# Boss attack patterns: source and verification The field guide covers all **234 boss variants** marked as bosses in the installed FINAL FANTASY RESONANCE DEMO data. They use **139 distinct AI classes**. The six additional monsters using Chimera's shared AI also have reviewed patterns. This includes records for encounters beyond the playable demo; presence in the files does not establish demo availability. Source: Steam build **24032336**. The descriptions come from the installed AI scripts, their starting values, each variant's move tables, and the game's native battle helpers. They are static code findings, not reports from playing every encounter. ## Reading the guide - A repeating pattern gives the order of the boss's own actions. A boss may act more than once in a battle turn. Rules using actual battle turns say so explicitly. A combination says when several attacks are used together. - A fixed move stays fixed even if other turns choose attacks randomly. Random choices name the available moves and their chances; each chance applies only when that choice comes up. - Percentages preceded by “about” are rounded. The game draws from weighted lists using an integer random value, so a displayed percentage should not be treated as an exact long-run measurement. - HP and remaining Break gauge are separate conditions. “Below” excludes the boundary; “at or below” includes it. - If two moves share a name but affect different numbers of characters, the guide identifies the one-person and whole-party versions. - Special responses describe what triggers them, whether they replace the next attack or add an attack, and what restarts the normal pattern. ## How the AI was decoded The cooked Unreal Blueprint bytecode was decoded with UAssetAPI, including function properties and instruction offsets. The default-value data was decoded separately, because variable names do not reliably describe the stored values. The boss roster was joined to each variant's actual AI and move table. Form changes that switch move tables were reviewed separately. All 139 boss classes were decoded. Their 926 functions contain 5,522 direct jump and continuation targets; every checked target resolves to an instruction boundary. All 234 guide entries pass move-reference validation with no missing attack names. These checks validate the decoded structure and references; they do not independently prove every interpretation of the battle logic. Shared functions were also checked in the installed executable: - The action counter starts at zero and increments before the boss makes its first decision. A three-action remainder test of one therefore corresponds to actions 1, 4, 7, and so on, until the script resets or changes the counter. - Random move selection uses the supplied weights. An attack with zero weight may still be selected explicitly by the script. Selecting a move and using it later does not imply a second random draw. - Move-list indexes start at zero. For both direct and random lookups, selector zero chooses the first list; any nonzero selector chooses the second. Out-of-range direct references return no move. - Queued attack combinations preserve their supplied order and skip IDs absent from the skill table. Counterattack arrays are a different mechanism; two entries do not establish two consecutive counterattacks. - HP and Break comparisons, relationship filters, level-divisibility checks, resistance enums, and party-alive filters were checked against the corresponding native functions where used. - Summoned-unit scripts were followed when their deaths or completed attacks alter a boss's behavior, including Shiva's spirits and Leviathan's orbs. - Incoming damage is collected into an array for each target and processed before invoking its AI once. A multi-hit move therefore does not necessarily add one damage-counter point per hit. The exact boundaries between these damage groups, and the frequency of the separate counterattack event, remain unconfirmed; relevant Rage entries retain this limitation. ## Evidence and limits The [detailed source notes](boss-ai-evidence.json) list the AI class, function offsets, starting values, and supporting findings for each reviewed boss. Exact native names are retained there for reproducibility. Raw extracted assets and disassembly remain in the local analysis workspace and are not included in the public site. The review distinguishes reachable decisions from unused helper functions. An unused attack-selection helper is not evidence that a boss actually uses that attack. Specific unresolved timing or eligibility details are retained where they affect a fight; code certainty should not be confused with observations of every interruption, multi-hit interaction, or possible battle state. Decoding tools: [UAssetAPI](https://github.com/atenfyr/UAssetAPI) and [UAssetGUI](https://github.com/atenfyr/UAssetGUI). The game behavior itself was derived from the installed build.