# Tartarus choices and combinations

Run `npx --no-install vite-node scripts/tartarus/tiers.ts` from the repository root to rebuild `website/src/guides/data/tartarus-tiers.json`. The generator uses the committed WASM engine, its legal actions, the static evaluator's heuristic policy and full playouts to the Riot. It rejects engine fallbacks and unfinished playouts. The output uses synthetic games and the repository's own rules and rollouts.

## What a point means

One point is one additional contained Chromatic under the default Chromatics goal. The engine also gives credit for the rarity of inmates still on the board at the Riot. Their weights are 0 for Normal, 0.002 for Magic, 0.01 for Rare, 0.05 for Legendary and 0.2 for Ultimate. The Objective's bonus cache belongs to a different scoring goal. These simulated gains are neither drop rates nor market prices. [Engine goals](../../src/core/tartarus/goals.ts) and [cache weights](../../src/core/tartarus/score.ts).

Each score is the mean final utility with an action or loadout minus the mean with its comparator on the same boards and future seeds. Every board has equal weight within its row. Scores and synergy are rounded to four decimals for display. Apply the following editorial cutoffs to the unrounded gain, including the boundary:

| Choice | S | A | B | C |
| --- | ---: | ---: | ---: | --- |
| Persistent run instruments and their pairs | ≥2 | ≥1 | ≥0.25 | <0.25 |
| One-off decisions, Chromatic picks and Opulence sequences | ≥0.5 | ≥0.25 | ≥0.05 | <0.05 |

`positiveBoards` counts boards on which the mean gain over the comparator was positive (floating-point tolerance 1e-9). For combinations it measures the total gain. Compare it with the row's `boards` to see how consistently that gain occurred. It is a descriptive count, not a confidence interval. A tier near a cut or an interaction measured on few boards can change with the seed set.

## Sample and comparators

Generate 40 full runs with seed `mix(503, run)`, starting at index zero. Cycle through None, Tyranny, Opulence, Iron Law and Cage Fight in engine ID order, eight runs each. Each run keeps its theme until the Riot, including Iron Law's 30-day duration. Equal representation intentionally gives unusual themes more weight than live offer frequency. A snapshot uses the seed `mix(runSeed, actionIndex)`. For each snapshot use eight future seeds `mix(snapshotSeed, 0..7)`. The scenario effects use the future seed and their continuations use a separate derived seed. All alternatives for a comparison share their seeds, though changes to the board or policy can change later draws.

| Family | Board and question |
| --- | --- |
| Run Treasure (Torture Instruments) | Use the first clear destination screen on days 2, 8 and 15 in each run (120 boards). Replace the held loadout with zero, one or two tested instruments. Apply each pickup through the engine, including its current-day destination penalty. Mark the Objective as completed in every arm to suppress further instrument offers. Compare with empty. The existing board, marks, destination hand and historical effects stay as sampled. This measures having the equipment from this morning onward. Actual start and Objective offers arrive at different times, and rare offers are harder to assemble. |
| Chromatic Treasure | Capture the first fresh Chromatic popup in each seven-day band per run. The four cards are the same on every popup. Compare each first pick with waiving only that pick. In Opulence, the policy still plays the two remaining picks in both arms. Extra Chromatics queue additional popups separately. Waiving a pick is an experimental comparator. |
| Opulence sequences | On the Opulence subset of those popup boards, compare skipping the current popup with applying exactly the named cards in order. Suppress unused picks from that popup to isolate each effect. The remaining queue and future rewards still play normally. The separate `opulenceSingles` rows use these very same boards and comparator for calculating synergy. Two-card results measure the isolated contribution of the first two picks. Three-card results cover the full popup. |
| Inmate decisions | For each acting Code, take its first opportunity per run with a legal Consume. Compare Consume (mean across legal targets) or Replace (only when still legal) with End. These are actions taken with that Code on the sampled boards. The model has no free choice of an arbitrary inmate species or rarity. |
| Destinations | For each Code take the first multi-card hand per run offering a visible inmate of that Code. Compare moving to a matching cell against the mean of other offered moves. Exclude hands with no alternative. Unopened cells can contribute to the comparator, but their hidden Code never identifies a candidate. |
| Other actions and room cards | Hunt uses a separate 20-run set starting with Blood-Stained Trap, seeded with `mix(0x503b, run)` and the same fixed theme cycle. At the first legal Hunt in those runs or first Chain in each main run, compare the mean across legal targets with End or Stop Chain respectively. At the first Interrogation Room popup compare Instigate against Bestow and vice versa. Each family uses its own board count and comparator. |

This seed set produced 88 fresh popup boards, including 16 Opulence boards. The 80 ordered Opulence sequences therefore share just 128 board/future combinations each. Selecting the largest measured interactions from that many candidates can overstate their advantage on new runs.

No step treats the popup as a fifth card or offers a third run instrument. Run Treasure is limited to the opening and the Objective. One Chromatic normally grants one pick. Under Opulence each Chromatic grants three picks from the same four cards, and repeats are legal. The engine also caps its pending queue. These figures describe the model's rules, policy and sampled boards; they do not rank a card regardless of board position, remaining days, current goal or held instruments. [Engine choices](../../src/core/tartarus/engine.ts) and [rules and assumptions](../tartarus-goals.md).

## Combination selection

Score all 441 distinct legal pairs of run instruments; the engine excludes incompatible species instruments. Keep the two highest positive synergies with positive total gain and the two largest total gains, removing overlapping entries. For Opulence evaluate all 16 ordered pairs and all 64 ordered triples with replacement. Keep up to two positive-gain, positive-synergy sequences of each length, ordered by synergy. A combination's synergy is its marginal gain minus the sum of all constituent single gains over the same baseline. Count repeated picks every time. The ten-Chromatic cap can make a high-gain pair subadditive. Triple synergy includes lower-order interactions.

The whyKeys identify what the effects can do together. They are mechanical interpretations, not measured causal explanations for the magnitude of a synergy estimate. The script never reads competitor tier lists or player logs.
