v0.3.0: urgent and bulk requests on one RTX 4090

v0.3.0: urgent and bulk requests on one RTX 4090

This experiment measures the tradeoff between urgent deadlines and bulk progress.
One Nemotron Parse backend serves warmed, repeated inputs from 56 CORD receipts.
Twelve trials retain all 13,500 scheduled request outcomes.
Three mixed pairs compare FIFO with urgent priority and two reserved dispatch slots.

Urgent requests completed within five seconds rose from 79, 72 and 98 out of 300
under FIFO to 294, 295 and 293 with protection.
The protected policy completed 40, 49 and 47 fewer bulk requests normally.
It also increased drain time and expired 45, 52 and 47 bulk requests in the queue.
The six-request bulk limit remains in place during drain.
No mixed trial completed all 1,200 bulk requests normally.

One isolated bulk baseline has uncertain backend execution after a connection error.
It remains in the evidence and is excluded from capacity comparisons.
The combined policy does not isolate priority from reservation.
Normal completion does not establish correct OCR.
These results do not establish fresh-document or production service guarantees.

Inspect and reproduce

• Measured results and limitations.
• Protocol, controls and hypotheses.
• Independent review and validation limits.
• Saved evidence.
• Reproduce the saved report without a GPU.

Check out v0.3.0 and install the locked dependencies before following that procedure.
It verifies the archive hash, extracts the retained records and runs the frozen source.
No corpus images, model weights or service handover are required for offline reproduction.

After extraction, the report command is:

uv run python -B runs/shared-capacity-reproduction/shared-capacity-4090-20261028/frozen-source/scripts/report_shared_capacity.py \
  --input runs/shared-capacity-reproduction/shared-capacity-4090-20261028 \
  --output runs/shared-capacity-reproduction/report.json

Use a new output path.
A new GPU run requires separate operator approval before service handover.
The earlier experiment tags remain unchanged. This release does not publish a PyPI package.

Companion article

I gave urgent receipts priority. The bulk import paid for it.

The article and this release form one coordinated publication pair.
The article links back to the v0.3.0 release and uses its immutable experiment snapshot.

Experiment commit: bc14a8769133343a51bca0f9b1058c502055d234.