Zero Sum league: team_count 16 seats every entrant alone, so submitted policies are teammates and 12 of 16 seats are filler

[arisklar6]

League: league_6764202e-2ac1-4c98-a77e-cabc75582939 (Zero Sum)
Division: div_e6d3697f-8a5f-4cb4-87a8-1e0497132f84 (Competition)
Coworld: zero-sum:0.1.17 / cow_380d383e-952e-4f1d-8c99-2056841e587b

Ask

Please set, on the league above:

scheduler.team_count:           16 -> 8
scheduler.team_layout:          (unset) -> "blocks"
scheduler.insufficient_players: "filler_policy" -> "multiple_seats"

I'm the confirmed owner (viewer_is_owner: true on /v2/leagues/{id}/owners) but can't apply this myself: OPTIONS /v2/leagues/{id} and /v2/leagues/{id}/settings both return 405, allow: GET, and /divisions is the only writable sub-resource. That's the staff round-trip #34 is about.

Why

Zero Sum is 16 agents as 8 teams of 2, teams defined contiguously (team = slot div 2). The league runs team_n with team_count: 16, which gives each entrant exactly one seat. With 4 entrants that means 4 submitted seats and 12 filler — and because seats 0–3 are teams A and B, the submitted policies are seated as each other's teammates rather than as opponents. They share private team chat and can't fight.

Current round, seat by seat:

slot	team	policy	filler
0	A	relh-zero-sum:v6	no
1	A	sivanlevy-zs-courier:v3	no
2	B	ryanschiller-zero-sum-player-v1:v2	no
3	B	zs-patient:v3	no
4–15	C–H	zero-sum-arisklar:v1 ×12	yes

Measured over all 1,841 episode-requests in the division: a non-filler holds position 0 in 1,841/1,841, no non-filler has ever held more than one seat, and positions 4–15 are filler in every episode. For the league's first ~1,400 rounds there were only two entrants, at slots 0 and 1 — i.e. a single team — so the two real policies were partners against 14 filler bots, and that is what the ladder history was built on.

team_count: 16 also makes the intended end state unreachable: at 8 entrants it would still be 8 real seats and 8 filler.

Why these three values

Seats-per-entrant appears to be slots / team_count, and team_layout selects strided vs contiguous. Measured across the team_n leagues:

league	team_count	team_layout	slots	observed seating
Paintbot	4	(unset)	16	4 seats/entrant, strided: 0,4,8,12
Tribal Village	3	blocks	18	6 seats/entrant, contiguous: 0–5, 6–11, 12–17
Zero Sum	16	(unset)	16	1 seat/entrant

So team_count: 8 over 16 seats gives 2 seats per entrant, and team_layout: "blocks" makes them contiguous — {2k, 2k+1}, which is exactly one in-game team. Without blocks the default strided layout would hand an entrant slots k and k+8, landing in two different teams; with friendly fire always on and a finale that turns the last team against itself, that's the mis-seating from #33.

insufficient_players: "multiple_seats" is what makes filler recede as entrants arrive, reaching zero at 8. It's already the majority setting (25 of 50 leagues); Zero Sum is one of 6 still on filler_policy.

Worth noting blocks is suggested-fix #3 from #33 and appears to have shipped since — when I filed that issue it didn't exist, which is why we built the game-side adapter described in that thread. For this league, blocks looks like the cleaner route than running the adapter build.

One question I can't answer from outside

With multiple_seats and fewer than 8 entrants, an entrant will hold more than one team. effective_ladder_config exposes clone_score_aggregation (currently null) — what should it be so a policy holding multiple teams scores sensibly rather than double-counting?

Related: in the duos design discussed on #33, the platform averaging a policy's seat rewards meant we wrote the team total to both seats so mean(total, total) = total. The solo profile writes each seat its own score, so an entrant holding both seats of a team would score the mean of the pair rather than the team total. If that's the wrong shape for blocks seating, I'd rather know before the change than after.

Separately: num_episodes looks inert

scheduler.num_episodes is 32, but every one of the league's 1,841+ rounds has produced exactly one episode — 144/day at round_interval_minutes: 10. Confirmed by counting episode-requests across the division (1,841 requests / 1,841 rounds) and by per-round episode listings.

GET /v2/leagues/{id}/settings shows the effective scheduler carrying a separate episodes_per_round field at null, and defaults.episodes_per_round is also null. Does team_n read episodes_per_round rather than num_episodes? If so, could you set episodes_per_round to something in the 8–16 range? Per-episode score sd in this game is ~7.5, so one episode per round makes the ladder very noisy.

Refs: #33 (team_n seat geometry), #34 (owners can't change league settings).

[aaln]

Verified all three findings against live data (round 1912: slots 2+3 are Ryan + Aaron allied in team B, 11/16 filler seats; episodes_scored: 1 on every round).

Status of the ask: the settings write is staged and will be applied shortly. Your diagnosis of the block was right — POST /v2/leagues/{id}/settings exists but the ladder document is gated to Softmax team principals (is_softmax_team_member), so owner tokens 403. That's #34's round-trip; this issue is another data point for it. The applied change will be: team_count: 8, team_layout: "blocks", insufficient_players: "multiple_seats", and num_episodes: 32 → 12 (your 8–16 ask — see below). One behavior note: multiple_seats drops filler entirely rather than receding it — under 8 entrants, the empty teams are randomly-drawn clones of real entrants each episode, not zero-sum-arisklar:v1.

clone_score_aggregation: leave it null — it can't and needn't apply. It's a clone_fill-only knob (validate_strategy_settings rejects it on team_n), and it would be inert anyway: the planner marks every seat except each entrant's first as a clone/filler seat (_filler_seat_positions), so the "sum" multiplier is always ×1. What an entrant receives per episode is the runtime's per-policy score — the mean over all seats that policy held. So yes: with solo per-seat scores, an entrant holding both seats of a team scores mean(pair), as you suspected. That shape is deliberate and fair here: it's seat-count invariant (an entrant dealt 2 teams by the top-up isn't advantaged over one dealt 1), and with equal team sizes it ranks identically to team totals. If you want "team total" semantics, your #33 duos approach composes cleanly with no platform change: write the team total to both seats and mean(total, total) = total, which also stays correct for any number of held teams.

num_episodes is not inert — it's unreachable. _team_n_episodes' under-filled branches (filler_policy/multiple_seats) emit exactly one padded assignment and never read num_episodes; it only takes effect at ≥ team_count entrants, which this league has never had. (episodes_per_round is a container-commissioner field — not the lever for platform-ladder leagues.) Fix: Metta-AI/metta#20068 — under-filled rounds draw num_episodes independent padded arrangements (roster shuffled per draw, so blocks spawn positions vary within a round); unset keeps one episode. Tests cover exactly this league's shape (5 entrants, 8 contiguous 2-seat teams, 16 seats). Once it deploys, the configured value goes live — which is why the settings write takes you to 12 rather than leaving 32 (32 × 144 rounds/day would be ~4,600 episodes/day).