rules.md

Chorus rules

Seats and voices

• Four seats, four voices: 0 Bass (base MIDI 36), 1 Tenor (48), 2 Alto (60), 3 Soprano (72).
• The seat-to-voice assignment is a seeded permutation, redrawn every episode. This is the cross-author seating the game is built on: a policy cannot specialise on one register across a league round, and the same four policies re-seated write a different piece.
• Nobody else can write a note in your voice and you cannot write a note in theirs.

The piece

• bars bars (default 8, range 4..16). Every bar is exactly 16 steps.
• Key: a root pitch class drawn from C, D, E-flat, F, G, A.
• Mode: one of ionian, dorian, aeolian, mixolydian.
• Tempo: 84, 90, 96, 102 or 108 BPM.
• Chord plan: one chord per bar, from one of four progressions tiled across the piece (I IV V I / I vi IV V / I V vi IV / vi IV I V). The chord tones of a chord rooted on degree r are r, r+2 and r+4. The whole plan is public from turn 0 - it is the shared score sheet.
• All of it is drawn from the seed, so a replay re-derives everything from the seed alone.

Notation

A bar is 16 integer tokens, one per step:

• -1 is a rest.
• 0..13 is a scale degree: 0 the tonic, 6 the seventh, 7 the tonic an octave up, 13 the seventh two octaves up. midi = base + 12 * (d div 7) + scale[d mod 7], where base is your voice's base MIDI note in the episode's key; every legal token lands in base .. base+23.
• Every note lasts exactly one step. There are no ties, sustains, velocities or rests of length. A step is an onset if and only if its token is at least 0.

A turn, in order

There are exactly bars turns, one per bar. Turn t:

1. Open. Every voice HOLDS its previous bar (grid[v][t] := grid[v][t-1], or 16 rests at t = 0). Last turn's messages become this turn's heard.
2. Observe. Each seat is shown its voice, the complete grid, the whole chord plan, the live piece score and its four components, its OWN counterfactual credit, the other seats' lines from last turn, its private notes, and the exact list of legal targets.
3. Decide. All seats decide simultaneously; nobody sees another seat's bar for this turn until it resolves. All the LLM calls go out as one parallel batch.
4. Validate. A reply that times out, fails to parse or is illegal is retried once with an explicit hint; a seat still open after that plays the arpeggio baseline, which is legal by construction.
5. Apply, in seat order. target must satisfy 0 <= target <= t and steps must be exactly 16 tokens, each -1 or 0..13. Each seat writes only its own voice, so the resulting grid is order-independent.
6. Resolve once all four bars are in.

Hold versus edit

target = t writes the new bar. target < t rewrites one of your own earlier bars, and the hold at grid[v][t] stands unchanged - spending the turn on the past costs you the new bar. That is the whole trade: repair a weak early bar, or add a new one.

Caps

steps is also accepted as a string of 16 whitespace- or comma-separated tokens where ., -, r, R and rest all mean a rest; floats are rounded and numeric strings are accepted. Every string that reaches the replay is truncated on rune boundaries.

What is hidden

Almost nothing. The metric, its weights, the chord plan, the key, the tempo and the whole grid are public by design - this is co-composition against a known ruler, not a guessing game. Hidden from a seat: the other seats' credits, the other seats' private notes, the other seats' move for the current turn, and every policy display name (seats know each other only as cog aliases).

Endings

There is no third path. An episode never ends because a seat disconnected (its bars simply hold), never because the LLM is unavailable (the scripted fallback plays), and never mid-turn.