Route simulation

Goal: find, for each faction and each race's start, the 1-60 leveling route that reaches 60 in the least /played time under one explicit model of the game, by simulating very large numbers of complete routes; measure how robust the winners are to the model's uncertainty, and score Polestar's own routes under the same model.

"Proven" here means: best found under the model, after a search large enough that more search stops improving it, and still best (or within a stated margin) across thousands of plausible variations of the model's parameters. Real-world proof comes from recorded play (record report); sim calibrate fits the model's kill time to the same recordings.

Code: tools/lib/sim/ (Node ≥24, CommonJS, zero npm dependencies, worker_threads for parallel search). CLI: node tools/polestar.js sim <cmd>. Outputs: data/sim/ and data/bench/ (both gitignored).

1. World model (built once per faction, cached in data/sim/world-<faction>.json)#

• Quests available to the faction (quests.json race/class masks; class quests kept with their class): level, minLevel, prerequisites, exclusive quests, next links, giver and ender positions (world coordinates via map-frames.json), objectives with workload:
  • kill N of a mob: N kills at the mob's spawn clusters (npcs.json levels). A kill objective on an NPC of the character's faction that gives or ends the quest, or whose objective text opens with speak/listen/talk/hear/ show/wait/watch, is the community quest database's kill credit for a conversation (Kalaran Windblade 3441/3453, Tirion Fordring 5742, the faction leaders of 93739/93963): one interaction at the NPC, never elite; in such a quest the other faction's NPC is that faction's counterpart and costs nothing. A deed on an NPC of the character's own faction proper (A or H; text opening with find/heal/use/locate/keep searching: healing Grunt Kor'ja, waking Lazy Peons) is one interaction per credit. A friendly-flagged NPC that is really fought (the AH Threshwackonator 2078) stays a kill;
  • collect N items: N / (drop chance) kills of the best dropper, or N object uses for object sources, or a purchase from the friendly vendor nearest the quest giver. A source on another landmass than the quest giver is used only when the giver's has none. Drops without a recorded rate (Forever creatures, null in items.json) count at 30%, quest-only items (questItem) at 60%, unless a dropper of the item with a recorded rate lives on the giver's landmass or has no open-world site (Kum'isha's sphere: 0.01% recorded, one null dropper, stays rare); on Zephras Isle, where every recorded strider is on another continent, Galestrider (null) is the source of 92553's meat;
  • quest-starting items: the expected kills of the best solo dropper; with no recorded rate for the item at all, its droppers count at the same 30% / 60% (91741 Nibbled-On Book, 97281 A Simmering Storm), elites still not. An item handed over by another quest's turn-in (the data has no quest rewards; startItemFrom in tools/lib/sim/solo.js: the Paladin Tome of Divinity 1642/1646 from 1641/1645) starts the quest at that quest's ender once it is done;
  • interact/use/event: one visit (plus the object's site); an event counted in its text ("Free 6 Webbed Forsaken", "Motivate 5 Deathguards") is that many interactions;
  • escort/event quests: a fixed duration (parameter); a rescue reads as an escort ("Help Perith Stormhoof escape Palemane Rock", 98430).
• Placementzone map frames overlap (the Barrens map covers Durotar), so one spawn can be recorded on two maps. Such a spawn counts once, on the map most spawn records around it use (every NPC and object, 150 yd cells), then the map most of the entity's own spawns use, then the map where it lies furthest from the edge. Two of the Valley of Trials' Lazy Peons are also recorded on the Barrens map; they stay in Durotar, so Lazy Peons is a Durotar 1-5 quest, not a Barrens one. Gryphon Master Talonaxe stays in The Hinterlands (9.75, 44.47, not Hillsbrad 94.3, 3.5).
• Givers in several places (Minor Manifestation of Earth: the Valley of Trials and Mulgore): the quest starts at the giver's site nearest its turn-in (Call of Earth 1518 in the valley, 1521 in Mulgore).
• Several givers or enders (Cenarion Aid 8254: Brother Joshua in Stormwind and High Priest Rohan in Ironforge; the two Wanted Posters of Tanaris): the route search uses the first friendly one the data lists, and the world also keeps every friendly one by entity (gAlt, eAlt: { "npc:<id>": point }, only when there is more than one), so a guide is scored at the one it names (§5).
• Child quests (the community quest database parentQuest): a child is offered only while its parent is in the log, so the parent is not its prerequisite and the child is accepted after the parent's accept and before its turn-in. A parent whose objective items are its children's rewards turns in after them (turninAfter in tools/lib/sim/solo.js: Stranglethorn Fever 348 after 349; The Green Hills of Stranglethorn 338 after chapters 339-342).
• Prerequisitesthe canonical blocking prerequisites of tools/lib/data/prereqs.js, the ones the guide compiler and validator use: preAll all-of, preAny (else the Era quest database's pre) any-of; breadcrumbs (an any-of group holding one is optional as a whole: 788 Cutting Teeth needs neither 787 nor 4641), exclusive quests, the parent and quests nobody can start (786's 785) never block. A DungeonJournal-only quest (no quests.json record) keeps its journal chain. A quest with no prerequisite data at all (no pre/preAny/preAll/breadcrumbs/breadcrumbFor) gets an implied any-of over every quest whose next it is (e.g. 5093/5094/5095 -> 5096, never only the first). world.preRules counts each rule.
• Exclusive quests are one-way: quests.json excl merges the Era quest database's close (symmetric, both sides list each other) and the community quest database's exclusiveTo (one-way). A quest is unavailable while any quest of its own excl is in the log or done; taking it closes nothing else. So 287 stays open after 282 and 322 after 324, while 526 (excl 322, 324) closes once 324 is taken and 5601/5142 close each other.
• Tasks group a quest's objectives by site (spawn clusters, 70 yd).
• Hubsgivers and enders clustered (150 yd), with flight masters, inns and class trainers nearby (trainers.json, TravelData).
• Travel graphwalking (the walking router's routed walks where sim walks has them, below; else the straight line × detour factor, per continent), flights between known flight paths (TravelData flight times), boats, zeppelins and trams (wait + ride, data/forever/travel.json), hearthstone (bind, 60 min cooldown, parameter). All-pairs shortest paths over the transport nodes are precomputed per known-flight-path set; a leg is walk-to-node + network + walk-from-node (the network part must use a flight or a ride), or walking, whichever is cheaper. A place the character cannot reach (another landmass with no transport, e.g. back to Zephras Isle) skips the step and is counted (unreachable).
• Walks (sim walks, tools/lib/sim/walks.js): the walking router the arrow uses (addon/Polestar_Nav, run standalone by nav eval's harness) walks every pair of the world's points (quest givers, enders, objective sites, inns, transport nodes, starts) up to 4,000 yd apart on the same landmass, as the faction (its hostile towns cost more) in the arrow's default route style (balanced): one Path.FindMany search per point, on worker threads (about 10 min per faction on 15 workers), into data/sim/walks-<faction>.json. A walk between two of those points is then the routed yards × routeMul (prior 1 ± 0.05: a player strays a little from the drawn line), and so are the walks to and from transport nodes, between nodes, and the order of a package's steps; a longer walk, or one the router could not find, keeps the straight line × detour. The file is used only on the world (data hash) and walking maps (PACK.json) it was built on, and a search result records whether it was searched with walks, so it replays only on the same (--no-walks walks the straight line × detour anyway). Measured on the 3,103 walking legs the default guides of the 8 starts use (2026-10-17): the straight line × 1.3 is right on average (routed 1.31 Eastern Kingdoms, 1.34 Kalimdor) but wrong both ways, about 2.1 h per start too slow within zones (open ground is nearer 1.1-1.2) and 2.1 h too fast between them, nearly all of it 97 legs straight through a mountain or a cliff (Silithus to Feralas: 1,785 yd straight, 13,807 walked; the Barrens to Thousand Needles: 325 and 4,168). A factor per zone or per pair of zones fixed neither (the error is the leg's, not the zone's), so the walks are routed one by one. With walks the search needs longer to settle: --paths 3000000 (the default 1,000,000 stopped Human at a route 5 h worse).
• Travel safetyeach zone's level is the P10 of its hostile normal-rank spawns (the maps build ZoneLevels heuristic, tools/lib/maps/zonelevels.js), and each 200-yard cell of a landmass belongs to the zone with the most NPC spawns in it (gaps filled from neighbours up to 600 yd). A direct walk whose straight line (sampled every 200 yd) crosses a zone, other than the ones it starts and ends in, whose minimum level is 5 or more above the character costs deathRisk seconds per such zone (deaths, corpse runs, detours; prior 300 ± 150 s), so the flight/boat path wins when it is safer. Crossed zones are sampled once per point pair and cached (measured cost: 0.22-0.29 ms per full route evaluation, vs 0.20-0.24 ms without).
• Packagesthe unit the route search orders. A package is one visit to one zone doing a set of quests: built by splitting each zone's quests by quest level into bands (default width 5, parameter) and keeping chains inside a package when their levels allow; every quest belongs to exactly one package per split. Each package precomputes: entry hub, internal task order (nearest neighbour + 2-opt, as draft does), internal walking distance, per-quest kill counts by mob level, quest XP tier estimates, and the flight paths and inn it offers.

Solo filter (tools/lib/sim/solo.js)#

The solo world keeps only quests a levelling character can do alone. Each exclusion has one category, counted by sim world (world.solo.counts):

Starts (route rules)#

2. Character and XP/time model (parameters with priors in data/sim/params.json)#

• XP to level: the Classic table (tools/lib/game/xp.js, shared by every tool).
• Quest XP: the client's QuestXP table (the game-table export DB2 QuestXP, cached under data/sources/<tables>) at the quest's level and an estimated difficulty tier: breadcrumbs (no objectives, talk to someone) low, normal kill/collect quests middle, chain ends, escorts and group quests high; recorded quest XP (quests.json xp) overrides the estimate. Reduced by level difference: 100% up to quest level + 5, then 80/60/40/20/10%.
• Kill XP: the Classic mob formula (45 + 5 × level, ±5% per level of difference above, zero-difference table below, nothing for grey), elites ×2, rested XP optional.
• Kill time: per-level base time (fight + downtime) × level-difference factor × class factor; travel between mobs inside a site from spawn density.
• Speeds: run 7 yd/s, mount ×1.6 from mountLevel (default 40), flight and transport times from TravelData.
• Grinding fills gaps: when no package is worth doing, the character grinds at the best site for its level (kill XP per hour at Δ = 0..+2). "Grinding" is only killing for XP with no quest involved; kills for objectives and drops are quest work (xpSplit.kill).
• Shared kill credit (game rule): a kill counts for every quest in the log that wants that mob. Kills (and drops) for one kill or collect objective also count for the other active quests' objectives on the same mob whose site is within 100 yd (SHARE_R, run.js), up to what each still needs; that objective then needs only the rest, and one fully covered needs no trip. Kills before a quest is accepted give it nothing. The kill XP is earned once (sharedKills in results: the kills saved).
• Route policies (not measurements; reported times stay real): routes are compared by time to 60 with every hour of grinding counted 2.5 times (grindWeight) and every move into another zone counted 7.5 minutes more (zoneChange, zoneChanges in results). So a route grinds only where no quest is nearly as good, takes a hub's quests, does them in one go and hands them in, and switches zones only when it is clearly worth it. With neither, the search ground wherever the next quests were barely better (Orc consensus: 46 h of grinding in 129 h) or flew across continents for a single quest (Orc at 40-60: 32 visits in 41-50 and 62 in 51-60, 58 hearthstone uses). The values come from a sweep on the Orc and Human searches (600k routes each): a stronger zone weight brought grinding back (Human 23 h at 600 s with weight 2), a stronger grind weight brought the single-quest trips back; 450 s and 2.5 gave Orc 121.5 h with 2.9 h of grinding and 22 visits for one quest or none, Human 120.7 h with 10.2 h and 23.
• Quest log limit 40 (Forever's client: MAXIMUM_NUM_QUESTS_LOG_CAN_ACCEPT = 40 for its Camelot game type, Blizzard_APIDocumentationGenerated/QuestConstantsDocumentation.lua; the quest log shows N/40; issue #30): a package that needs more open quests is split. It is a world option (maxOpen, recorded in the world file; a world with another limit is rebuilt). The results up to 0.2.5 were searched at 20. The evaluator holds the log to it across the whole route: an accept with the log full waits (deferred like an accept above the character's level: taken in the same pass once a turn-in frees a slot, else in the package's second pass or a later return; still full there, it is skipped and counted). A stopgap until sim v2 plans the log: when the log is full, the character abandons the lowest-level quest in it at least 6 levels below it (ABANDON_GAP, evaluate.js; its reward is already reduced), losing that quest's progress, and waits only when none is that low. Without it, quests a route accepts and never hands in (65 in the 200k-route Orc time route) filled the log for good: today's best routes went from 123.5 to 173.3 h. Results report logPeak (at most the limit), logDeferred (accepts that waited), logSkipped, logAbandoned and acceptedNeverTurnedIn (accepted at some point, not turned in by 60: wasted accepts).
• Every parameter has a prior (mean and spread). Calibration (§4) narrows them.

Objectives (--objective time|smooth)#

What the search minimises is set by the policy parameters above; the objective chooses them (params.js OBJECTIVES, one place). Reported times stay real under either.

• zoneReturn: seconds per move back into a zone the route was in earlier and left (the first entry is free).
• spotReturn: seconds per return to a spot already worked: work (accept, objective, turn-in) in a 250 yd cell (SPOT_CELL, run.js) where work was done in an earlier stretch. A stretch ends when the character moves more than 600 yd (SPOT_LEAVE, evaluate.js) from where it began.
• grindWeight 4 under smooth: grinding never moves, so with the return penalties alone (2.5) the search ground instead of questing. Horde Orc, 200,000 routes, seed 1: time 123.5 h to 60 with 9.3 h of grinding, 105 zone changes, 73 zone returns, 507 spot returns; smooth at 2.5: 131.3 h, 52.4 h, 38, 8, 219; at 4: 127.4 h, 28.8 h, 57, 25, 365; at 6: 126.5 h, 16.3 h, 85, 50, 433.
• wastedAccept: seconds per quest accepted and never handed in (given up for a full log, or still in the log at 60). Taking a quest only to drop it is not smooth: Horde Orc, 100,000 routes with walks, the smooth objective without it gave up 42 quests and left 63 never handed in.

Smooth results are their own files (<faction>-<race>-<class>-smooth.json, consensus-<faction>-smooth.json, compare-…-smooth.json), so the two objectives never overwrite each other. Results, sim compare rows and the sim run report carry the smoothness counters next to the hours: zoneChanges, zoneReturns, spotReturns, sharedKills and the quest log counters above.

--param key=value (repeatable) sets any parameter of params.js after the objective's preset, fixed at that value (also in robustness draws); an unknown key is an error. A variant gets its own result files only with --tag <name>: sim run --objective smooth --param grindWeight=6 --tag g6 writes horde-orc-any-smooth-g6.json, and robust, compare and consensus with the same --tag read and write that variant. Results record overrides and tag.

3. Route evaluation and search#

• A route is an ordered list of packages starting at the race's start zone at level 1. evaluate(route, params) -> { seconds, levelAt[t], timeTo[10..60], xpSplit {quest, kill, grind}, travelSeconds, perPackage[] }, deterministic for given params, fast (target ≤ 1 ms per full route) so millions run in minutes across 16 cores.
• Feasibility: prerequisites done earlier (or inside the same package), level ≥ minLevel on arrival (else grind first, counted). A package runs up to three times: a return does only what opened since, so quest chains that go to another zone and back are kept (Scholomance and the Argent Dawn turn-ins, Loramus in Azshara); home packages run once. A package runs in its tour order (by place); an accept still too low for the character is taken as soon as the level allows, within the same walk, and does right away the objectives and turn-in the tour has already passed, and any passed accept its turn-in opens (Lazy Peons, then Thazz'ril's Pick at the same foreman). A second pass does only what is still too low after the whole tour (grinding first).
• Search, per faction and start race (class factor as a parameter):
  1. randomized greedy construction (next package chosen with probability ∝ exp(XP per hour of package incl. travel to it / τ); a package already run is a candidate again and scores only what opened since);
  2. local search on the best routes (swap, insert, drop, add, replace, merge; add and replace also bring returns) and simulated annealing;
  3. restarts until the best route stops improving over the last N% of samples (report the convergence curve);
  4. at least 10^6 evaluated routes per faction by default (--paths).
• Robustnessdraw K parameter sets from the (calibrated) priors (default K = 2000), re-evaluate the top routes under each, and report each route's win rate, mean and 90% interval of time to 60, and which parameters flip the ranking.

4. Calibration#

• From quest data: quests.json rewards against the QuestXP-table estimate (questEstScale) and Forever-observed rewards from the installed ForeverDungeonJournal (--fdj <dir>; questXpScale).
• From recordings (data/recordings): kill rates (killSec1, killSec60).
• extraKill (incidental kill XP, the main unobserved XP source) is a fixed prior in params.js (0.043 ± 0.038, fitted 2026-10-12 against level checkpoints).

Only the fitted parameter values are kept (data/sim/params.json).

5. Baselines (scored under the same model)#

• Polestar: the compiled guides (default route and every route choice), against the best route found (solo and with dungeons).
• Guides are scored on what they require.A guide's optional goals are left out: an optional quest (its accept is optional) whole, any other optional goal by itself. Most are quests started by rare drops (the OOX beacons, A Strange Red Key, the ritual totems at about 3%), the Green Hills of Stranglethorn pages, profession and bag quests and rares: counting them as done assumed every player gets every rare drop. Until 0.2.4 they were counted, which made the guides look about 8 h (Horde) and 5 h (Alliance) faster than what a player following them can expect; joelwurtz found it (issue #30). sim compare --with-optional (and sim bench --with-optional) scores them as done, the old way, to show what the extras can save. Elite and dungeon quests are outside the solo world, so they stay out either way. Search results (best found, consensus) are not guides and are unchanged.
• Guides are scored at the giver and ender they name.A compiled accept or turn-in carries its NPC or object (accept 8254 npc 11406, or the one the compiler placed it at), and a quest with several is scored there, not at the world's first. Guides used to be scored at the first, a trip the player never makes: Horde guides were charged the walk to the Steamwheedle Port poster for the two Tanaris wanted quests they take in the Gadgetzan square, and class guides the trip to another city's trainer (Rohan for Cenarion Aid was scored by moving him in the data). About 0.1 h to 60 for most starts and up to 0.9 h for Horde Warriors.
• A pack (sim compare --pack <dir>, or the pack folder as the one argument): its chain, after our chain up to the pack's first level when it starts above 1, under the same model (data/bench/compare-<faction>-<race>-<pack id>.json).

6. CLI#

sim world   [--faction F] [--rebuild]               build/cache the world model (data/sim/world-<faction>.json)
sim walks   [--faction F] [--radius Y] [--workers W]  route the world's walks with the walking router
            (data/sim/walks-<faction>.json; run, robust, compare, bench and consensus use them)
sim calibrate [--fdj DIR]                            fit parameters to quest data and recordings
sim run     --faction F [--race R] [--class C] [--paths N] [--workers W] [--seed S] [--dungeons]
            [--jobs J] [--anneal A] [--max-paths M] [--objective time|smooth] [--param K=V ...] [--tag T]
sim robust  --faction F [--race R] [--class C] [--draws K] [--top N] [--objective O]   robustness of the top routes
sim compare --faction F [--race R] [--class C] [--guides DIR] [--pack DIR | DIR] [--to 10,20,30] [--with-optional]
            [--objective O]   Polestar's routes vs the best found, same model
sim bench   --faction F [--race R] [--class C]      bracket table: zones, levels, quests per 10 levels
sim export  --result <file> --out <dir>             zone-by-zone skeleton guides + visits.json for authors
sim consensus --faction F [--dungeons] [--evals N] [--objective O]   one 12-60 route shared by the faction's starts (+ exports)
(robust, compare and consensus take --param and --tag too)

Less common options: --fdj <dir> (calibrate: the ForeverDungeonJournal folder, default the installed one), --band <levels> (level width of the world's quest packages, default 5; a new width rebuilds the world), --jobs <n> (run: search jobs per round, default 32), --anneal <n> (run: annealing evaluations per job, default 6000), --max-paths <n> (run: hard cap on evaluated routes, default 3x --paths), --guides <dir> (compare: the guides folder to score, default guides/), --pack <dir> or a pack folder as the one argument (compare: also score a pack's chain, after our chain up to its first level when it starts above 1; written to data/bench/compare-<faction>-<race>-<pack id>.json), --to <levels> (compare: levels to show, default 10,20,30,40,50,60), --top <n> (robust: routes re-evaluated), --dungeons (run, robust and consensus: the dungeon-inclusive family, needs data/forever/dungeons.json), --offline (never download the QuestXP table), --radius <yards> (walks: route every pair of points up to this far apart, default 4000), --no-walks (walk the straight line × detour even with a walks file), --quiet (less output); --data <dir> and --sim <dir> move the reference data and simulator outputs (default data/forever, data/sim).

7. Consensus route#

sim consensus --faction F builds one route from level 12 to 60 that every start of the faction can follow: each start keeps its own best route up to its level-12 point (and all its home packages); the shared part is one package sequence run in order by every start (a package a start already ran or does not have is skipped for it; no deferral).

Outputs: data/sim/consensus-<faction>.json and .md, and per start an export of the consensus route under data/sim/export/consensus-<faction>-<race>/. Export place names (hub, grind near) are the nearest named map area (subzone or town, client WorldMapOverlay/AreaTable) on the point's own zone map.

8. Results: world f7bd5e2cccab (final run, 2026-10-13)#

Every number below comes from one world: data hash f7bd5e2cccab91eb0a8f2bd4b4b673b58eb8af13, both factions, builder 18 (builder 17 plus friendly kill credit for find/heal/use/locate objectives, escorts named "escape", counted events as use, quests started from an item another quest rewards), with the route policies of §2 (grinding counted 2.5 times, 7.5 minutes per zone change), run from a frozen copy of the repository. Results record the hash; anything from another world (bde7513ac7b8 and earlier) is refused. The Forever rank corrections that came after the run (22 NPCs, see Known limitations) are in the current data but not in this world.

These are model estimates in /played hours, not measurements: no recordings exist yet, so time parameters keep their priors (only quest XP is calibrated; extraKill is a fixed prior, §4), and Forever's 30-60 content is not in the data. Rankings under the model are what the numbers support; absolute hours are not.

Solo filter counts#

Per start (hours to 60, solo)#

Guides: each race's default chain as written (sim compare --guides), counting their optional goals as done (the scoring until 0.2.4, §5), grinding at their own level steps. The leveling guides were rebuilt from this run's consensus exports with a flow pass the model does not make: side trips for one or two small quests behind a flight, boat, hearth or long ride were dropped, every hub is one loop, and where the level then falls short a level step names a camp (with an optional dungeon past 20 minutes at 40+). That costs 2.0-2.7 h (Horde) and 9.1-9.8 h (Alliance) against the consensus, almost all of it grinding: the Alliance 44-50 has few solo quests left once the Booty Bay, Shimmering Flats and OOX trips are gone (the longest stop is 48 to 50 at the Un'Goro gorillas). The previous guides (written against builder 15, scored on this world) are 3.8-5.5 h slower with 29-34 h more grinding.

Consensus time to 10/20/30/40/50/60 (h):

On world bde7513a (builder 17, grinding and zone changes at face value) the consensus was 128.7-128.9 h (Horde) and 129.7-130.5 h (Alliance) with about 46 h of grinding: returns to a zone and the grind policy save 8-11 h.

Consensus and convergence#

• Horde: 173 shared packages from level 12; bound (Skyborne's own 12-60) 542.7 summed policy cost, consensus 542.6. Alliance: 169 packages; bound (Dwarf's own) 591.9, consensus 591.9. 1.2M evaluations each.
• Searches: 3.46-9.03M routes per start. Human and Skyborne H reached the route cap (3x the 3M paths) still improving (last improvement at 99% and 96% of the evaluations); the others converged (last improvement at 72-98%).

Robustness (2000 parameter draws, top 10 routes per start)#

Costs include the route policies (hours plus the weighted grinding and zone changes), so they sit above the hours. The top routes are close: the best found route is within 1% of each draw's best in 30-99% of draws, and no route wins a majority.

Dungeon family#

Horde only (the Skyborne and Alliance dungeon searches were stopped): Orc 120.0, Tauren 120.4, Undead 121.8 h, never ahead of the solo best found. Under the model dungeons do not beat questing solo.

Second search (verification)#

The same pipeline run again to test whether the first search's routes are as good as the method gets: another random seed (2), up to twice the routes (--paths 6000000, cap 18M), and the current data with the 22 rank corrections, world 165d88c5eaaa. Best found / consensus, hours to 60:

• Searches: 6.54-18.07M routes per start; Horde Tauren, Undead and Skyborne and Alliance Skyborne reached the cap still improving (last improvement at 95-100% of the evaluations).
• Consensus: Horde 167 shared packages, 531.4 summed policy cost (bound 531.6); Alliance 169 packages, 586.9 (bound 587.3).
• Robustness: the best found route stays within 1% of each draw's best in 48-100% of draws; win rates 9-78%.
• The two searches' consensus routes differ by at most 2.8 h (Horde) and 1.0 h (Alliance), every start within 117.3-121.2 h: the top of the search space is flat, and more search still finds small gains. The guides score the same on both worlds (within 0.1 h); against the second consensus they cost 5.1-5.6 h (Horde) and 9.9-10.5 h (Alliance), almost all of it the grinding the flow rules add.

Third and fourth searches (0.2.0)#

Two more independent searches checked the routes again before 0.2.0, with the guides as they ship scored on the fourth world:

• Third: seed 3, --paths 10000000 (cap 30M), world 5a8503dbb43a; 17.7-30.2M routes per start, seven of eight reached the cap still improving.
• Fourth: seed 4, --paths 6000000 (cap 18M), world 329281f957a5, the data 0.2.0 ships with; every start reached the cap still improving.

Guide columns count the optional goals as done (the scoring until 0.2.4; the hours for what the guides require are in the next section).

• Robustness: the best found route stays within 1% of each draw's best in 87-100% (third) and 76-100% (fourth) of draws.
• Across the four searches every start's best found route lies within 116.3-121.2 h and every consensus within 116.3-120.6 h.
• Against the fourth consensus the guides cost 6.4-6.9 h (Horde) and 9.3-10.7 h (Alliance). The Horde guides lost 0.7-0.8 h against 0.1.0 by dropping quests a player cannot take (a chain started by a dungeon drop, quests removed from the game, a chain without its first quest) and moving the level-35 Dustwallow quests from 30 to 39; the Alliance guides gained about 0.5 h from a northern Stranglethorn loop and the Shimmering Flats.
• The third world predates the data correction that places Sulhasa in the southern Barrens, so it charges the Horde guides a trip to Un'Goro Crater for Field to Clear (about 0.9 h); the fourth world has the correction.

Guides scored on what they require (0.2.4)#

Since 0.2.4 the guides are scored without their optional goals (§5, issue #30). On world 5eef72140b61 (the data 0.2.4 ships with), each race's default chain (sim compare, and --with-optional for the last column; --guides with the guides of tag v0.1.0 for the middle one):

• The optional extras are worth 7.8-8.3 h (Horde) and 4.5-5.0 h (Alliance) to a player who gets every one; almost all of it comes back as grinding (about 20 h more for the Horde, 11 h for the Alliance) where the extras' experience is missing. The largest for the Orc chain: The Green Hills of Stranglethorn and its chapters (2.4 h together), A Strange Red Key (0.9 h), the OOX-09 and OOX-17 chains (0.9 and 0.8 h), Winterfall Firewater (0.8 h).
• Against the fourth consensus (its own world) the guides cost 13.3-14.1 h (Horde) and 13.6-15.1 h (Alliance).
• With the optionals the guides of 0.2.4 and 0.1.0 are within 0.6 h (Horde); without them 0.1.0 is 1.8-2.6 h faster for the Horde, because the later guides marked more of their quests optional. The Alliance guides of 0.2.4 are 0.4-1.0 h slower without and 0.8-1.6 h faster with.
• The search's best and consensus routes were searched on earlier worlds, so the README's charts show only the guides.

Known limitations#

9. Refresh (launch day)#

Results are labelled "Era + known Forever quests": Forever's 30-60 content (Riverglades, Shen'dralas, Mount Hyjal, the rest of the Wetlands expansion) is not in the data yet (the beta caps at 30). When it lands:

node tools/polestar.js data build
node tools/polestar.js sim world --rebuild
node tools/polestar.js sim calibrate
node tools/polestar.js sim run --faction horde
node tools/polestar.js sim run --faction alliance
node tools/polestar.js sim run --faction horde --dungeons && node tools/polestar.js sim run --faction alliance --dungeons
node tools/polestar.js sim robust --faction horde && node tools/polestar.js sim robust --faction alliance
node tools/polestar.js sim consensus --faction horde && node tools/polestar.js sim consensus --faction alliance

then re-export (sim export) the starts the guides follow.

Keeping the guides' fixes#

The guides are hand-written from the exports, and a refresh re-applies a new route to them. It keeps every fix a guide's header documents (its "Order changes against the export", "Deviations from the export" or "Local order fixed by hand" list): those orders answer player reports and things the model does not see, such as a cave's mouth, a quest item that vanishes after a while, or a delivery that belongs on the way out of a starting glen. A refresh that changes such a stretch changes the header too, and says why the new order still holds.

Two checks enforce it:

• tests/guides-reported.test.js (and tests/guides-caves.test.js for cave mouths) pins each player-reported fix on the compiled guide: goal order and where the arrow leads, one case per report. A refresh that undoes one fails the suite.
• Lint warns when a guide carries a timed quest item (items.json duration, the Forever client's ItemSparse DurationInInventory) longer than it lasts: from the goal that gets it (an accept that hands it over, a collect, a buy) to each later goal that spends it (a turnin that needs it, a use), timed from the latest goal that got more of it. The time is this model's (§2, at the priors): the straight line × the detour factor at run (or mounted) speed, flights from travel.json plus flightOverhead, the hearthstone cast, talkSec and interactSec, and kill time with the walk between mobs of the goal's spawn areas for kills and drops; other characters' and optional goals, grinding and legs across continents are left out, so the estimate errs short (tools/lib/dsl/timed.js).

When a guide is wrong because the data is (a giver, a spawn, a drop chance), the fix belongs in tools/lib/data/corrections.json as well as in the guide: the simulator builds its world from the data, so only a corrected record keeps the next search, export and refresh from reproducing the mistake.