Historical construction comparison. See the current offensive draft and results.
Selected experiment: Faster construction. Faster construction improved both Companions in screening and fresh-seed confirmation, with gains in all six holdout matchups. The combined package added only 0.23 percentage points in screening and slightly trailed construction alone for Keldrax, so the simpler package was preferred. Durability and offense alone showed smaller screening gains. All factions, Companions, card identities and illustrations are retained. Both Prismforge Companions still start at Guard 5. This is a provisional AI-tested balance draft; human balance and real-world play time remain unvalidated.
Mean rounds rose 6.568 → 6.595. Round-12 games rose 23/4608 → 61/4608; monitor the pacing tail in human testing.
Eight distributed seeds, both Prismforge Companions against both Companions from each other faction, both seat orders and starts, and four scoring/aggressive policy pairings. Each candidate has 3,072 games using the exact same matchup/seed matrix. The combined package tests whether the two approaches reinforce one another.
| Package | Wins / games | Win rate | Change | Mean rounds | First build¹ | Spire Renown / game |
|---|---|---|---|---|---|---|
| Unchanged baseline | 1227 / 3072 | 39.9% | 0.00 pp | 6.41 | 4.52 | 0.25 |
| Faster construction | 1276 / 3072 | 41.5% | 1.60 pp | 6.42 | 4.05 | 0.57 |
| Durable infrastructure | 1240 / 3072 | 40.4% | 0.42 pp | 6.40 | 4.52 | 0.26 |
| Offensive pressure | 1241 / 3072 | 40.4% | 0.46 pp | 6.41 | 4.52 | 0.25 |
| Construction + offense | 1283 / 3072 | 41.8% | 1.82 pp | 6.42 | 4.05 | 0.58 |
Twelve unused, distributed seeds; the same matchup matrix; 4,608 games per candidate. Exact 30-card inventories and JSON save/resume round trips were checked after every action. This confirmation set was not used to choose the initial screening leader.
| Package | Wins / games | Win rate | Change | Mean rounds | First build¹ | Spire Renown / game |
|---|---|---|---|---|---|---|
| Unchanged baseline | 1380 / 4608 | 29.9% | 0.00 pp | 6.57 | 3.91 | 0.30 |
| Faster construction | 1640 / 4608 | 35.6% | 5.64 pp | 6.60 | 3.64 | 0.73 |
¹ Mean owner turn of first construction, among games where a Construct was built. This is not an unconditional timing estimate; see construction frequency below. Rounds are engine rounds, not minutes.
| Companion | Baseline | Selected draft | Change |
|---|---|---|---|
| Vitra | 31.9% | 39.8% | 7.90 pp |
| Keldrax | 28.0% | 31.3% | 3.39 pp |
| Opponent | Baseline | Selected draft | Change |
|---|---|---|---|
| Hearthbound | 27.0% | 40.6% | 13.67 pp |
| Skyrunners | 25.0% | 27.5% | 2.47 pp |
| Veilborn | 27.1% | 33.1% | 5.99 pp |
| Ironmaw | 10.4% | 13.9% | 3.52 pp |
| Hourglass Covenant | 32.8% | 38.0% | 5.21 pp |
| Revelwild | 57.4% | 60.4% | 2.99 pp |
| Prismforge | Baseline | Selected draft |
|---|---|---|
| Starts first | 33.0% | 38.1% |
| Starts second | 26.9% | 33.1% |
| Measure | Baseline | Selected draft |
|---|---|---|
| Games with a Construct | '+pct(base.gamesWithBuild/base.games)+' | '+pct(chosen.gamesWithBuild/chosen.games)+' |
| Builds per game | '+num(base.buildsPerGame)+' | '+num(chosen.buildsPerGame)+' |
| Spire bonus Renown per game | '+num(base.spireRenownPerGame)+' | '+num(chosen.spireRenownPerGame)+' |
| Defeated Construct lifetime² | '+num(base.defeatedLifetimeOwnerTurns)+' | '+num(chosen.defeatedLifetimeOwnerTurns)+' |
| Games reaching round 12 | '+base.round12+' | '+chosen.round12+' |
² Owner-turn difference between deployment and defeat, for defeated Constructs only. Survivors are censored and excluded; this cannot establish overall survival or compare durability without that caveat. Returning a Construct with Salvage is not counted as defeat.
The original implementation trial used only two adjacent seeds. A second two-seed pilot produced a dramatically different baseline. We therefore replaced adjacent seed sampling with eight distributed screening seeds and twelve separate distributed confirmation seeds. The older figure was a small-sample result, not a reliable estimate of faction strength. This expanded report supersedes that estimate.
24,576 distributed-seed candidate games are included here, plus the retained initial pilot. Screening save round trips occurred every 25 actions, with exact inventories at turn transitions and match end. Confirmation round-tripped and checked inventories after every action. All runs checked Charge, action-loop bounds, defeated units, Construct slots and legal Route locations.
Games sharing a seed, mirrored seats or policies are dependent. We report paired exact-match comparisons and per-seed changes rather than treating every game as an independent observation. The same limited AI is used across candidates; it does not plan complete multi-action combinations. No claim of statistical certainty, universal balance, or human win rates is made.
Next: human playtests of construction timing, opponent counterplay and companion choices. Revisit further changes only with a separate comparison against this draft.
Browser layout was not tested in this environment. Automated interface checks execute the real launch, private ordering, upgrade, Construct and resume controls in a minimal DOM.