Jira Metrics
Projects/Test AI Agents(TAA)

Cycle Time

Days from first move into an active status until the issue is resolved.

How fast we actually workactive · business hrs ÷ pauses

2.1d

Avg

1.3d

Active Time (÷ pauses)

3d

P75

3.6d

P85

4.4d

P90

Wall-clock (start → done)raw · calendar days

2.7d

Avg

2.3d

Calendar Time (raw)

3d

P75

6.5d

P90

12

issues in range

Weekly Median trend

Contributing tickets (12)

🔍
12
KeySummaryAssigneeCycle (active)Completed
TAA-1Orchestrator autoplay UI: always show the real/actual seed for RL agentsOliverAI4.9d7/1/2026
TAA-2Root-cause the argmax-collapse: why does greedy (deterministic) eval score ~0% when stochastic scores ~43%?OliverAI4.5d(+0.4d paused)7/1/2026
TAA-11Rule-server loss-heuristic false-positives on draw-heavy-but-winnable positionsOliverAI3.1d7/3/2026
TAA-12Deterministic (argmax) eval arm logs NaN/empty win_rate instead of a numberOliverAI3d7/3/2026
TAA-6Background solvable-seed collector (Azure CPU job)OliverAI3d7/3/2026
TAA-18Model registration silently didn't happen on completed training jobs with register_model=trueSemenyuk Dmitriy1.6d7/7/2026
TAA-20Fail fast when a training config needs the solver but the job image doesn't have itOliverAI1d7/7/2026
TAA-19Training runs don't log the demonstration-imitation signal (bc_loss / bc_weight) — add the instrumentationOliverAI1d7/7/2026
TAA-16Autoplay orchestrator: serve the new 3900-input observation schema (blind_v8) alongside the current 1011-input oneOliverAI1d7/7/2026
TAA-7Make the canonical lever-screen statistically valid (train-to-budget + N≥2000), forbid underpowered verdictsOliverAI1d7/4/2026
TAA-13Wire parallel eval to MLflow and verify it matches serial on the real RL championSemenyuk Dmitriy0.5d(+0.5d paused)7/4/2026
TAA-5TAA-2 follow-up: run the argmax-collapse E1-E7 investigation and deliver the findings doc (PR #34887 added only the tooling)Semenyuk Dmitriy0d7/1/2026
Showing 112 of 12
Page 1 of 1