Reproducible summaries for the public AscensionAI snapshot. Raw logs, checkpoints, and rollouts stay out of git; this page links to the publishable reports and registry.
Diagnostics ruled out deeper search (88% of Act-2 deaths are "lost on arrival") and raw deck-power (losing decks aren't damage-deficient), leaving upgrade efficiency. The smith was upgrading a basic Strike 97% of the time; targeting the highest-tempo card instead yields 9 wins/150 (6.0%), floor 26.1 — the biggest single-lever floor gain so far, at no HP cost.
Search-combat → first wins. A diagnose-first stack of macro levers (deck hygiene, varN Act-2/3 combat retrain, AoE/draw role scorer, conservative smith) yields mainline_v2: 7 wins/150 (4.7%), floor 24.9 — the project's first winning agent.
Headless-sim variant series (Jun 10–14). Combat was the wall: 1-ply search nearly doubled run depth (floor 14.7→26.4, Act 3 reached), and deck hygiene on top produced the project's first 3 wins.
19,400+ games, network upgrade to (512,256,256) GELU, HP-urgency heal reward. 200-game eval: 38.1% boss WR, 20% Act 2 reach.
12,088 PPO games, 1,488 updates, boss reward shaping for all bosses, accelerated BC anchor decay, and the May 20 200-game evaluation.
5,146 PPO games, 641 updates, and the May 18 200-game evaluation (peak pre-shaping).
4,136 PPO games, 515 trainer updates, and the May 16 150-game evaluation.
~2,500 PPO games, 311 trainer updates, and the May 14 fixed-seed comparison.
Heuristic, BC, and PPO checkpoint comparison on deterministic seeds.
Parallel rollout collection, offline trainer updates, and stale-rollout behavior.
Behavior cloning collection and supervised warm-start training metrics.
| Run | Games | Average Floor | Avg Reward | Boss WR | Act 2+ | Notes |
|---|---|---|---|---|---|---|
| Heuristic 150-game eval | 150 | 15.78 | 8.44 | 39.0% | 26.0% | Reference policy baseline. |
| BC checkpoint 150-game eval | 150 | 12.81 | -0.55 | 30.6% | 12.0% | Playable warm start below heuristic. |
| PPO 5k (200-game eval) | 200 | 15.44 | 4.03 | 31.1% | 20.0% | Peak PPO pre-shaping. Within 0.34 floors of heuristic. |
| PPO 12k (200-game eval) | 200 | 14.83 | -0.95 | 21.7% | 13.5% | Exploration valley post-BC-decay. 49.5% deaths on floor 16. |
| PPO 19k (200-game eval) | 200 | 14.66 | 9.60 | 38.1% | 20.0% | Live-era peak. Network upgrade + HP-urgency heal; boss WR nearly matches heuristic. |
| Sim plateau (varB, headless) | 150k | ~14.7 | — | — | ~26% | Headless-sim era. Plateau confirmed not capacity/entropy (and pre-fix deck-blind). |
| Sim + search combat (48-seed full run) | 48 | 26.35 | — | — | 83.3% | 1-ply search at combat decisions. Act 3 reached (14.6%), max floor 50 — combat was the wall. |
| Sim + search + deck override (48-seed) | 48 | 23.3 | — | — | ~73% | First wins in project history: 3/48. Deck hygiene layered on search-combat. |
| mainline_v2 (150-seed) | 150 | 24.9 | — | — | — | First winning agent: 7 wins (4.7%). varN combat + deck hygiene + AoE/draw scorer + conservative smith. |
| mainline_v3 (150-seed) | 150 | 26.1 | — | — | — | 9 wins (6.0%); biggest single-lever floor gain. + tempo-targeted smithing (Bash+/attacks vs Strikes), no HP cost. |