Choose better cuts. Then test again.
9,296 simulated candidate evaluations across deck selection, fresh validation, Advanced checks, usability prototypes and optimized package matchups. See the sections below for the sample size of each comparison.
What we learned
Seven of the nine full packages gained wins against optimized base opponents in the initial validation; Revelwild was flat and Prismforge was slightly down. Cathedrax alone was a positive Prismforge result. A further cut-refinement check won 34 / 64 versus 31 / 64 for its base control and 33 / 64 for the earlier package—one extra win from the revised cuts on that same sample. The usability prototypes did not show a reliable win-rate improvement and have not been adopted. Individual cards still miss the target in several factions.
What changed in the test
Bounded selection: test legal base-only copy-count variants, individual rare substitutions, and sampled packages. Prismforge tests every distinct base card name as each single-card cut. Other factions use up to seven cuts per rare (cost/role matches, expensive cards, redundant copies). Package candidates include combined individual winners and reproducibly sampled cuts. Rank on eight training games against two starter opponents, then freeze one base and each rare build. Validate on 64 fresh paired scenarios against the selected base-only decks of all eight other factions, both own and enemy Companions and both starting seats. Select the overall rare build using training results only; do not pick the best validation outcome.
This is not an exhaustive optimal-deck search. Only cards already present in faction base lists are eligible for base copy-count variants. Some factions have just one legal distinct base candidate under those limits. Intermediate AI plus a lab-only Prismforge adapter is used for selection and validation; conclusions are about these simulations, not human win rates. The adapter evaluates Blueprint using unordered remaining Construct counts and chooses Observatory Routes using public board state. Source rules/stats remain unchanged. Seed-bootstrap intervals reflect simulation variability only. There are only 64 games per build and rare exposure can be low. No 2v2 evidence.
Fresh validation: the rare build chosen before testing
Each rate has 64 games. “Selected rare build” is locked from training and may contain one rare or all three. The full package is also shown separately. Draws count as half a win.
| Faction | Base-only control | Selected rare build | Change | All three rares | Package change |
|---|---|---|---|---|---|
| Hearthbound | 50.0% | 51.6% package | +1.6 pp 95% seed interval -9.4 pp to +12.5 pp | 51.6% | +1.6 pp |
| Skyrunners | 51.6% | 51.6% tactic | +0.0 pp 95% seed interval -4.7 pp to +4.7 pp | 56.3% | +4.7 pp |
| Veilborn | 39.1% | 43.8% package | +4.7 pp 95% seed interval +0.0 pp to +10.9 pp | 43.8% | +4.7 pp |
| Ironmaw | 39.1% | 46.9% package | +7.8 pp 95% seed interval +1.6 pp to +14.1 pp | 46.9% | +7.8 pp |
| Hourglass Covenant | 60.9% | 67.2% low | +6.3 pp 95% seed interval +1.6 pp to +12.5 pp | 67.2% | +6.3 pp |
| Revelwild | 70.3% | 70.3% package | +0.0 pp 95% seed interval -7.8 pp to +7.8 pp | 70.3% | +0.0 pp |
| Prismforge | 56.3% | 54.7% low | -1.6 pp 95% seed interval -7.8 pp to +4.7 pp | 54.7% | -1.6 pp |
| Vigilant Array | 43.8% | 53.1% package | +9.4 pp 95% seed interval +1.6 pp to +18.8 pp | 53.1% | +9.4 pp |
| Elderwake | 48.4% | 57.8% package | +9.4 pp 95% seed interval +1.6 pp to +18.8 pp | 57.8% | +9.4 pp |
Prismforge: deck choice and AI handling
The previous policy tended not to cast Blueprint because its one-action evaluation stopped before the private search resolved. The lab adapter can cast it when remaining Construct density, an open build slot and Charge justify searching, using deck composition rather than hidden draw order. It also places Cathedrax around friendly Constructs and public blockers.
On the same selected package and 64 validation scenarios: reference policy 50.0%; card-aware policy 54.7%. Difference: +4.7 pp. This comparison keeps the deck fixed; it is separate from changing which cards are removed.
Hearthbound
Selection tested 8 base-only candidates, 6 / 6 / 6 cuts for the high-cost character / tactic / low-cost character, and 20 distinct packages. Original starter validation: 42.2%; selected base-only control: 50.0%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Hearthroot Matriarch | Lampkin | 43.8% | -6.3 pp |
| tactic | Reclaim the Hearth | Lampkin | 43.8% | -6.3 pp |
| low | Kindleseed Apothecary | Cindertusk | 54.7% | +4.7 pp |
| package | Hearthroot Matriarch Reclaim the Hearth Kindleseed Apothecary | Lampkin Turn Back Cindertusk | 51.6% | +1.6 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Hearthroot Matriarch | 30 | 18 | 18 |
| tactic · Reclaim the Hearth | 30 | 29 | 0 |
| low · Kindleseed Apothecary | 23 | 23 | 23 |
| package · Hearthroot Matriarch | 30 | 16 | 16 |
| package · Reclaim the Hearth | 26 | 24 | 0 |
| package · Kindleseed Apothecary | 23 | 23 | 23 |
Skyrunners
Selection tested 9 base-only candidates, 5 / 7 / 6 cuts for the high-cost character / tactic / low-cost character, and 22 distinct packages. Original starter validation: 43.8%; selected base-only control: 51.6%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Auric Cloud Regent | Veilwing | 51.6% | +0.0 pp |
| tactic | Windward Procession | Thundercrest | 51.6% | +0.0 pp |
| low | Zephyr Needle | Reedrunner | 56.3% | +4.7 pp |
| package | Auric Cloud Regent Windward Procession Zephyr Needle | Veilwing Thundercrest Reedrunner | 56.3% | +4.7 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Auric Cloud Regent | 24 | 7 | 7 |
| tactic · Windward Procession | 16 | 16 | 0 |
| low · Zephyr Needle | 18 | 18 | 16 |
| package · Auric Cloud Regent | 25 | 7 | 7 |
| package · Windward Procession | 15 | 15 | 0 |
| package · Zephyr Needle | 18 | 18 | 16 |
Veilborn
Selection tested 5 base-only candidates, 4 / 6 / 5 cuts for the high-cost character / tactic / low-cost character, and 20 distinct packages. Original starter validation: 32.8%; selected base-only control: 39.1%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Maw Beyond the Map | Grinshade | 40.6% | +1.6 pp |
| tactic | Veil of Reversal | Unseal the Netherway | 39.1% | +0.0 pp |
| low | Threadblind Acolyte | Vesperant | 42.2% | +3.1 pp |
| package | Maw Beyond the Map Veil of Reversal Threadblind Acolyte | Grinshade Unseal the Netherway Vesperant | 43.8% | +4.7 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Maw Beyond the Map | 21 | 8 | 8 |
| tactic · Veil of Reversal | 27 | 26 | 0 |
| low · Threadblind Acolyte | 30 | 30 | 23 |
| package · Maw Beyond the Map | 21 | 8 | 8 |
| package · Veil of Reversal | 25 | 24 | 0 |
| package · Threadblind Acolyte | 30 | 30 | 23 |
Ironmaw
Selection tested 9 base-only candidates, 4 / 7 / 6 cuts for the high-cost character / tactic / low-cost character, and 19 distinct packages. Original starter validation: 34.4%; selected base-only control: 39.1%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Thronetusk Colossus | Ridgebreaker | 40.6% | +1.6 pp |
| tactic | Banner of Predators | Ridgebreaker | 32.8% | -6.3 pp |
| low | Ashfang Whelp | Snapmaw | 43.8% | +4.7 pp |
| package | Thronetusk Colossus Banner of Predators Ashfang Whelp | Gorequill Bloodrush Snapmaw | 46.9% | +7.8 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Thronetusk Colossus | 32 | 25 | 20 |
| tactic · Banner of Predators | 32 | 32 | 0 |
| low · Ashfang Whelp | 26 | 25 | 23 |
| package · Thronetusk Colossus | 33 | 24 | 21 |
| package · Banner of Predators | 35 | 34 | 0 |
| package · Ashfang Whelp | 26 | 25 | 24 |
Hourglass Covenant
Selection tested 8 base-only candidates, 4 / 6 / 5 cuts for the high-cost character / tactic / low-cost character, and 19 distinct packages. Original starter validation: 50.0%; selected base-only control: 60.9%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Aevora, Keeper of Unlived Hours | Secondhand Sentinel | 60.9% | +0.0 pp |
| tactic | The Hour Nobody Remembers | Secondhand Sentinel | 57.8% | -3.1 pp |
| low | Minutehand Scribe | Pendulum Warden | 67.2% | +6.3 pp |
| package | Aevora, Keeper of Unlived Hours The Hour Nobody Remembers Minutehand Scribe | Secondhand Sentinel Eonbreaker Pendulum Warden | 67.2% | +6.3 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Aevora, Keeper of Unlived Hours | 21 | 5 | 5 |
| tactic · The Hour Nobody Remembers | 21 | 21 | 0 |
| low · Minutehand Scribe | 21 | 21 | 21 |
| package · Aevora, Keeper of Unlived Hours | 21 | 5 | 5 |
| package · The Hour Nobody Remembers | 20 | 20 | 0 |
| package · Minutehand Scribe | 21 | 21 | 21 |
Revelwild
Selection tested 7 base-only candidates, 3 / 6 / 5 cuts for the high-cost character / tactic / low-cost character, and 20 distinct packages. Original starter validation: 59.4%; selected base-only control: 70.3%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Orabella, Carnival Crown | Crescendo Hulk | 71.9% | +1.6 pp |
| tactic | One More Impossible Night | Crescendo Hulk | 71.9% | +1.6 pp |
| low | Mirthspark Firewalker | Grumbleglow | 70.3% | +0.0 pp |
| package | Orabella, Carnival Crown One More Impossible Night Mirthspark Firewalker | Crescendo Hulk Empty Seats Grumbleglow | 70.3% | +0.0 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Orabella, Carnival Crown | 22 | 8 | 6 |
| tactic · One More Impossible Night | 22 | 22 | 0 |
| low · Mirthspark Firewalker | 26 | 26 | 14 |
| package · Orabella, Carnival Crown | 22 | 10 | 10 |
| package · One More Impossible Night | 18 | 18 | 0 |
| package · Mirthspark Firewalker | 27 | 27 | 15 |
Prismforge
Selection tested 9 base-only candidates, 17 / 17 / 17 cuts for the high-cost character / tactic / low-cost character, and 24 distinct packages. Original starter validation: 51.6%; selected base-only control: 56.3%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Cathedrax, Living Observatory | Salvage Order | 62.5% | +6.3 pp |
| tactic | Blueprint of Tomorrow | Salvage Order | 54.7% | -1.6 pp |
| low | Facetwisp Caretaker | Shard Courier | 54.7% | -1.6 pp |
| package | Cathedrax, Living Observatory Blueprint of Tomorrow Facetwisp Caretaker | Facetbreaker Lattice Vanguard Shard Courier | 54.7% | -1.6 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Cathedrax, Living Observatory | 28 | 21 | 21 |
| tactic · Blueprint of Tomorrow | 28 | 28 | 14 |
| low · Facetwisp Caretaker | 26 | 26 | 25 |
| package · Cathedrax, Living Observatory | 28 | 19 | 19 |
| package · Blueprint of Tomorrow | 23 | 23 | 10 |
| package · Facetwisp Caretaker | 23 | 23 | 23 |
Vigilant Array
Selection tested 1 base-only candidates, 5 / 7 / 5 cuts for the high-cost character / tactic / low-cost character, and 23 distinct packages. Original starter validation: 43.8%; selected base-only control: 43.8%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Last Horizon Executor | Orison, Last Directive | 46.9% | +3.1 pp |
| tactic | Terminus Protocol | Intercept Protocol | 48.4% | +4.7 pp |
| low | Orison Gatelet | Orison, Last Directive | 50.0% | +6.3 pp |
| package | Last Horizon Executor Terminus Protocol Orison Gatelet | Latchguard Intercept Protocol Orison, Last Directive | 53.1% | +9.4 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Last Horizon Executor | 11 | 4 | 3 |
| tactic · Terminus Protocol | 22 | 22 | 1 |
| low · Orison Gatelet | 11 | 11 | 10 |
| package · Last Horizon Executor | 25 | 8 | 5 |
| package · Terminus Protocol | 22 | 22 | 1 |
| package · Orison Gatelet | 11 | 11 | 10 |
Elderwake
Selection tested 1 base-only candidates, 5 / 7 / 5 cuts for the high-cost character / tactic / low-cost character, and 22 distinct packages. Original starter validation: 48.4%; selected base-only control: 48.4%.
| Build | Add one copy each | Remove one copy each | Validation rate | Change vs base |
|---|---|---|---|---|
| high | Itharyn, Continental Warden | Cragling | 45.3% | -3.1 pp |
| tactic | World Before Footsteps | Seismic Roar | 51.6% | +3.1 pp |
| low | Mosswake Larvalith | Hollowpeak Ancient | 57.8% | +9.4 pp |
| package | Itharyn, Continental Warden World Before Footsteps Mosswake Larvalith | Oruun, the Walking Horizon Deep Nourishment Hollowpeak Ancient | 57.8% | +9.4 pp |
Ability exposure and usability gaps
Counts below are games, not raw triggers. “Affordable” means the rare was in hand during at least one main-phase decision with enough Charge. The raw hand-action offer counter includes paid discards and is not treated as a legal-cast counter. Ability counters are included in the raw download.
| Build / card | Reached hand | Affordable | Played |
|---|---|---|---|
| high · Itharyn, Continental Warden | 19 | 11 | 8 |
| tactic · World Before Footsteps | 22 | 22 | 0 |
| low · Mosswake Larvalith | 26 | 26 | 24 |
| package · Itharyn, Continental Warden | 16 | 8 | 5 |
| package · World Before Footsteps | 17 | 17 | 0 |
| package · Mosswake Larvalith | 26 | 26 | 24 |
Prismforge follow-up: keep the strong character
With the selected single-card cut, Cathedrax won 40 / 64 validation games versus 36 / 64 for the optimized base control (+6.3 pp). It replaced Salvage Order. That is a positive lead for the existing character, not a reason to lower its cost automatically.
Original Advanced policy, fresh scenarios
Four opposing factions, both Companions and both starting seats: 32 games per build. The selected rare build and full package were locked from training. These rates are not directly comparable to the all-opponent Intermediate sample.
| Build / variant | Wins (draws count half) | Rate | Change vs control |
|---|---|---|---|
| optimizedBase | 11 / 32 | 34.4% | +0.0 pp |
| selectedRare | 13 / 32 | 40.6% | +6.3 pp |
| package | 11 / 32 | 34.4% | +0.0 pp |
Resolve conflicting cuts before changing cards
Both Cathedrax and Blueprint selected the lone Salvage Order as their preferred individual removal. The refinement tested 50 valid combinations drawn from the four best cuts for each rare. The next package was locked before another fresh 64-game comparison.
| Build | Remove one copy each | Fresh wins | Rate |
|---|---|---|---|
| base | Emergency Assembly | 31 / 64 | 48.4% |
| oldPackage | Facetbreaker Lattice Vanguard Shard Courier | 33 / 64 | 51.6% |
| resolvedPackage | Salvage Order Emergency Assembly Shard Courier | 34 / 64 | 53.1% |
Usability prototypes: publish the unsuccessful tests too
These are experiments, not changes to the public card definitions. Costs, stats and illustrations remain as published. Each prototype comparison uses the same deck cuts and paired seeds within its own experiment.
First prototype: cheaper search and arrival-turn Insight
Blueprint costs 1 instead of 2 Charge; Caretaker gains +1 Insight for its arrival turn if a friendly Construct is already at its Route. Neither change improved win outcomes in this 576-game experiment.
| Build / variant | Wins (draws count half) | Rate | Change vs control |
|---|---|---|---|
| base | 33 / 64 | 51.6% | +0.0 pp |
| lowArrivalInsight | 36 / 64 | 56.3% | +4.7 pp |
| packageBoth | 31 / 64 | 48.4% | -3.1 pp |
| lowOriginal | 36 / 64 | 56.3% | +4.7 pp |
| packageArrivalInsight | 31 / 64 | 48.4% | -3.1 pp |
| tacticCost1 | 33 / 64 | 51.6% | +0.0 pp |
| packageCost1 | 31 / 64 | 48.4% | -3.1 pp |
| tacticOriginal | 33 / 64 | 51.6% | +0.0 pp |
| packageOriginal | 31 / 64 | 48.4% | -3.1 pp |
Second prototype: free revealed build and persistent Insight
Blueprint retains its 2-Charge cost and may build one revealed Construct at an empty normal Route for free. Caretaker retains its printed stats and has +1 Insight while sharing a Route with a friendly Construct. Targeted checks verify card conservation, slots, Charge, and losing the Insight bonus when the Construct leaves.
| Build / variant | Wins (draws count half) | Rate | Change vs control |
|---|---|---|---|
| base | 33 / 64 | 51.6% | +0.0 pp |
| lowPersistentInsight | 36 / 64 | 56.3% | +4.7 pp |
| packageBoth | 31 / 64 | 48.4% | -3.1 pp |
| lowOriginal | 36 / 64 | 56.3% | +4.7 pp |
| packagePersistentInsight | 34 / 64 | 53.1% | +1.6 pp |
| tacticFreeBuild | 33 / 64 | 51.6% | +0.0 pp |
| packageFreeBuild | 31 / 64 | 48.4% | -3.1 pp |
| tacticOriginal | 33 / 64 | 51.6% | +0.0 pp |
| packageOriginal | 34 / 64 | 53.1% | +1.6 pp |
Combined second prototype with Advanced AI
Four opposing factions and fresh seeds; 32 games per build. This small check tests the combined prototype, not individual card contributions.
| Build / variant | Wins (draws count half) | Rate | Change vs control |
|---|---|---|---|
| base | 12 / 32 | 37.5% | +0.0 pp |
| packageBoth | 12 / 32 | 37.5% | +0.0 pp |
Optimized packages against one another
576 additional games compare optimized base/base with optimized three-rare package/package matches. Each faction has 64 games per format. Rates average 50% in both formats; these AI results describe this bounded set of decks and policies, not human balance.
| Faction | Optimized base field | Optimized rare-package field | Relative change |
|---|---|---|---|
| Hearthbound | 57.8% | 46.9% | -10.9 pp |
| Skyrunners | 46.9% | 42.2% | -4.7 pp |
| Veilborn | 45.3% | 42.2% | -3.1 pp |
| Ironmaw | 42.2% | 43.8% | +1.6 pp |
| Hourglass Covenant | 48.4% | 50.0% | +1.6 pp |
| Revelwild | 67.2% | 68.8% | +1.6 pp |
| Prismforge | 51.6% | 54.7% | +3.1 pp |
| Vigilant Array | 50.0% | 48.4% | -1.6 pp |
| Elderwake | 40.6% | 53.1% | +12.5 pp |
Release target
Every faction should gain useful deck-building options from its rares. A forced three-rare package is not assumed optimal. Cards that fail after sensible substitutions and adequate use need further timing, targeting, cost or reliability experiments. These results do not justify declaring every rare balanced or changing cards solely to make this sample win.
Download all candidate lists, selections, exact decks, seeds, ability counters and results