected.
What the worker and the auditor saw
