Findings from a fight-lab prototype: six switchable fight-rule variants, about 40 fights played by hand and hundreds of thousands simulated. The code is on branch claude/fight-lab (no PR); play it with scripts/microslice.sh fight-lab.
Shape under test: gear has integer numbers and runners have none. A bump rolls d10 + weapon aim against a target, then d6 + weapon force against the armour where it lands. If it gets through, it deals wound cards: a severity (light, serious, critical) at a location, flavoured by damage type (stab gives a puncture or a deep stab, blunt gives a bruise or a broken arm). Keywords pick which drawn card lands: Precise (attacker picks), Brutal (worst), Spread (all land, one step lighter), Hooked (limbs first). The aim stance is set once and steers every bump. Everything is seeded, and every preview and popup percentage is enumerated through the same code that plays the bump.
The variants #
- Gate and draw. A hit draws one card from the whole deck, the card names the location, and armour there can stop it. Aim does nothing.
- Two margins. Every 3 points of aim margin draws one more card (max 3). The first card goes where you aim. A force margin of 3+ lifts light to serious. Keywords pick; otherwise the defender takes the lightest.
- Wound ladder. No deck. A hit lands light where you aim, force 5+ past armour makes it serious, and the same place again is one step worse.
- Openings. Plain hits are light, at the aimed location, and the same place again escalates. Once the enemy is hurt, an aim margin of 5+ (or his fumble, or an ambush) opens him up: a narrative popup of 2-4 options, once per side per fight. When the enemy gets an opening on you, you get a reaction popup instead.
- Defender's choice. A hit draws 2-3 cards and the defender picks which one lands; keywords override. A popup appears only for a real dilemma (no option dominates).
- Draw and open. Like 2, but the cards are light location cards: keywords choose where, same place escalates, and openings work as in 4.
Measurements #
Fair fight: pistol against a pistol guard in a vest, aiming at the body, 2000 fights per variant.
| variant | bumps | in 3-5 bumps | 1-2 bump fights | cards | first serious card decides | popups per fight | options per popup | words per popup | runner wins |
|---|---|---|---|---|---|---|---|---|---|
| 1 Gate and draw | 5.2 | 33% | 26% | 3.2 | 84% | 0 | - | - | 51% |
| 2 Two margins | 3.3 | 34% | 48% | 2.0 | 85% | 0 | - | - | 45% |
| 3 Wound ladder | 4.4 | 64% | 11% | 2.5 | 100% | 0 | - | - | 33% |
| 4 Openings | 5.1 | 65% | 0% | 2.9 | 95% | 0.6 | 3.1 | 38 | 48% |
| 5 Defender's choice | 4.1 | 41% | 33% | 2.6 | 80% | 0.01 | 2 | 23 | 46% |
| 6 Draw and open | 4.4 | 45% | 26% | 2.6 | 88% | 0.5 | 3.1 | 38 | 45% |
- Popup answers matter. Answering popups well instead of at random gives Openings 47% against 37%, and Draw and open 44% against 33%.
- Armour grades the fight. In Openings, with a pistol, the runner wins 70% against clothes, 48% against a vest and 36% against plate.
- The opening budget barely matters past one. Budgets of 1, 2 and 3 give 0.6, 0.7 and 0.8 popups per fight, because fights end before a second opening comes.
- An ambush decides the fight in variants 4 and 6 (100% for win and escape goals). Its choice only matters under a quiet goal (74-100%), where a strike can kill and a choke can fail.
Aim and goals #
Aim now pays off differently by location:
- Leg: a serious leg stops pursuit, not the fight.
- Arm: makes a pistol guard drop the gun and draw a backup baton; a second arm wound disarms him.
- Head: hard to hit and crit-heavy, and plate brings a helmet.
Goals are win, escape, quiet (no alarm, no body) and kill. The best aim across 80 cells (5 weapons, 4 enemies, 4 goals), in Openings: body 33, arm 26, leg 11, head 5, where it hurts 5.
- Escape against an armed guard: leg is best for every weapon.
- Win or kill against a vested guard: body for club, pistol and shotgun; head for the knife.
- Toolbot: arm (the manipulator) is best for every goal, so it is still dominant there.
- Near-ties: about a third of cells are within 3 points of the runner-up.
Within one enemy and goal, one aim usually wins whatever the weapon. The aim decision is "read his armour and know your goal": legible, rarely agonising.
Cognitive load #
| variant | decisions per fight | options per decision | tracked across turns | preview honest | log explains it | learnable in 3 fights |
|---|---|---|---|---|---|---|
| 1 | 0 | - | wound list | yes | yes | yes, nothing to learn |
| 2 | 1 (aim) | 6 aims | per-location wounds, card count rule | yes | needs the numbers line | barely: two margins plus keyword pick |
| 3 | 1 (aim) | 6 aims | pips per location | yes | yes | yes |
| 4 | 1 aim + 0.6 popups | 6 aims, about 3 options | pips per location; opening once hurt | yes | yes | yes |
| 5 | 1 aim + rare popup | 6 aims, 2 options | per-location wounds | yes | needs the numbers line | no: the defender's pick is invisible most fights |
| 6 | 1 aim + 0.5 popups | 6 aims, about 3 options | pips; card count; keyword pick | yes | needs the numbers line | borderline |
Same place again, one step worse is the most legible rule tried. The pips show it, players see the next hit coming, and it creates the "one more on that leg" moment with no hidden state.
Promising #
- Light plain bumps, an escalation ladder, and one opening popup once he is hurt (variant 4). It gives the arc: trade, wound, opening, decide.
- Popup odds as enumerated plain-word outcomes, e.g. "you go down 60% · he loses a bump 40%". The first version said "serious 100%" and hid that a third of those were criticals; hand play caught it.
- Pareto-pruned choices. Dominated options are dropped, so a dilemma is a real trade-off (leg: can't run, or arm: drop the pistol). Without pruning I misclicked into a loss.
- Showing every aim's odds at once instead of cycling six stances.
- Making hidden triggers visible: "the guard fumbles: your next hit finds an opening".
- A backup weapon and lamed-but-fighting made aim choices differ by goal.
- A non-lethal ambush option (choke) made ambush popups depend on the goal.
Good on paper, not fun #
- Defender chooses which drawn card lands (5): almost always an obvious pick, so the popup never earns its place. Auto-picking is invisible, so players can't learn it.
- Aim margin draws more cards (2, 6): hard to read, and it produces one-bump kills (48% of fights in 2 end within two bumps).
- The opening budget per fight by tension: it isn't a lever, because fights end first.
- Wound ladder (3) with a low strong-hit line: coin flips. With a high one it's legible, but without openings it's a race to the second hit on the same spot.
- Opening popups inside a bare duel: when every outcome counts as a win, "strike" always dominates. Popups need goals (quiet, escape, kill) to be real choices.
- Six aims break the four-options rule. Body, limbs, head and "where it hurts" would cover nearly every winning aim in the tables.