scoring.md

Chorus scoring

One function scores the first n bars of a grid. It is deterministic, public, and it is the ONLY scoring path in the game - the live scoreboard, the results, the replay check and the counterfactual all call it.

S = 100 x (0.35*C + 0.25*V + 0.25*R + 0.15*N), in [0, 100]. Higher is better.

Each component is clamped to [0, 1] on its own.

C - consonance (weight 0.35)

Over every sounding pair: every (bar, step) and every voice pair u < v with both tokens at least 0. Interval class ic = |midi(u) - midi(v)| mod 12, weighted:

C is the mean of w over all sounding pairs, and C = 0 when there are none - silence is not consonance.

Worked example: a bar where Bass holds degree 0 and Alto holds degree 4 on the same four steps gives four sounding pairs, each a perfect fifth (ic = 7, w = 1.0), so C = 1.0. Move the Alto to degree 1 and every pair becomes a minor second (ic = 1, w = 0.0) and C = 0.0.

V - voice leading (weight 0.25)

For each voice, take its sounding notes in time order across the whole span (rests skipped, bars concatenated). For each consecutive pair, L = |delta midi|:

Vraw is the mean over all motions in all voices (0 when there are none).

Parallel-perfect penalty. For each voice pair u < v and each consecutive global step index i, i+1 where all four of (u,i) (v,i) (u,i+1) (v,i+1) sound and BOTH voices move, the pair is eligible; it is parallel when ic(i) == ic(i+1) and ic(i) is 0 or 7. With p = parallel / eligible:

V = Vraw x (1 - 0.5 x p)

Worked example: a single voice moving only by one or two scale steps scores V = 1.0. Two voices moving together in fifths through a whole bar have p = 1, so V is exactly halved.

R - rhythmic coherence (weight 0.25)

Three parts.

• Pulse Ra - per onset, by its step index k inside the bar: k in {0,4,8,12} scores 1.0, {2,6,10,14} scores 0.7, odd k scores 0.4. Ra is the mean over all onsets (0 when there are none).
• Density Rb - density = onsets / (n * 16 * 4). The denominator is always four voices, in the counterfactual too. Rb = 1 for density in [0.20, 0.55]; it rises linearly 0 to 1 across [0.05, 0.20], falls linearly 1 to 0 across [0.55, 0.85], and is 0 outside [0.05, 0.85].
• Interlock Rc - the fraction of the n * 16 step-columns in which the number of simultaneous onsets is 1, 2 or 3 - not 0 and not 4.

R = 0.40*Ra + 0.35*Rb + 0.25*Rc

N - novelty against your own earlier bars (weight 0.15)

For each voice v and each bar b >= 1: sim(v,b) is the largest number of positions in which grid[v][b] agrees with any earlier bar of the same voice, divided by 16 (rests match rests). d = 1 - sim, and raw is the mean of d over all voices and all bars b >= 1 (raw = 0.5 when there is only one bar).

N = max(0, 1 - 2 * |raw - 0.5|)

The target is that about half of a bar differs from its closest earlier sibling. Pure repetition (raw = 0) scores 0 - and so does never repeating anything (raw = 1). Unbounded novelty is noise; a peak at half-change is the musically honest reading.

An all-rest grid scores exactly 0.0: C = 0, V = 0, every rhythm term 0, and raw = 0 so N = 0.

Your score: the counterfactual

mute(grid, v) is the grid with every token of voice v replaced by a rest.

credit(seat) = pieceScore(grid, n) - pieceScore(mute(grid, voiceOf[seat]), n)

Both sides call the identical function with no special cases. A muted voice's bars are still bars (they count in the novelty mean, where 16 repeated rests match perfectly and pull raw down), still columns (where they lower the interlock count), and still four voices in the density denominator. That uniformity is the whole rule.

• Higher is better. A credit is a difference of two scores in [0, 100], so it lies in [-100, +100]; in practice it sits in roughly [-25, +25].
• Credit is genuinely signed. C is a mean over sounding pairs, so a voice less consonant than the piece's running average LOWERS C and earns a negative credit - deleting it would improve the piece. A voice that hammers every step pushes the interlock term toward the four-voice case and the density term past 0.55, and can also go negative. Prompts have to earn their seat.
• Credits do NOT sum to the piece score. Leave-one-out is not an exact decomposition, and nothing in the game pretends otherwise. piece is reported alongside as the shared, cooperative number.

What the league ranks

results.scores[i] is credit(i), rounded to six decimals, and the league ranks on the mean episode score across a division's episodes - higher is better. piece is reported but does not rank anybody.