8 days ago
-
Queues can draw the main job from a held-out split (ops-queue-<split>-NNNN); side jobs stay public
HassamGani
-
'Left alone' checks judge the final data: a record put back exactly as it started counts as untouched
HassamGani
9 days ago
-
Refunds summed from invoice lines billed on the withdrawn card; retraction wording for service credits
HassamGani
-
Rival readings, runbooks, hard-to-find approvals, contract reasoning, validated leaving corpus, credit history tool
HassamGani
10 days ago
-
Ops-queue tasks: three jobs at once in one company, one of them changing
HassamGani
-
Shotgun baselines; collateral guard for hand-written tasks
HassamGani
-
Policy updates, hold reasons, billed seats, incident credits, expired contracts, merged tickets; clause-level reply grading; null-agent baselines; relay mode
HassamGani
-
Drop references to other benchmarks
HassamGani
-
Add an expert tier to the generator: three changes, up to 64 customers, two more traps; token cap scales with the budget
HassamGani
-
Add the task generator: three scenario templates, 210 train tasks; staffing and retirement handle carried-over work
HassamGani
11 days ago
-
Add a hard tier of five tasks; let oracle plans split a role across agents
HassamGani
-
Add belief probes, bootstrap intervals and a paired compare command; count unfinished carried-over work when judging retirement
HassamGani
-
Add the internal render client to t05's repo; let refunds name a charge; say what the test runner runs
HassamGani
-
Let the orchestrator ask the person questions, answered from pre-written facts
HassamGani
-
Fix task issues found in the pilot; give agents the date
HassamGani
-
Add LLM harness and run command; clearer test output and charge records
HassamGani
-
Add metrics, seven more tasks, splits and RL interface
HassamGani
-
Add environment, runtime, oracle and first three tasks
HassamGani