At a glance
AI Cluster compares Magic Roll, Harvest, Fossil and Rare Roll for each cluster type. Your optimization goal determines which method leads: expected profit per hour, expected profit per finished item, or profit relative to craft cost.
The result depends on the active trade snapshot, resource prices, cluster base price and item level. Calculate with the inputs you intend to use, then review the Run choices. A high displayed rate describes a model; the metrics guide shows its assumptions and a reproducible example.
Follow one plan through the workspace
- Open AI Cluster and choose Plan. In Trade Scan, select the league and inspect the active scan's date and coverage. Trade login is required for a new scan.
- In Best Method, choose the chaos threshold and Optimize for goal. Use Run calculation to compare methods and refresh the craft package used by Run. The calculation refreshes currency prices and uses the active trade data.
- Expand a cluster result and compare its revenue, cost, profit per hour and profit percentage across methods. Check which Harvest reforge or fossil combination the result names.
- Open Run. Choose cluster quantities, whether to buy bases, and the destination after crafting: List in shop, Dump stash, or Neither. Review material requirements and the Live price check and Exalt 4th mod options.
- Start with a small batch. Confirm that the item type, material use and destination match the plan. Save the run's actual quantities and elapsed time before increasing it.
Changing Optimize for can restore an earlier result saved for that goal. Use Run calculation after changing prices or goals when you need a newly computed plan. Run becomes available once its craft package and required run choices are ready.
Calculator scope: this guide uses the Best Method table inside AI Cluster. The Check Item dialog called Compare craft methods evaluates separate node targets.
What the planner is trying to keep
A notable combination has value when the active price catalog assigns it a price. The planner uses spawn weights and mod pools to estimate how often outcomes appear, then models what happens after each outcome.
The base's item level matters. The calculator filters the mod pool using the scan's minimum item level for that cluster, with item level 84 as its default when no minimum is provided. A plan for one item level can give different odds from a plan for another.
The modeled notable limits are one for Small, two for Medium and three for Large clusters. Rare jewels still have separate prefix and suffix limits. A Large cluster's notable model allows up to two prefix notables and one suffix notable. Those restrictions affect whether a continuation can reach the wanted outcome.
Prices and the keep threshold
The chaos threshold selects which priced combinations count toward the plan. For runtime keep eligibility, the comparison also accounts for the base:
catalog value - base cost >= keep threshold
With a 1c base and a 30c threshold, a 30c catalog value leaves 29c after the base. A 31c value leaves 30c. Crafting materials are included in the planner's expected-profit calculation; they are separate from this keep-threshold test.
The catalog can include direct listing observations, cheaper superset prices and estimates for missing combinations. For example, a listing containing A, B and C also satisfies a search for at least A and B. The price preparation code can use that cheaper superset to lower the A+B catalog value. Missing prices can be estimated from neighboring combinations. An estimated row needs more scrutiny than a recent observed listing.
When price evidence is missing
Without usable trade prices, the calculator can fall back to a cost comparison and affinity-based targets. Its fallback takes the top half of notables by affinity when the pool has at most ten notables, or roughly the top third for larger pools, and includes unusually low-weight notables. There is also a specific fallback target pair for Totem Damage.
That fallback can suggest a way to obtain desirable mods. Profit ranking needs sale-price evidence. Treat a result with missing revenue or cost as incomplete for an economic decision.
Why the four methods differ
| Method | Main materials and model | What changes its result |
|---|---|---|
| Magic Roll | Alterations and useful Augmentations, with Regal and reset steps where the path needs them | The chance to create a useful magic item, promotion costs, and the number of individual orb applications |
| Harvest | A named reforge and its lifeforce cost, plus preparation and any modeled continuation | The guaranteed tag, coverage of target notables, lifeforce price and remaining mod pool |
| Fossil | The named fossils and resonator, plus preparation and any modeled continuation | Tag boosts, blocked or reduced tags, fossil prices, resonator price and preparation time |
| Rare Roll | Alchemy and Scouring to repeat rare rolls, plus any modeled continuation | The unrestricted rare-roll pool, material cost and two core orb applications per cycle |
Cheap materials alone do not establish the best method. A method can cost more per use and still reach a valuable accepted item sooner. Another method can produce a high-value average item but take enough rolls to lose on profit per hour.
Reading a Harvest choice
Harvest chooses a reforge by its tag. A Life reforge guarantees a Life-tagged modifier somewhere in the result; it does not guarantee the exact notable pair you want.
The current calculator selects its Harvest candidate by coverage of target notables first, then modeled profit among candidates with equal coverage. It compares that selected Harvest path against the other method families. Read the named reforge in the result instead of assuming that the cheapest reforge wins.
In the historical metrics example, the calculator selects reforge_life for a Medium Damage Over Time cluster. Its input cost is 1.56375c per reforge, while the expected material cost per accepted item is 138.42c. The gap comes from repeated attempts. Those values belong to that captured input set.
Reading a fossil choice
Fossils change relative weights. In the checked-in model, Scorched multiplies fire weights by ten and blocks cold; Aetheric multiplies caster weights by ten and reduces attack weights to 0.15 of their original value. A blocked tag can remove a wanted outcome as well as unwanted competition.
Use the exact fossil combination shown by the planner and check its resonator cost. A blank or unavailable fossil result provides no usable economic forecast for that cluster under the selected inputs.
Choose the goal that matches the decision
| Optimize for | Score used for ranking | Use it to examine |
|---|---|---|
| Profit / hour | Expected profit divided by modeled duration, displayed in div/hour when a divine rate is available | How quickly the model turns active crafting time into expected item value |
| Max absolute EV | Expected profit per finished item, in chaos | The expected value from each completed base |
| Max profit margin % | Expected profit divided by craft material cost | Expected profit relative to material spending |
EV means expected value across repeated outcomes. Individual jewels can cost much more or less than the average.
Expected profit subtracts both craft materials and the cluster base. The displayed margin denominator is craft cost, so it is not a return on all capital committed. Compare base cost separately when bases are expensive. A high percentage on a small profit can still produce a lower hourly result.
The hourly model includes craft actions, continuation orbs, a shared item-handling allowance and a listing allowance. It uses fixed timing assumptions. Your layout, trade availability and interruptions can change actual throughput.
What sell, reroll, exalt and annul mean
The continuation model examines the current notables and available affix slots. It prices continuation orbs and the possible resulting states.
| Decision | Meaning |
|---|---|
| Sell or keep | The item meets the value rule. The selected Run destination determines whether it is listed, dumped or left after crafting. |
| Reroll | Start another crafting attempt on the current base. |
| Exalt | Add an affix when the modeled continuation justifies the orb cost and applicable time cost. |
| Annul | Remove an affix when the modeled path from the resulting states justifies that cost and risk. |
The model protects a keep-eligible item from destructive reroll or annul paths. For an item below the keep threshold, continuation in Profit / hour mode must first beat the forward value of continuing to craft after orb cost. Time also affects the choice. The current price map and orb prices matter more than a rule such as always exalting a certain notable pair.
Exalt 4th mod is a separate user choice. When enabled, it tells the run to fill the remaining mod before sell or dump even when that filler exalt lacks a positive EV case. It spends extra currency by instruction. Record the option when comparing a run with the planner's forecast; the calculator does not take that Run toggle as an input.
How live price checks change a run
With Live price check enabled, the run can recheck a candidate even when the destination is a dump stash. Successful checks update the active price catalog and can change keep or continuation decisions. Newer checks are preserved when bundled price snapshots are merged by date.
With Live price check off and a dump-only destination, the run uses the catalog. A listing destination can still trigger a live confirmation for qualifying items. The option therefore does not act as a global switch for every trade request.
A lower confirmed price can make an item fail the keep threshold and return it to crafting. A lookup with no usable price, or an unavailable trade service, can retain the catalog estimate on fallback paths. Check the run log for the resolved price and any fallback before treating it as refreshed evidence.
Live checks also consume time. The planner has a fixed handling allowance rather than a separate live-on and live-off throughput model. Measure your own elapsed time before attaching an hourly claim to a run.
Check the first batch
Confirm the cluster's enchant, passive count and item level against the planned base. Make sure Ctrl+Alt+C is free for PoE item copy; Windows Copilot or another global binding can steal it. Complete the condensed setup checks for the panels, pins, pixels and OCR the selected path uses.
After the batch, compare material consumption, kept items and elapsed time with the forecast. A completed listing records an asking price. Record a sale when the item actually sells, and calculate realized profit from the currency received and costs paid.
Use Reprice when finished stock needs another price review. Refresh the plan after meaningful market or resource-price changes. Continue with Reading cluster profit per hour for the calculation and sensitivity checks.
Source notes
Behavior was checked against the source available on 20 October 2026. Maintainer references:
| Subject | Source |
|---|---|
| Workspace stages and run readiness | electron-app/src/components/ClusterWorkspace.tsx |
| Calculation, modes and price refresh | electron-app/src/components/BestCraftingMethodDialog.tsx, runCalculation |
| Method selection, inputs and profit | scripts/calc-best-crafting-method.js, main, getTargetMods, selectBestHarvestCraftByProfit |
| Pre-run ranking and units | electron-app/src/utils/clusterPreRunMetrics.ts, rankingScoreForCountParam, rankingMetricForCountParam |
| Continuation and time model | scripts/lib/clusterRuntimeContinuationEv.js, chooseActionProfitPerHour, estimateMethodDurationMs |
| Price preparation and recency | scripts/cluster-trade-utils.js, estimateMissingComboPrices, mergeClusterTradePricesByRecency |
| Runtime live checks and filler option | CraftAutomation/TreeExecutor.Conditions.cs, CraftAutomation/TreeExecutor.ClusterContinuationEv.cs, CraftAutomation/TreeExecutor.cs |
| Tier access | electron-app/src/types/appTier.ts, canUseClusterAi |
For the app and licenses, join the official Discord and DM PathOfRolling. Check the name carefully and keep keys private.