Typed Tibia-rules simulator targeting WGSL for the OT Agent Arena benchmark

A typed Tibia-rules simulator in Spiral, compiled to C, Rust and WGSL (GPU) from one source, used to train agents that are measured against the best open-source Tibia bots on a public, reproducible benchmark: the OT Agent Arena.

27 .spi modules in src (plus src/standards.spi, the package's standards checked by eoie standards), 15 files in apps/arena (9 Lua probes and their shared client modules) and the server-map tool in apps/map.

• Rules: geometry, vitality, coins, rng, cooldowns, damage, combat, creatures, experience, skills, blessings, wire.

- **Typed domain:** quantity (phantom units), readiness (GADT witnesses), vocations (vocation-indexed
spells), catalog (existential spawns), views (HKT`mapped` and`counted` ), terrain (layouts, incl. the scanned arena spot).
- **Decisions:** policy, network (integer neural policy), search (evolution).

• Simulators and entries: hunt with its hunt_sim tests, hunting_pack (the multi-creature simulator) with its pack_sim and gpu_sim tests, andlive.spi (the Lua entry for OTClient).

Tests name their targets with a reason: C as the oracle, WGSL for the GPU path, Lua for the live client, and Rust only where a test needs the training host.

Correction 2026-10-14: every live spiral-* run from 2026-10-13 11:48 until the fix ran with vBot's HealBot and TargetBot active alongside our policy (a bot setting the vBot probe left enabled; BENCHMARK.md, PR-1). The findings below come from the re-measurement on the fixed arena (PR-4, 30 runs per competitor, every run checked for foreign actions); the earlier live claims are withdrawn.

1. The simulator predicts live clear times and, mostly, potion use. Clear times matched in every live check. Potions took four corrections, each measured live or read from Canary's source: login protection keeps monsters off a fresh knight for 10 s; a simulator bug fielded three cyclops in small camps; monsters surround the knight instead of queueing (occupancy-aware routing, PR-7 holds out of sample); and monsters wander while the knight is protected, so every clear starts from a different geometry (before, every seed replayed one), with the cyclops swing pace measured at 2166 ms (mirroring_the_live_arena ). Potions per unit of damage were always right; the misses were in hits taken. Against pooled live means the model now sits within the noise for rules and retreat at two and three cyclops (in-sample). Its out-of-sample check failed in one direction (PR-10): it predicted that preferring a monster already in reach costs 0.8 potions, and live it costs nothing (12.3 vs 12.5 for rules). Live, our policies drink alike. Since 2026-10-20 retreat steps cancel the pending swing as every walk does in Canary and the knight's think runs at the measured ~1083 ms (PC-6 confirmed both on fresh logs); with that the model gives ReachFirst 12.47 against rules 12.30, close to live, though that comparison is in-sample. Where nothing protects the fresh login (every

arena since 2026-10-15), the model still predicts too many potions: rules 15.3 vs 14.4 (PR-11), retreat about 14.5 vs
12.8 (PR-12), the hunting ground 15.5 vs 12.7 (PR-15, whose kill pace it predicts well: 12.4 vs 12.8 kills). The

retreat gap traces to crowding: the model's knight is pinned in the spot's dead end with three cyclopes, the live one keeps stepping (4.0 vs 11.5 retreat steps per clear). Update 2026-10-21: the arena bundlemirroring_the_current_arena (Canary's knight step speed, dance step and step formula on the map's ground speed, with the knight acting on the client's last view) closes the cyclops gap out of sample. PR-17, 40 fresh runs: rules 13.90 live against 14.55 predicted (t -0.87), retreat 13.25 against 12.81 (t 0.74); retreat steps 12.4 against 11.7 live. Since 2026-10-22 the bundle also carries two hunter-arena facts, each adopted by a rule written before measuring: three Canary timing facts (all3; PR-18 holds at 14 of 20 clears, where the bundle predicted 99 of 100), and arrows that stop only at projectile blockers, which lift predicted hunter hits from 14.9 to 21.0 per run (PR-19, 20 fresh runs: 20.45 live, t -0.40, and 15 of 20 clears; all3 alone rejects at t 3.86). The hunting ground's potion gap closed as noise on fresh hunts (PR-23, 2026-10-22: 14.8 live against 14.84 predicted at shield 11, t -0.05; PR-15's 12.7 was a low draw). Still open: the hunter clear time (55.9 s vs 60.5 s in PR-19), where live hunters walk faster at two to three tiles from the knight than the model's (about 1 tile per second against 0.44-0.66). A hunter run is decided by its first chase: live, the first target leads the knight 20 or more tiles away in about half the runs, and the run fails when the other hunters lose contact on such a chase; in the model such long chases are rarer (20%). Seven Canary movement facts were tested against this on 2026-10-21 and none passed in sample (TIBIA-FIDELITY). Since then, the model's long chases were traced to one path: once the knight lines up on a hunter's row, the hunter runs straight along it, the same way in every seed, while live hunters break off and still stay ahead (half of the live long chases end south, none of the model's do). Letting the model's hunters break off at random makes the knight catch every one, so the missing piece is what keeps a live hunter ahead after it turns. The likeliest answer is on the knight's side: when its target steps away from it, the live knight is still on its tile a second later in 98% of cases (the model's in 57-68%). Our competitor turns chase mode off while its target is adjacent, so Canary re-plans the knight's path only at the player's next one-second think and then waits a full step (TIBIA-FIDELITY note 85; a two-second follow timer, rule D30, was rejected). Modeled (rule D32), it reproduces the stalls (91% against 98% live) but the model knight then clears only 53 of 100 hunter runs against 72-95% live, so it stays off and the clear rate is the open question (notes 99-101). Retreat steps (15.1 vs 11.3 per clear) close when the knight's retreat steps follow the live competitor's own 650 ms pace: 11.7 predicted, 10.1 on 20 fresh runs (PR-20, t -0.54; potions 12.80 vs 13.32), adopted with the rest on 2026-10-22. The step count's spread is wide, so PR-20 could not reject the old 15.1 either, and its shape still differs: live runs are first pinned in the pocket at 1936,1349 after about 5 steps, the model's knight further north at 1937,1342 after about 8. The difference starts before any retreat: live, the cyclopses move off their summon tiles before the knight acts; the model starts them frozen there, so its opening is the same every clear. 2. The policy designed in the simulator does not win live. Stepping back when two cyclops reach the knight saved about 4 potions per clear in the first simulator; with monsters surrounding the knight as they do live, the saving shrinks to about 1 potion (5%), and live it uses 1.4 more than the plain rules (t 1.8, not significant). The earlier live win came from runs with vBot healing alongside. 3. Our rules policy is level with vBot. With enough runs (PR-9, 50 each, pre-registered) the Spiral rules policy drinks 12.5 potions per cyclops clear against vBot's 14.7 (difference -2.2, t -3.8), but the whole difference sits in the first 10 s: Canary ends a fresh login's protection when the player moves, turns, talks or uses an item, which vBot does and our probe does not, so vBot takes 9.1 hits in that window against our 1.6. After

it the two take the same hits (60.1 vs 60.4). Re-tested with the protection ended for both (PR-11, 50 each): 14.4 vs
14.8 potions, no difference shown (t -0.6), identical hits (69.1 vs 69.2). Our three policies drink alike live (rules

12.5, ReachFirst 12.3, retreat 12.2). Against nExBot the retreat policy was narrowly better (12.2 vs 14.6 potions, PR-3a, t -2.14), but about half of that is the same login rule (nExBot talks right after login). The fair re-test (PR-12, 40 each, protection ended and both starting 2.5 s after login) shows no difference: 12.75 vs 13.93 potions, difference -1.18 (interval -2.42 .. 0.07, t -1.88). nExBot runs only with three arena shims (BENCHMARK.md PR-3a). 4. Learning wins little at this level of detail. In every scenario tried (open field, corridor, room, arena spot, mixed cyclops and trolls, potion budgets, respawning hunts) tuned threshold rules match or beat neural and scored learned policies. On a respawning hunting ground the rule that sweeps to the next spawn due beats every fixed route at a 60 s respawn, the evolved one included, but not at 30 s. With more monsters than the knight can fight at once, the tuned sweeping cavebot leads and the retreat rules trade experience for cost, and two knights coordinate best by simply following one shared rule together. That party finding rests on a duo model the first live duo hunts did not confirm (PR-24, 2026-10-22: 5.1 party potions against 7.56 predicted, 16.6 kills against 15.36): the model sent every cyclops at the first knight, while live the hits split about evenly. Canary's target order (the last knight to log in first, later arrivals in map-scan order) fixes the potions and the split in the model; it still kills about one troll per hunt too few, so the party sweep is rerun only after that.

Every mechanic comes from Canary's source or a live measurement, and the model is checked against both:

Live-mirroring model (2026-10-15, first swing folded in 2026-10-18): mirroring_the_live_arena turns on the five rules above that only matter live (wake, idle walk, swing period, first swing after targeting, swing after a step), and live_retreat_tactic steps exactly like the live probe. Before it, every seed replayed one geometry (fixed spawns, monsters frozen while the knight is protected), and the simulated knight escaped along the corridor that the live knight rarely reaches; live, the retreating knight sits cornered in the dead-end pocket at 1936,1349 for 43% of the clear. Healing was never the gap: potions per 1000 hp of damage are about 5.8 live and simulated. Against pooled live means (in-sample, these data shaped the rules): 3 cyclops rules 12.30 vs 12.22 potions, ReachFirst 12.47 vs 12.30, retreat 11.76 vs 12.62; 2 cyclops rules 4.72 vs 4.85, retreat 4.91 vs 5.45 (with the 2026-10-20 swing fixes; before them 12.51, 13.22, 11.46, 4.65 and 4.59). Scored against the six live means with their standard errors, the sum of squared t fell from 17.5 to 8.8 with the fixes, none beyond 2 (before: both retreat means). Folding in the first swing moved the simulated clear times by at most 0.6 s. The 3-cyclops ranking of retreat vs rules is still off by about 0.9 potions (model: retreat 0.54 better; live: 0.40 worse), the crowding at the pocket (BENCHMARK, PR-12 post-hoc). Under this model a lower heal threshold or a better standing tile saves at most 0.3 potions per clear, so no live-runnable change to the rules policy is predicted to show live. The 10 s wake and the idle walk describe the arena before 2026-10-15 14:17, when our probe kept the login protection. Since then the parity layer ends the protection for every competitor at login, so the matching scenario is mirroring_the_live_arena |> waking_at 0 (no wake, so no idle walk); its first prediction (rules 15.6 potions, 72 hits) ran about 5% high (PR-11: 14.4, 69). Since 2026-10-20 that scenario has its own bundle, mirroring_the_current_arena: on top of it, monsters re-path only on their think (Canary Creature::onThink, following_on_think) and swing only on their attack rounds, losing a round spent out of reach (Monster::canUseSpell, meleeing_on_canary_rounds). Chosen overnight by a rule written before measuring, it cut the misfit over PR-11, PR-12 and PR-15 from a sum of squared t of 36.9 to 12.6 (rules 14.63 vs 14.42 live). On 40 fresh runs (PR-16, 2026-10-21) it holds for the rules policy (14.35 live vs 14.63, t -0.43) but not for the retreat policy (12.45 live vs 13.74, t -2.39): the model's retreat steps more and strikes less than the live one (94.7 s vs 87.4 s per clear), the same bias it showed in sample. The cause was the knight's fixed 650 ms step; Canary steps a level-30 knight in 500 ms on arena ground, and the bundle now adds paced_like_canary (the knight steps by creatures.player_step_ms), which brings the retreat to 12.53 (in sample, rules unchanged) and hunting-ground kills to 12.73 per hunt (live 12.80). It also adds Canary's dance step (dancing_like_canary: an adjacent monster with no path step sidesteps 10% of its thinks, staticAttackChance 90), which moves retreat to 12.87 (the hunter clear rate went back inside its band at that step, before the last-tick view pushed it to 99 of 100). Two more server-and-client facts close the retreat step gap (20.6 steps per clear down to 12.4, live 11.7): monsters and the knight step by Canary's formula on the map's own ground speed (stepping_by_canary_formula; the scaled grass step had turned a cyclops's 700 ms into 800 on the 100 ms tick), and the knight acts on what the client saw a tick earlier (seeing_a_tick_late), so a cyclops can take the last free tile beside it first. All of this was fitted in sample overnight under rules written before each measurement (BENCHMARK, PR-16 rows), and it holds on fresh clears: PR-17 (2026-10-21, 40 runs) gave rules 13.90 against 14.55 (t -0.87) and retreat 13.25 against 12.81 (t 0.74).

The multi-creature simulator is bit-exact on WGSL too, checked since 2026-10-20 when Cells began honoring a WGSL test's status (before, every GPU test passed whatever it returned; all 28 GPU twins pass under the fix) (gpu_sim.spi: a full troll hunt, five arena-spot cyclops clears, and a five-minute hunting-ground hunt with the shortest-path router and its scratch arrays). With parallel lanes (one seed per GPU thread, 2026-10-14) every lane matches C exactly; a launch costs about 4 s of fixed setup, then about 0.23 ms per arena clear against 1.6 ms on one C core, so the GPU pays off from about 4000 clears per launch. Large lane counts now run in chunks of 1024 (the 16384-lane hang, WGSL-LANES-LARGE-DISPATCH-HANG, is fixed in the pack-1e compiler).

Protocol: everything is tuned on training seeds and judged on held-out seeds; the seed blocks are typed, so tuning on held-out seeds does not compile. Claims are Pareto dominance on (kills, profit) or a cost within a stated time cap. Profit books gold looted, 50 gp per potion and 14000 gp per death. The control is a random policy (each decision tick it drinks, engages, steps to a random tile or stands, with equal odds; once engaged the server keeps swinging, as live): on held-out seeds it earns 54310 experience on the hunting ground at -1.15 million gp, and clears the arena spot in 94 s with 72 potions per clear (one potion per second at most, as the server allows), far below every rule.

• Targeting (three-troll pack, 20 seeds): finishing the current target first gets 330 kills; fighting whatever is in reach gets 75 kills and 381 escapes. The live client switched to that order and killed all three trolls, where the old order killed one. A chasing policy gets 356 kills against 34 for a standing one (hunt_sim.spi ).
• Cyclops pack, open field (100 training and 100 held-out seeds): healing below 130 dies; 170 and above tie at 950 kills. Only the network warm-started from the cavebot tuned to heal below 160 dominates it, by 1550 gp at the same 955 kills. Kiting now walks the knight out of the cyclops' sight (monsters see 11 tiles), so they stop following and it kills only 300.
• Terrain is the lever. In a one-tile corridor the same cavebot clears three cyclops with 6.5 potions and no deaths, against 15.3 potions and 3 deaths in 20 clears in the open field. In a room with a corridor, rules evolved on 100 training seeds (ReachFirst, heal below 223 + 3 per attacker, step back at three attackers) dominate the fully tuned cavebot on 100 held-out seeds: 946 kills and -288217 gp against 942 and -477181.
• Arena spot (the server map, radius 45, monsters routing around each other): rules evolved for the cheapest clear within the tuned cavebot's time +15% choose ReachFirst targeting, heal below 88 + 26 per attacker and stepping back as soon as one monster reaches the knight. On 100 held-out clears they use 1070 potions against the tuned cavebot's 1702 (heal below 90), with one death against none, at 95.6 against 90.1 s per clear: cheaper, slower, not dominant. Before road speed applied at the spot (radius 30) the same search chose 96 + 8 per attacker for 1069 potions, no deaths, at 90.9 s.
• Mixed camp (one cyclops, two trolls, arena spot): the cavebot tuned on this camp (heal below 80) uses 85 potions in 100 held-out clears at 37.7 s. Rules evolved on this camp without knowing kinds use 74 potions at 37.6 s with no deaths, dominating it; kind-aware rules use 68 but die once. Rules evolved for three cyclops use 190 potions; counting only attackers that stand and fight (policy.temperate : trolls flee, cyclops do not) cuts them to 135.
• Hunting ground (two cyclops and four trolls respawning every 60 s, the server map around the arena, five-minute hunts, 100 held-out seeds): greedy one-tile steps stick behind walls there, so knight and monsters route with a shortest-path search (straight steps only: a diagonal costs three straight ones, and Canary's A* agrees), and every step costs what the server's ground under it says (roads, speed 100, are 1.6 times faster than the arena ground, speed 160; measured live in PC-1, confirmed by the pre-registered PC-2). With the cavebot's heal tuned on training seeds (120), standing at the start earns 104920 experience; a hand-drawn route 108270, a loop by the spawns 114470, a route evolved on training seeds with roads (a two-point shuttle between 121,86 and 117,102) 119280, and sweeping to the nearest live or next-due spawn 119950 (with the protection zones of the east buildings walled off since 2026-10-20). The rule that reads respawn timers leads at this 60 s respawn and ties at 120 s, but at 30 s the fixed routes beat it (hand route 133170, shuttle 133200, sweeper 126980): which walker is best depends on the respawn time.
• Six cyclops on the hunting ground (more than the knight can fight at once; 100 held-out hunts, each cavebot's heal tuned on training seeds): standing (heal below 100) earns 112700 with 2 deaths at -211081 gp; sweeping the spawns (heal below 130) earns the most, 126900, with 3 deaths at -296687 gp. Stepping back at two attackers while sweeping (the live retreat rules) earns 121050 at -166786 with no deaths: about 5% less experience for 44% lower cost, a choice of objective rather than a better policy.
• Two knights on the six-cyclops ground (100 held-out hunts, party experience): the second knight acts through the same code (the simulator swaps which knight is acting), and monsters chase and strike the nearer one. Both knights sweeping together earn 259350 at -180187 gp; splitting the ground into two territories earns less (248550) at 30% more cost (-233638), because each knight then faces the cyclops alone. Rules evolved for the duo (heal below 117 + 6 per attacker, never retreat) reach 261750 at -180890: within 1%, so one shared rule already captures the coordination.
• Learned policies: a 12-gene scored policy (heal, retreat, targeting weights, step weights) whose start genome reproduces the evolved rules found nothing better in 640 evaluations. With a 6-potion budget it clears 5.6% faster at the same deaths and potions, below the 5% cost bar. A respawning hunt with a potion budget does not separate policies either: above about 40 potions the budget never binds; below it the only decision that matters is when to leave, which a one-number rule captures.

PR-4 (2026-10-14, fixed arena): cyclops clear, BENCHMARK pack v1, full gear restored before every run, the same 250 ms action gate and 100 ms decision tick for every competitor, heal below 160, 30 runs each, alternating, every run cleared, no deaths, no arena failures. These runs predate the login-protection fix: vBot ended its protection at login and our probes did not, so the vBot row carries about 7 extra hits per clear (BENCHMARK.md PR-9 addendum; level on the fixed arena, PR-11).

Pre-registered tests: retreat vs vBot, no difference shown (t -1.45); retreat vs rules, no difference shown (t 1.82). Rules vs vBot, pre-registered as a fresh replication (PR-8, 30 runs each): 12.9 vs 14.2 potions, no difference shown (t -1.47); PR-4's 11.5 for the rules did not replicate. Rules vs vBot with 50 runs each (PR-9, 2026-10-15): 12.5 vs 14.7 potions, significant (t -3.80), clear time 88.7 vs 87.9 s, damage taken 2145 vs 2479; the whole difference comes from vBot ending its login protection in the first 10 s (BENCHMARK.md PR-9 addendum). With the protection ended for both (PR-11, 50 runs each): 14.4 vs 14.8 potions, no difference shown (t -0.58). Retreat vs nExBot 4.0.4 (PR-3a, 20 runs each, 2026-10-15): 12.2 vs 14.6 potions, significant (t -2.14), clear time 90.9 vs 88.4 s, damage taken 2030 vs 2421 (16 and 18 of 20 logs matched), no deaths. With the protection ended for both and synchronized starts (PR-12, 40 runs each, 2026-10-18 and 2026-10-20): 12.75 vs 13.93 potions, no difference shown (t -1.88), clear time 90.3 vs 89.2 s, damage taken 2230 vs 2369, no deaths; one retreat run ended without a clear, standing in the protection zone east of the spot (the retreat now skips zone tiles). The random control (PR-5, 10 runs: drink, engage, random step or stand with equal odds) clears in 102.7 s with 94.1 potions, far below every competitor.

The earlier table, kept as a record of the faulty arena: Measured with vBot's HealBot and TargetBot running alongside every spiral-* competitor (see the correction above); kept as a record until PR-4 replaces it. Cyclops clear, BENCHMARK pack v1: full gear restored before every run, the same 250 ms action gate and 100 ms decision tick for every competitor, heal below 160 unless stated, no deaths in any round below.

The first retreat-vs-vBot round alone looked significant (t -2.14); later rounds regressed toward a tie, which is why comparisons are now pre-registered with a fixed run count.

Rounds before these measured arena faults rather than policies (and these rounds themselves carry the last one found, vBot left enabled during our runs): the forge inflating every monster, gear dropped on deaths and never restored, vBot's HealBot ignoring a closed backpack, our probe deciding every 250 ms, the probe leaving chase mode on during retreats, and the vBot probe sampling only every 10 s, which rounded vBot's clear times up by about 5 s.

The probes are in apps/arena/ and the arena scripts in live/; the live world they drive (Canary 3.6.1, OTClient Redemption 4.1, run

off-screen and muted) stays in a local folder outside the repo (see Running).

• Set up: pwsh scripts/init.ps1 clones polyglot (which brings spiral) next to this repo, links both underdeps/ , installs Mesa's software Vulkan on Ubuntu, and builds the WGSL runner intotarget/ .-fast 1 only links and builds what is missing.

• Test: pwsh scripts/build.ps1 runs everysrc/*.spi file with tests throughdeps/spiral/apps/cells/spi/run_notebook.ps1 and fails unless every test executed and passed (no failed, unsupported, skipped or missing rows);-Files geometry,views runs a subset. Notebook outputs go tonotebooks/ (ignored by git, published byscripts/publish.ps1 ).

• Standards: scripts/build.ps1 first runsSpiralCompiler format src apps/arena apps/map --check and stops on any layout change; locally,eoie standards src apps/arena apps/map also checks the rulessrc/standards.spi declares (no comment lines, LF only, a target above every test, 120 columns, no floating point in 26 integer-only modules).

• Live arena: the arena needs a local world outside the repo: Canary, MariaDB, OTClient, logs (default ~/tibia-local , orTIBIA_WORLD ; MariaDB binaries fromTIBIA_MARIADB_BIN , default scoop's). For examplepwsh live/run-arena.ps1 -Runs 1 -Monster cyclops -MonsterCount 3 -Level 30 -AxeSkill 60 -HealBelow 160 -StrikeAbove 0 -FullExperience 675 -Competitors spiral-policy,vbot -Out results/arena-results-try.tsv . The hunting ground swaps in its own spawns and runs for a fixed time:-Scenario hunting-ground -SpawnFile live/hunting-ground-spawns.xml -KnightAt 1936,1336 -RunMs 300000 -MonsterCount 0 -FullExperience 1000000 . nExBot is not part of the repo (no license);-Competitors nexbot expects it under the world'sbots/ .

• Pre-registered blocks: pwsh live/block.ps1 -Ledger <ids> runs the block, then computes each test witheoie experiment analyze againsttarget/ledger , eoie's working copy (git-ignored). The record is the BENCHMARK section: register its test there witheoie experiment register target/ledger ... right after writing it.

• BENCHMARK.md : the OT Agent Arena (fidelity, baselines, scenarios, metrics, pre-registrations, publication).

• One novelty at a time, each with its own benchmark: fidelity before training, training before scale.

• CPU and GPU runs of the same seed must agree bit for bit (integer simulation, counter-based RNG).

• Every claim names its baseline, its control (random policy) and its budget; nothing is promoted on rhetoric.

• CipSoft-derived data (map, sprites, items) stays local; the arena publishes code, scenario specs and seeds.

• Each test target has a reason: ///> c is the semantic oracle on every rules test,///> wgsl the GPU simulator,///> lua the live client, and///> rust only the simulators' golden seeds, because a Rust compile costs about 10 s per test (662 s of a 1,164 s suite, 2026-10-13). A module'sstatic test only type-checks.

• Tibia dogfoods the tooling: a missing workflow feature goes into Cells, the compiler or eoie as a backlog item before anything works around it, never imitated in Tibia code.

• Every backlog item names the claim or milestone it serves; one that serves none is dropped, not done.