ESODecoded

BuildNet — a neural net trained on real ESO parses

We trained a chess-engine-style value network on 15,236 real builds from the ESO Logs API — every player's actual sets, slotted abilities, class and the DPS they did. It predicts a build's top-parse ceiling (held-out Spearman 0.76), and it learned the meta from data nobody hand-wrote. Trained on Update 50 parses, 2026-07-23.

Analyze a real log

Paste an ESO Logs report URL (optionally with #fight=N; otherwise we take the last boss kill). The net scores the build and searches every supported single swap:

The learned set tier list (5-piece, DPS)

Ablation attribution on held-out builds: how much the net thinks each worn 5-piece is worth vs the average 5-piece in top parses. Negative = costs personal DPS (support sets — working as intended).

Namevs category averagebuilds seen
Corpseburster+17.8%375
Azureblight Reaper+17.0%98
Savage Werewolf+13.6%119
Perfected Kazpian's Cruel Signet+9.7%79
Aegis Caller+5.3%151
Deadly Strike+0.0%146
Perfected Mora Scribe's Thesis-0.4%58
Spattering Disjunction-1.0%282
Aetheric Lancer-4.4%42
Perfected Ansuul's Torment-4.5%92
Tide-Born Wildstalker-5.1%84
Perfected Slivers of the Null Arca-6.3%957
Z'en's Redress-11.4%110
Roar of Alkosh-12.5%155
War Machine-15.3%63

The learned ability values

Same method, per slotted ability. Class-signature skills carry class identity in these numbers — read them as "what builds slotting this achieve", not pure skill power.

Namevs category averagebuilds seen
Flame Lash+29.8%83
Winter's Revenge+29.8%47
Ritual of Retribution+24.5%74
Blighted Blastbones+22.6%393
Burning Talons+19.1%45
Power of the Light+15.7%81
Pummeling Goliath+15.5%50
Exhausting Fatecarver+15.0%88
Subterranean Assault+14.7%46
Vampire's Bane+11.6%58
Cephaliarch's Flail+11.4%617
Cruxweaver Armor+11.2%259
Werewolf Berserker+9.6%144
Relentless Focus+9.6%165
Killer's Blade+8.7%229
Volatile Armor+8.1%134
Camouflaged Hunter+8.0%758
Debilitate+6.4%141
Mystic Orb+6.0%46
Anti-Cavalry Caltrops+5.8%359
Unnerving Boneyard+5.7%249
Haunting Curse+5.7%62
Arctic Blast+5.6%49
Shadowy Disguise+5.5%210
Bound Armaments+5.4%60
Ulfsild's Contingency+5.0%260
Igneous Weapons+4.9%121
Elemental Rage+4.7%59
Banner Bearer+4.4%502
Shooting Star+3.9%278
Quick Cloak+3.2%795
Noxious Breath+3.1%97
Solar Barrage+3.0%174
Power Extraction+2.9%46
Barbed Trap+2.8%888
Flames of Oblivion+1.1%145
Incapacitating Strike+1.0%170
Elemental Susceptibility+1.0%46
Detonating Siphon+0.8%399
Venom Skull+0.7%295
Soul Harvest+0.3%121
Scalding Rune-0.1%76
Ghostly Embrace-0.8%51
Twisting Path-1.0%48
Fulminating Rune-1.7%494
Shattering Knife-1.7%117
Pestilent Colossus-2.2%46
Pragmatic Fatecarver-2.5%519
Blazing Spear-2.9%77
Flawless Dawnbreaker-3.9%1,002
Carve-4.1%416
Dark Shade-6.2%198
Skeletal Arcanist-6.3%87
Stampede-6.6%1,185
Inspired Scholarship-7.1%599
Recuperative Treatise-7.3%49
The Languid Eye-7.6%548
Avid Boneyard-7.8%146
Radiant Glory-8.1%154
Venomous Claw-8.5%113
Liquid Lightning-9.1%60
Blue Betty-9.2%47
Surprise Attack-9.3%182
Hurricane-9.6%77
Crystal Fragments-10.2%63
Stone Giant-10.4%88
Skeletal Archer-10.8%264
Glacial Colossus-11.0%158
Biting Jabs-11.1%76
Echoing Vigor-12.2%167
Whirling Blades-12.3%40
Siphoning Attacks-13.2%88
Merciless Resolve-14.3%122
Standard of Might-19.1%104
Resolving Vigor-21.4%109
Elemental Blockade-24.2%313
Wild Guardian-35.0%44
The Tide King's Gaze-52.7%40

FAQ

What is this, exactly?

A small neural network (an NNUE-style accumulator net, the architecture family chess engines use for evaluation) trained on thousands of real parses from the ESO Logs API: each training example is a player's actual sets, slotted abilities, class and boss — and the DPS they actually did. It learned to predict DPS from build alone.

How accurate is it?

On held-out fights it never saw in training, its DPS ranking of players correlates at Spearman 0.76 with reality, with roughly 30% median error. Build alone — no Champion Points, no food, no player-skill rating — carries that much signal. All numbers are from the report-level split so no fight leaks between training and evaluation.

Why do support sets show negative numbers?

Because they cost personal DPS — that's what they're for. Sets like Roar of Alkosh exist to buff the group, and the net measured the personal price of wearing them. That number was folk knowledge for years; this is it measured from data.

Should I regear based on this?

Treat it like a very well-read friend, not an oracle. The numbers are correlational — they say what top parses run, at what results. They're blind to CP, mundus, food, rotation and your content (a trial dummy ceiling is not a vet-dungeon pull). Verify in game.