Safety override may over-suppress Past in Flames combo branches #13

Closed
opened 2026-07-06 17:51:07 +00:00 by bstein · 1 comment
Owner

What happened

In deployed Draw-Go vs Storm g04, the safety layer replaced model-selected Past in Flames lines with land or creature development because the same-turn permission evaluator saw no visible follow-up mana. The game still ended in a Storm win later, but the override may be too coarse when reducers and graveyard resources make the combo branch plausible.

What should have happened

The same-turn permission safety rule should distinguish generic no-follow-up flashback grants from Storm combo turns where reducers, graveyard spells, floating mana, and Past in Flames can make the follow-up line real.

Category

  • kind: misplay
  • severity: debatable

Context

  • correction target: ["safety_override", "same_turn_permission_evaluator", "storm_combo_context"]
  • evidence: [{"frames": "161/165", "game id": "edge_1v1_storm__1_g1_2026070604", "level": 4, "local log": "artifacts/deployed_batteries/edge_case_0521_live_20260706T124234Z/logs/dashboard/edge-case-smoke-deployed-proof-20260706t101310z-edge-1v1-draw-go-vs-storm-g04-level-4/edge_1v1_storm__1_g1_2026070604/game.replay_events.ndjson", "run group": "edge-case-smoke-deployed-proof-20260706t101310z-edge-1v1-draw-go-vs-storm-g04-level-4", "summary": "Safety overrides replaced model-selected Past in Flames lines with land/creature development under casual_same_turn_permission_no_visible_followup_mana."}]
  • feedback issue key: explicit:same_turn_permission_safety_override_suppressed_past_in_flames_branch
  • recommended source files: ["src/veles/priority_windows/temporary_combat_modifiers.py", "src/veles/strategy_policy_runtime/policy_ranking.py", "src/veles/codex_safety/runtime_flags.py", "src/veles/codex_decision_worker/legal_action_context.py", "src/veles/playtest_context/permission_summaries.py"]
  • root cause: same_turn_permission_safety_override_suppressed_past_in_flames_branch
  • source findings file: reports/edge_case_atomic_eval/0.5.21_live/draw_go_storm_unique_feedback_findings_20260706.json
## What happened In deployed Draw-Go vs Storm g04, the safety layer replaced model-selected Past in Flames lines with land or creature development because the same-turn permission evaluator saw no visible follow-up mana. The game still ended in a Storm win later, but the override may be too coarse when reducers and graveyard resources make the combo branch plausible. ## What should have happened The same-turn permission safety rule should distinguish generic no-follow-up flashback grants from Storm combo turns where reducers, graveyard spells, floating mana, and Past in Flames can make the follow-up line real. ## Category - kind: misplay - severity: debatable ## Context - correction target: ["safety_override", "same_turn_permission_evaluator", "storm_combo_context"] - evidence: [{"frames": "161/165", "game id": "edge_1v1_storm__1_g1_2026070604", "level": 4, "local log": "artifacts/deployed_batteries/edge_case_0521_live_20260706T124234Z/logs/dashboard/edge-case-smoke-deployed-proof-20260706t101310z-edge-1v1-draw-go-vs-storm-g04-level-4/edge_1v1_storm__1_g1_2026070604/game.replay_events.ndjson", "run group": "edge-case-smoke-deployed-proof-20260706t101310z-edge-1v1-draw-go-vs-storm-g04-level-4", "summary": "Safety overrides replaced model-selected Past in Flames lines with land/creature development under casual_same_turn_permission_no_visible_followup_mana."}] - feedback issue key: explicit:same_turn_permission_safety_override_suppressed_past_in_flames_branch - recommended source files: ["src/veles/priority_windows/temporary_combat_modifiers.py", "src/veles/strategy_policy_runtime/policy_ranking.py", "src/veles/codex_safety/runtime_flags.py", "src/veles/codex_decision_worker/legal_action_context.py", "src/veles/playtest_context/permission_summaries.py"] - root cause: same_turn_permission_safety_override_suppressed_past_in_flames_branch - source findings file: reports/edge_case_atomic_eval/0.5.21_live/draw_go_storm_unique_feedback_findings_20260706.json
bstein added the
misplay
battery-0.5.21
manual-review
draw-go-vs-storm
labels 2026-07-06 17:51:07 +00:00
Author
Owner

Verified in Cassandra 0.5.30.

Release proof: source bebd4c7 (Prepare Cassandra 0.5.30 feedback backlog closeout), Jenkins Veles #209 SUCCESS published version-only backend/frontend/sim-worker 0.5.30 images, Titan-IAC 1bf84977 deployed them, live /api/v1/app-config reports backend/frontend/sim_worker 0.5.30, /api/v1/ready is green, backend pod and both frontend pods are Running with 0 restarts, and VELES_SIM_IMAGE is registry.bstein.dev/veles/veles-sim-worker:0.5.30.

Local/source gate on the deployed source: bridge/runtime 42 tests OK; evaluator/audit/policy/API suite 504 tests OK; frontend maybeboard/compact-replay/prompt Playwright 22 tests passed; frontend build and git diff --check passed.

Issue-specific proof: policy/evaluator coverage keeps legitimate Past in Flames combo branches from being over-suppressed while preserving safety filtering. Closing.

Verified in Cassandra 0.5.30. Release proof: source bebd4c7 (Prepare Cassandra 0.5.30 feedback backlog closeout), Jenkins Veles #209 SUCCESS published version-only backend/frontend/sim-worker 0.5.30 images, Titan-IAC 1bf84977 deployed them, live /api/v1/app-config reports backend/frontend/sim_worker 0.5.30, /api/v1/ready is green, backend pod and both frontend pods are Running with 0 restarts, and VELES_SIM_IMAGE is registry.bstein.dev/veles/veles-sim-worker:0.5.30. Local/source gate on the deployed source: bridge/runtime 42 tests OK; evaluator/audit/policy/API suite 504 tests OK; frontend maybeboard/compact-replay/prompt Playwright 22 tests passed; frontend build and git diff --check passed. Issue-specific proof: policy/evaluator coverage keeps legitimate Past in Flames combo branches from being over-suppressed while preserving safety filtering. Closing.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: cassandra-alpha/feedback#13
No description provided.