DeepDesk-Bench

DeepDesk-Bench tests whether an AI agent can perform a realistic European gas analyst workflow: inspect a synthetic LNG operations portal, reconcile vessel and cargo evidence, and update an existing Excel supply-and-demand model without damaging it.

This first release focuses on LNG scheduling and spreadsheet execution. All market data are synthetic.

• Source repository
• DeepDesk-Bench article and results
• Browse LNG-PortalSim
• Download the starting Excel workbook

Tasks

Task	What the agent must produce
01-vessel-recon	A JSON list of material inbound cargoes and supporting evidence
02-cargo-reconciliation	A JSON decision table for the LNG schedule
03-schedule-update	An updated workbook using supplied cargo decisions
04-full-update	An updated workbook produced from the portal and prior-close model

The first three tasks isolate evidence gathering, analytical judgment, and spreadsheet execution. Task 04 tests the complete workflow end to end.

Run the benchmark

The canonical Harbor registry is in the GitHub repository. Run all four tasks with:

harbor run --repo qrlow/evals -d snd-floaters -a <agent> -m <model>

Run the bundled oracle solutions from a checkout of the source repository:

harbor run -p snd-floaters -a oracle

Each task starts its own local copy of LNG-PortalSim at http://localhost:8000 inside the candidate environment. The task instructions allow LNG-PortalSim as the only external information source.

What is included

• Four self-contained Harbor task packages
• The synthetic LNG-PortalSim portal and frozen evidence records
• The starting Excel workbook and deterministic scoring code
• Oracle solutions and verifier tests
• A compact result from the August 2026 campaign
• benchmark_tasks.jsonl, a small browser-friendly index of the four tasks

The maintained source assets live under shared/. Task-local copies are kept because Harbor mounts each task as an independent package.

Scoring

Scoring is deterministic; no language model acts as a judge. JSON tasks are checked against frozen reference data. Workbook tasks are inspected directly for correct cargo decisions, numerical outputs, evidence, formula preservation, and safe handling of uncertainty.

Task 04 uses a 100-point score:

Category	Points
Operational LNG schedule decisions	40
Numerical balance accuracy	25
Provenance and cutoff discipline	20
Workbook integrity and formula preservation	10
Uncertainty handling	5

Critical penalties cover post-cutoff evidence, unsupported or duplicate cargoes, unresolved cargoes included in the balance, and hardcoded LNG output cells.

Notes

This is an early proof of concept, not a general ranking of models for commodities work. The included campaign ran each model-task pair once, so the published scores should be read as preliminary diagnostic results rather than stable estimates of model quality.