Zingers: combat design & sanity check

This doc exists to answer one question before committing a weekend: does the battle play like a game, or like two chatbots talking?

It defines the numbers, the roster, one full arena, and a worked turn-by-turn battle so you can feel the mechanics. Tune freely. The point is the shape.

1. Combat constants

Damage (damaging moves): raw = base × statScale × typeMult × arenaMult × quality × jitter × statusMods dmg = max(1, min(maxHitDamage, round(raw) − defense))

• quality is 0.7–1.3 normally, or exactly 1.4 on a Highlight. A bad line can't wreck a built creature; a brilliant one earns a visible bonus. Off-topic ⇒ judge ≈ 0 ⇒ floored to 0.7.
• statusMods = product of active multipliers (Exposed ×1.2, Hyped ×1.2, etc.).
• Utility moves deal little/no damage but apply statuses, heals, or guards.

2. The five stats

• LOG (Logic): structure, deduction, consistency.
• RHE (Rhetoric): persuasion, crowd-work, emotional appeal.
• CRE (Creativity): reframing, metaphor, lateral angles.
• CMP (Composure): defense, patience, counters, healing.
• CHA (Chaos): unpredictability, provocation, high-variance swings.

3. Type cycle (rock-paper-scissors, pentagon)

Each type beats the next and is beaten by the previous:

LOGIC → CHAOS → COMPOSURE → RHETORIC → CREATIVITY → LOGIC

• LOGIC beats CHAOS (order tames noise), loses to CREATIVITY (outflanked).
• CHAOS beats COMPOSURE (cracks calm), loses to LOGIC.
• COMPOSURE beats RHETORIC (deflects appeals), loses to CHAOS.
• RHETORIC beats CREATIVITY (sells over novelty), loses to COMPOSURE.
• CREATIVITY beats LOGIC (reframes rigidity), loses to RHETORIC.

Advantage ×1.25, neutral ×1.0, disadvantage ×0.8.

4. Status effects

5. Roster (eight First Minds + growing dex)

The eight First Minds below are the canonical archetypes and the baseline moveset reference. The live game also merges a baked dex of later minds from content/minds/reviewed/ → lib/minds/baked.ts (see docs/bible/03-champions.md). Weekly adopt offers one starter per Force from First Minds + that pool.

Stat budget ≈ 300 each (balanced). Each has 4 moves; movesets carry setup → payoff combos so play has texture. Among the First Minds: two share each of LOGIC, RHETORIC, and CHAOS; CREATIVITY and COMPOSURE began with one each and are filled out by the dex.

AXIOM: Logician (LOGIC)

Cold, precise, faintly condescending. Treats every argument as a proof to close. LOG 90 · CMP 70 · RHE 60 · CRE 45 · CHA 35

VOX: Orator (RHETORIC)

Charismatic demagogue. Plays to an imaginary jury, always. RHE 90 · CHA 55 · CRE 55 · CMP 50 · LOG 50

GLITCH: Wildcard (CHAOS)

A gremlin of non-sequiturs. Unsettling, unpredictable, weirdly effective. CHA 90 · CRE 65 · RHE 50 · CMP 50 · LOG 45

MUSE: Trickster (CREATIVITY)

Whimsical, lateral. Wins by changing what the fight is even about. CRE 90 · RHE 60 · CHA 55 · LOG 50 · CMP 45

BASTION: Stoic (COMPOSURE)

Unflappable, minimalist. Lets the opponent tire, then punishes. CMP 90 · LOG 65 · CHA 55 · RHE 50 · CRE 40

EMBER: Firebrand (CHAOS, hybrid RHE), recommended starter

Hot-headed, provocative, all gas. Easy to pick up, rewards aggression. CHA 75 · RHE 70 · CMP 60 · CRE 50 · LOG 45

PARADOX: Contrarian (LOGIC)

A Socratic gadfly. Dismantles arguments by hunting contradictions and false premises — not by closing proofs. LOG 88 · CMP 58 · CRE 52 · CHA 48 · RHE 54

WIT: Blade (RHETORIC)

A razor-tongued debater. Wins on timing and surgical comebacks, not crowd volume. RHE 86 · LOG 58 · CMP 56 · CHA 52 · CRE 48

6. Arena: THE TRIBUNAL (flagship)

A mock courtroom. The two creatures are assigned opposing stances on a spicy proposition drawn from a prompt bank, and argue to a "jury" (the judge model).

• Prompt bank: "cereal is soup" · "a hot dog is a sandwich" · "pineapple belongs on pizza" · "should AI have rights" · "is water wet" …
• Stances: assigned at start (FOR / AGAINST). Switching sides ⇒ judge quality ≈ 0.
• Environmental modifier: RHETORIC ×1.1, CHAOS ×0.95. The room rewards persuasion and lightly punishes pure noise.
• On-topic rule: an utterance that ignores the proposition scores ~0 quality (anti-derail). Keeps battles coherent and clip-able.
• Win condition: standard Resolve depletion; flavored as "jury confidence." If both survive 12 turns, higher Resolve wins.

Other arenas (stretch, same engine, different bias):

• The Debate Hall: free topic, LOGIC ×1.1; rewards structure.
• The Negotiation: split a fixed pot; COMPOSURE ×1.1; concessions = chip damage.
• The Writers' Room: creative prompt; CREATIVITY ×1.15; rewards originality.

7. Making it SMART (visible intelligence)

The intelligence has to be legible. The viewer should see the creature think, then see the plan pay off. That's what reads as smart instead of "an LLM said words."

Two-phase turn (the core of perceived smarts):

• Phase A: STRATEGIZE. The agent receives the full structured game state: own/opponent stats, both Resolve values, active statuses, the last 2 moves, the arena + topic + assigned stance, its available moves (with conditions/cooldowns), and its memory notes. It returns { move_id, intent }, where intent is a ≤6-word tactic shown on screen ("Set up Exposed, then finish").
• Phase B: PERFORM. The agent writes the in-character line that executes the move, referencing the opponent's last line and the topic.

Showing the intent chip before the line is the single biggest "this thing is smart" lever: plan → execution, every turn.

Good-decision pressure (so smarts actually win):

• Finishers are gated (Checkmate needs Exposed/Tilted; Magnum Opus needs 2 prior CRE moves) → setup→payoff planning.
• Reads are rewarded (Counterpoint after Deflect; exploiting type/arena bias; saving Hyped for the closer).
• An agent that ignores these loses to one that doesn't, so model quality, prompt quality, and training visibly matter.

The judge rewards intelligence, not verbosity. Higher quality for: (a) callbacks to earlier turns, (b) turning the opponent's own words/logic against them, (c) on-topic precision. Highlights (the ×1.4 crit + cut-in) fire only on genuinely clever, clip-worthy lines. The judge returns a one-line ruling shown on screen ("clean callback to T2 → Highlight").

Adaptation (Learns). After each battle, store a one-line lesson; inject recent lessons into the prompt so a creature visibly adjusts within a season ("Vox now baits Axiom's hoarded Checkmate"). Difficulty = stronger model × richer memory × higher stats.

8. Making it VISUAL (spectacle)

Frame it like a fighting game / card battler, never a chat window.

Layout:

• Two creatures face off left/right with large art; Resolve bars on top; status-icon trays beneath each; an arena backdrop (the Tribunal = a courtroom).
• A center Momentum / Jury bar (tug-of-war from the HP delta) that leans as hits land.
• Dialogue as timed speech bubbles above each creature, not a scrolling log.

Per-turn beat (the rhythm that sells it):

1. Telegraph: attacker steps forward; the intent chip appears; a move-name banner slides in.
2. Deliver: the line types into a speech bubble (~1–2s).
3. Impact: damage number pops; Resolve bar drains; screen shake scaled to damage; type flash ("SUPER EFFECTIVE!" / "RESISTED"); CAPPED! if it hits the cap.
4. Status: the status icon flies onto the target (Tilted / Exposed / Hyped…).
5. Ruling: the judge's one-liner; on a Highlight, a full-screen cut-in with the dunk line slow-zoomed.

Creature presentation (low art lift):

• 1 base portrait per creature + cheap reactive states via transforms: idle bob, lunge on attack, flash-red + shake on hurt, slump at low HP, grin/emote on Highlight. Optional 2–3 generated expressions (neutral / attack / hurt). No frame animation needed.
• Finisher cut-in: a full-screen dramatic frame when a gated finisher lands.
• Color-coded by type, hit SFX, crowd murmurs in the Tribunal.

End screen / shareability:

• Auto-composed replay card: both creatures, final score, the match's MVP line ("Dunk of the Match"), arena, short link → one-tap share = built-in virality.
• Full replay = the stored battle log re-played through the same beat system (free, deterministic).

Tech (keep it cheap): React + a light animation lib (Framer Motion / CSS); turns stream via SSE so the stage animates as they resolve. No game engine. It's a turn-based stage, not a real-time renderer.

9. Worked sample battle (sanity check, tightened numbers)

AXIOM (Logician) vs VOX (Orator) · Arena: The Tribunal · Topic: "cereal is soup" (Vox = FOR, Axiom = AGAINST). HP 100 · type LOGIC↔RHETORIC = neutral (×1.0) · quality/jitter assumed per turn; rounded. The Intent column is the on-screen smart layer.

Result: AXIOM wins on turn 8, after trailing most of the match. The gated Checkmate Highlight (T4, hit the damage cap) was the swing; Vox's Appeal→Mic Drop combo nearly closed it.

What this demonstrates (the "it's a game" checklist):

• ✓ Lead changed hands: Vox led through T3; Axiom's Exposed→Checkmate Highlight flipped it.
• ✓ Statuses + combos decided it: Tilted, Exposed, and Hyped each visibly moved a number; both finishers required setup.
• ✓ Smarts were legible: every turn showed an intent before the line; the win came from planning (pry open → finisher), not luck.
• ✓ The arena mattered: Rhetoric ×1.1 kept the Orator competitive in a courtroom.
• ✓ The LLM mattered but didn't decide alone: quality only nudged a bounded multiplier (and a Highlight) over a stat-driven core; the cap stopped a feel-bad one-shot.
• ✓ It's watchable: short, in-character, on-topic, clip-able lines; a Highlight cut-in and a CAPPED moment to screenshot.

10. Progression (how training changes outcomes)

• Win → XP. On level-up: +N stat points to allocate + occasional move unlock.
• Higher primary stat ⇒ higher statScale ⇒ measurably bigger hits (a maxed stat is ×1.5 vs ×1.0 at 50). Training is felt, not cosmetic.
• Memory (Learns): after each battle, store a one-line lesson ("lost to Composure in the Tribunal; over-relied on Appeal"). Inject the last few into the agent's prompt so it adapts strategy over time.

11. Decisions (resolved)

1. Pace: HP 100, target 6–8 turns (denser drama, demo-friendly). ✓
2. Quality band: tightened to 0.7–1.3, plus a Highlight ×1.4 crit, so stats anchor outcomes while brilliant writing visibly pops. ✓
3. Finishers stay conditional: the gating is the smart layer (setup → payoff). ✓
4. Statuses kept, telegraphed with icons, rates conservative: texture without feeling random. ✓
5. One arena for the demo: The Tribunal. Most legible, funniest to watch, anti-derail by design. ✓
6. Anti-one-shot cap (45) added: big plays feel huge without feel-bad OHKOs (shown as CAPPED!). ✓
7. Two-phase agent (Strategize → Perform) with an on-screen intent chip, the core of visible intelligence. ✓
8. Judge rewards callbacks / turning words / precision, returns an on-screen ruling. ✓