A visual guide to what a13 is, how one move travels through it, how it was selected, where the intelligence actually lives, why it is expensive, and how its ideas apply to mountains, ranged armies, waiting, Luck Shield, level-4 units, and finishing before Armageddon.
Version note: v0.8 is the AI policy version, not the game release number. This guide is a source snapshot audited on July 21, 2026.
The name means adaptive candidate number 13—the fourteenth zero-based child in a tournament campaign. It is not “Algorithm 13,” not a neural-network architecture, and not an acronym.
A13 is model-based rollout search with a learned linear value function, wrapped around a hand-authored tactical policy and protected by hard gameplay rails.
These labels appear throughout the guide. They matter: an automatic engine effect is not proof that the AI planned it, and a learned correlation is not a hard tactical rule.
Calling the whole thing “the model” hides the most interesting part. Most of a13’s competence comes from how four different systems cooperate.
Inherits years of v0.7/v0.6 behavior and proposes the incumbent move. It also contains the new ranged-posture, productivity, target-pressure, and finish logic.
Enumerates moves, melee stand cells, shots, area throws, and spells that obey current board and unit mechanics. Candidate zero is always the pilot’s choice.
Snapshots the fight, applies a candidate through the real engine, lets both sides play forward, scores the result, and restores the exact original state.
Turns 60 measurements of a future board into an estimated P(win). It is a logistic linear model: cheap, inspectable, and much smaller than a neural net.
StrategyV0_8 runs. The full rollout search lives in the Node/simulation runtime. A browser-only caller can therefore use the v0.8 behavior improvements without necessarily performing a13’s expensive future simulations.
Imagine a melee screen whose inherited v0.7 logic wants to strike a mountain while an enemy can be attacked or approached. Here is the complete path, simplified without changing its essence.
The inherited policy’s exact action array becomes candidate zero. Search is an editor, not a replacement policy conjuring every decision from nothing.
For production a13 this can include one move destination, up to six melee target/stand-cell pairs, four shot aims, two area-throw cells, and every currently legal spell. Caps keep the branching factor finite.
Every candidate is applied to a snapshot and evaluated immediately. Candidate zero is retained, then the two strongest legal challengers survive. This is fast relative to rolling every option twelve turns into the future.
Rollout one for every candidate uses the same private deterministic seed; rollout two also shares a different seed. This “common random numbers” design reduces the chance that one move wins only because it rolled luckier dice.
The real engine resolves mechanics. Each side follows its actual non-recursive native policy during the imagined future. Search does not call itself again inside search, avoiding an exponential hall of mirrors.
At the horizon, 60 board measurements become a probability-like value. The mean of the two futures is the candidate score.
Normally, a challenger must improve the estimated win probability by at least 0.03—three probability points—to replace a legal incumbent. Hard passive/endgame cases have stricter productivity priorities.
Illustrative scores, not a replay. An ordinary non-posture wait is treated more aggressively: a productive action can replace it on a tie. The special stronger-ranged wait retains the normal +.03 gate.
A13 does not build a giant tree to the end of the match. It spends a tightly bounded budget around the current decision and uses an approximate critic at the edge.
Every extra candidate, rollout, or future turn multiplies engine work. A full battle tree is far beyond a live decision budget.
It is a stable anchor and safe fallback. A learned critic with only two samples should not casually overturn a competent policy on tiny differences.
Abilities, movement, status effects, stolen skills, damage, and queues interact. A simplified simulator would be faster but could teach search the wrong physics.
The critic is logistic regression. It receives a vector x, multiplies each feature by a learned coefficient, adds a bias, and compresses the result into a value between zero and one.
HP advantage, stack-count advantage, attack/firepower advantage, ranged advantage, and wounded fractions.
Lap progress, seat advantage, exposed units, hourglass queue share, and remaining up-next queue.
Advancement, nearest-enemy distance, army spread, and distance from the narrowing-safe center.
Ranged, flyer, and caster fractions for both armies, plus the fraction of HP invested in ranged stacks.
Remaining shot advantage and an interaction between allied ranged fraction and distance to the enemy.
The raw 30 features are copied and multiplied by average board rangedness, yielding 30 shared terms + 30 ranged-board interaction terms.
Magnitude is shown for orientation. Logistic coefficients are correlated predictive signals, not independent rewards or causal truths.
hpAdv, cntAdv, atkAdv, rangedAdv, woundedOurs, woundedEnemy, advOurs, advEnemy, lapNorm, seatAdv, enemyExposed, ourExposed, hourglassFrac, upNextFrac, nearEnemyDistOurs, nearEnemyDistEnemy, spreadOurs, spreadEnemy, centerDistOurs, centerDistEnemy, ownRangedFrac, enemyRangedFrac, ownFlyerFrac, enemyFlyerFrac, ownCasterFrac, enemyCasterFrac, rangedHpFracOurs, rangedHpFracEnemy, shotsAdv, xRangedDist. The deployed basis adds xRg_ copies of all 30.The complaints that motivated v0.8 are handled mostly by policy rails around search. That is intentional: obvious bad actions should not depend on a noisy approximate value estimate noticing them every time.
If a damaging enemy attack is legal, take one. Otherwise, if a legal advance exists, advance. A mountain action can survive only when no enumerated productive alternative survives.
Preferred when it can reduce enemy HP.
Preferred when no attack lands yet.
Fallback only when no productive candidate exists.
Honest limitation: this is “no avoidable mountain attack,” not a theorem that the AI never hits a mountain. It can also be too aggressive: tactically valuable mining may be rejected whenever any move is available.
The screen should not charge merely because “aggressive” sounds good. If allied shooters have the stronger remaining ranged output, an idle movable pure-melee screen can spend the hourglass to let them fire and force the enemy to close.
What the formula misses: accuracy, armor, line of sight, area damage, two-shot costs, matchup access, and the positional value of one particular screen. Search can correct some mistakes inside twelve future turns, but not every long-term misclassification.
Direct v0.8 replaces an avoidable defend/Luck Shield with attack, spell, or move. If truly immobilized, hourglass is preferred when legal. Shield remains a last resort, so the correct promise is “zero avoidable shields,” not “mathematically zero shields.”
An empty or end-only decision becomes an explicit engine-valid fallback. The productivity layer then searches for damage or movement. Forced locks and truly actionless states can still fall back safely.
Blindly banning wait previously damaged decisive strength. Ordinary waits can be replaced by a productive action on a rollout tie; the explicit stronger-ranged posture wait keeps the normal +.03 protection.
The first Armageddon wave begins at lap 12. A short search horizon cannot reason all the way there from an early fight, so v0.8 adds explicit deadline-aware behavior.
From lap 7, if living non-summoned original-stack HP is at least twice the enemy’s, direct combat becomes lexicographically stronger. Prefer a kill, then maximum immediate damage, then an attack without movement; if no attack exists, advance.
From lap 9, every surviving v0.8 army must deal positive enemy damage when possible, otherwise move toward the enemy. Stronger-ranged patience ends here.
Removing an activation is stronger than merely spreading damage.
Do not spend several laps cutting one tank and then open a fresh target.
Estimate remaining attack-work from HP and current delivery damage, then focus the stack most likely to become impossible to finish later.
Reserve future Wild Regeneration HP and add a melee tax for permanent Dulling Defense attack loss. Avoid unsafe fresh Dulling melee when another positive target exists.
Equal damage without spending movement is more stable and preserves positioning.
This lowers Armageddon risk; it does not guarantee zero Armageddon. The formal campaign ceiling was 0.1%, and exact a13 did not pass it. Broad artifact, synergy, and roster distributions remain an active measurement problem.
The engine handles automatic effects during every real and simulated action. The AI must separately choose targets, stand cells, shots, and spells well enough to exploit those effects. Coverage proves the paths run safely; it does not prove optimal mastery.
A 2×2 melee unit with Tie up the Horses Aura, Crusade, and Rapid Charge. Candidate damage estimation prices charge distance; real rollouts resolve the full engine effects.
A 2×2 melee unit with Web Aura, Infest, and Predatory Assimilation. It can steal an active ability and become a genuine runtime caster or shooter.
A 2×2 melee protector. Dense Flesh makes aimed ranged attacks consume extra shots; Flesh Shield Aura absorbs most direct damage to protected allies and recalculates defense.
A 2×2 melee unit with AI Driven, Magic Shield, and Boar Saliva. Magical resistance and physical miss effects resolve through the engine and candidate damage estimates.
The forced L4 panel checks operational coverage: all four units as candidate-owned and opponent-owned, both physical seats, across the live map rotation; acting turns; no candidate rejections; no raw candidate end-turns; no candidate-owned L4 mountain attacks; no stuck fights. It uses Black Dragon as a fixed control. This is a behavior/safety census, not a balanced L4 win-rate ranking.
The word training is useful shorthand, but it can imply the wrong mechanism. A13 emerged from offline simulation, value-model fitting history, finite candidate mutation, and tournament selection—not from gradient-updating a deep policy during play.
A13’s campaign label was adaptive-a13-from-c43-gate-search-gate. It changed its parent c43’s ordinary challenger threshold from 0.025 to 0.03.
The fixed leaf belongs to a line of offline outcome predictors and evolutionary/CEM research. Candidate campaigns could blend value vectors by 15% or 25% as well as flip discrete controls.
Research ranking was lexicographic: candidate win rate, then draw rate, then non-loss Armageddon rate. Fresh validation seeds checked screen overfitting.
| Term | Accurate for a13? | Why |
|---|---|---|
| Deep reinforcement learning | No | No deep policy/value network or policy-gradient loop. |
| Online learning | No | Weights and rules are fixed during a match. |
| Supervised value learning | Yes | The logistic critic maps simulated board features to eventual win outcomes. |
| Evolutionary / black-box optimization | Yes | Finite genomes were mutated, screened, and selected by tournament outcomes. |
| Model-based planning | Yes | The real game engine is the transition model used to imagine consequences. |
| RL-adjacent self-play optimization | With explanation | Reasonable umbrella language, but it must not be confused with PPO/Q-learning. |
A13 ranked first for strength on both independent hosts in the campaign evidence. It did not pass the formal Armageddon promotion rule. Both statements are important.
| Host | Validation games | Wins | Losses | Draws | Raw win rate | Decisive win rate |
|---|---|---|---|---|---|---|
| M4 | 61,440 | 36,006 | 25,078 | 356 | 58.60% | 58.95% |
| Ryzen 7 9800X3D | 27,648 | 16,249 | 11,232 | 167 | 58.77% | 59.13% |
| Combined | 89,088 | 52,255 | 36,310 | 523 | 58.66% | 59.00% |
Campaign validation was versus v0.7 under the frozen training alias and campaign binding. It is strong evidence, but not a located post-promotion production-rebound tournament artifact. The raw dual-host output bundle is not currently included in the public source repository, so these figures are an audited snapshot rather than a downloadable public dataset.
A13 ranked first for strength on the M4 and the 9800X3D. On each host it improved the stable c48 anchor by roughly +0.76 percentage points raw. Cross-host agreement is valuable because it is harder to explain as a single-machine accident.
M4 Armageddon reached: 502 / 61,952 = 0.81%. 9800X3D: 233 / 28,160 = 0.83%. Required ceiling: 0.1%. New-L4 panels were higher still. Final campaign leaderboards therefore recorded no formally eligible promotion candidate.
Honest conclusion: a13 was later promoted manually after ranking first for strength on both hosts. It would be inaccurate to say it passed the campaign’s formal near-zero-Armageddon gate.
| Evidence family | Approx. raw result vs v0.7 | Interpretation |
|---|---|---|
Earlier d174… stable candidate | ≈62.4% | Earlier candidate family; not exact a13. |
| Later r3 qualification of that family | ≈61.34% | Still not exact a13’s frozen genome. |
| Exact a13 dual-host validation | ≈58.66% | The directly attributable campaign number. |
AI meta cohorts answer a different question: which units, artifacts, augments, synergies, maps, and roster types perform under v0.8+a13. A non-mirrored balance study does not establish an AI-version win rate against v0.7.
The report renderer is not the bottleneck. The difference is that plain v0.7 makes one native decision, while v0.8+a13 may execute dozens of hidden engine turns to decide one visible action.
Evaluate heuristics, emit an action, resolve it. Historical 1.05-million-fight report throughput was on the order of hundreds of fights per second with 12 M4 workers distributed across cohorts.
Enumerate candidates, run an immediate census, snapshot and restore state, simulate up to about 72 future unit turns, extract value features, and repeat for many real decisions in every fight.
Hybrid systems are powerful because rules cover known hazards and learning covers fuzzy trade-offs. They also inherit the blind spots of every layer.
Twelve unit turns is roughly local tactics, not a lap-12 plan from the opening. Fortresses, long bodyguard setups, ammunition wars, and delayed regeneration can sit beyond the leaf.
Caps can omit the best stand cell, aim, throw, or move. Only two challengers survive the immediate shortlist, so a setup move with weak instant value may disappear.
The critic sees aggregates, not explicit unit, ability, artifact, augment, or synergy identities. Novel combinations can be out of distribution.
Two paired rollouts reduce noise but do not reliably sample low-frequency steals, misses, resistances, or high-impact luck tails.
“Advance instead of mine” stops obvious waste but may suppress a genuinely useful obstacle plan. Every invariant trades tactical freedom for reliability.
Deadline fallback depends on hardware and load. If a difficult decision repeatedly exceeds the budget, the shipped AI effectively becomes more of its baked pilot.
Rollout policies do not invoke another full a13 search. This is necessary for speed but means the imagined opponent can be weaker or stylistically different from deployed search.
A coefficient or meta-report lift can reflect which armies create a state, not what changing that feature would cause. Prediction and balance causality are different problems.
A unit acting without rejection proves integration. It does not prove target selection, positioning, or synergy use is near-optimal.
A practical diagnostic rule: first ask whether the desired action was absent from the candidate set, removed by the shortlist, scored badly by the critic, rejected by a hard rule, or lost to a deadline. “The AI made a bad move” can originate at five very different layers, and each needs a different fix.
The best next step is unlikely to be “turn every knob up.” More rollouts can improve decisions and simultaneously destroy tournament throughput or trigger more deadline fallbacks.
For every bad replay, record incumbent, candidate set, shortlist, rollout means, override reason, deadline/circuit state, and executed fallback. This turns anecdotes into labeled failure classes.
Collect a13 recommendations offline, then train a compact action-ranking model or bake stable rules. Let cheap decisions bypass rollout; reserve search for uncertain states.
Spend little when the incumbent is obviously legal and strong. Expand rollouts when top candidates are close, the state is stochastic, an L4 mechanic is active, or the endgame is near.
Preserve diversity by action class and target, not only immediate value: one movement setup, one kill line, one control spell, one target-coverage delivery. Measure whether the eventual winner was pruned.
Represent protection, control, regeneration debt, Dulling exposure, remaining high-impact casts, shielded HP, current lanes, and Armageddon finish slack—carefully, with out-of-sample validation.
Use one model for the next-lap fight and another for long-horizon terminal risk. Blend by lap, board openness, rangedness, and ability state instead of forcing one linear leaf to express everything.
Profile battle snapshot/restore, pathfinding, repeated legal queries, and invariant feature blocks. A faster engine buys stronger search without changing the live deadline.
Require strength confidence, avoidable passive rate, mountain/shield rate, action rejection, Armageddon reach/decision, L4 coverage, cross-host reproducibility, and timing-fallback rate—not one headline win rate.
| Question | Metric | Best test | Failure interpretation |
|---|---|---|---|
| Does search find better tactics? | Paired score-rate delta | Same seeds, seats, armies, host | Critic/candidate/search issue |
| Does AI stop wasting turns? | Avoidable passive per 1k decisions | Action audit with legal-alternative replay | Productivity rail gap |
| Does it finish cleanly? | Armageddon reached and decided | Long balanced + dominant-state panels | Horizon or endgame policy gap |
| Are new units safe? | Rejections, raw no-ops, acting coverage | Forced L4 ownership/seat/map matrix | Engine/candidate integration gap |
| Is it deployable? | p50/p95/p99 decision time and fallback rate | Production-like fixed host load | Compute budget mismatch |
| Did it generalize? | Fresh-seed cross-cohort lower bounds | Unopened seed banks and diverse drafts | Panel overfitting |
You do not need to begin with advanced neural RL. A13 sits at the intersection of game AI, simulation, statistics, optimization, and systems performance.
Model the board as state; legal moves as actions; the engine as transition dynamics; elimination/score as objective. Learn Markov decision processes as a vocabulary, even when the practical agent is not pure RL.
Understand branching factor, depth, opponent models, stochastic outcomes, evaluation leaves, alpha-beta intuition, Monte Carlo estimates, and why bounded search must prune.
Learn logits, sigmoid, cross-entropy, regularization, feature scaling, interactions, calibration curves, leakage, and distribution shift. You can understand a13’s critic completely with these tools.
Study common random numbers, paired seat swaps, confidence intervals, multiple comparisons, holdout seeds, preregistered gates, and why sub-1% deltas demand many games.
Explore finite candidate arms, mutation, crossover/blending, cross-entropy method, successive halving, multi-armed bandits, and tournament selection when gradients are unavailable.
Learn why a learned scorer should not own legality. Study constrained optimization, action masking, lexicographic objectives, fail-closed fallbacks, and regression/property tests.
Measure snapshots, memory churn, pathfinding, worker oversubscription, cache behavior, deadlines, deterministic RNG, replayability, and host-bound performance.
Once search logs are trustworthy, learn behavior cloning, ranking losses, uncertainty estimation, ensembles, policy/value networks, and selective search. Deep RL becomes relevant later—not automatically better.
Best practical exercise: take ten visibly bad fights. For each decision, reconstruct the five checkpoints—candidate availability, shortlist survival, rollout score, rule arbitration, timing fallback. Then propose the smallest change that fixes the correct layer and test it on fresh paired seeds.
These values are pinned in the published a13 profile as of July 21, 2026. The hash makes “a13” an exact reproducible genome identity rather than a loose nickname.
src/ai/versions/v0_8_a13_profile.ts — exact identity, weights, search and policy controls.src/ai/versions/v0_8.ts — productive actions, stronger-ranged posture, native policy integration.src/simulation/search_driver.ts — enumeration filtering, shortlist, rollouts, gate, deadlines, circuit.src/simulation/value_features.ts — 20 base + 10 composition + 30 ranged interaction features.src/ai/versions/v0_8_dominant_finish.ts — 2:1 HP and lap-based finish states.src/ai/versions/v0_8s_finish.ts — target-work scheduler and universal sprint behavior.src/simulation/v0_8_aggressive_12h.ts — candidate generation, screening, validation and promotion gates.src/simulation/v0_8_l4_coverage.ts — Champion, Arachna Queen, Abomination and Frenzied Boar coverage.src/ai/candidates.ts — legal candidate census and capability routing.src/constants.ts — narrowing, Armageddon, morale, shield and wait constants.| Behavior | Confidence type | Precise claim |
|---|---|---|
| Generated mountain/shield challengers | Hard filter | Search does not introduce them for v0.8. |
| Inherited passive repair | Hard policy rail | Replace with legal damage first, otherwise nearest advance, when found. |
| Stronger-ranged wait | Heuristic + search | Uses ammo × damage × living amount; retains +.03 rollout protection. |
| Target choice | Scheduler + search | Kills, wound focus, remaining work, regeneration and Dulling tax guide selection. |
| New-L4 mechanics | Real engine | Automatic effects resolve in live actions and rollouts; tactical mastery is not guaranteed. |
| Near-zero Armageddon | Not achieved by campaign gate | Exact a13 strength passed; formal ≤0.1% promotion ceiling did not. |
| ≈58.66% vs v0.7 | Frozen campaign evidence | Dual-host validation under the training binding; not a universal matchup constant. |