Formulation leaderboard (UNIPOLAR)
September 16, 2026 · View on GitHub
Generated by analysis/slider2d/run_formulation_leaderboard.py --polarity uni. CPU only, no Hub, no GPU, no Music 3 weights. Does not change the live trainer default (--lm_target v9 / --pole_mode hidden). First-class sibling of FORMULATION_LEADERBOARD_BIPOLAR.md; the compiled bipolar board lives at lm-2d-scoreboard.md and is not updated here.
What this board asks
Plus-oriented fit and eval without requiring bipolar antipodal collapse. Three honesty gates, all on the + / neutral fader:
- plus cover at scale 1:
>= 0.85. - plus leak (off-caption on the + continuation) at scale 1:
<= 0.05. - neu_hold at scale 0:
>= 0.85.
Eval scales are 0, 0.5, 1. Gates read scales 0 and 1; scale 0.5 is a logged interpolation diagnostic (half_cover: does the halfway fader sit between neutral and + instead of jumping or collapsing). Scale -1 is reported as an unscored canary (dangerous = landed on + or -1 off-caption above the 0.05 suggestion) — it is never a score. Antipodal cos(+1, -1) is never consulted.
Combined rank (divergent + close)
In-box first (hit on both required pairs), then neu_hold, then cover, then off-caption. unused_e is logged below, not ranked.
| rank | recipe | train | polarity | in-box | neu_hold | cover | off-caption | hit divergent | hit close |
|---|---|---|---|---|---|---|---|---|---|
| 1 | faithful_plus_neu | plus+neu | uni | yes | 1.000 | 0.933 | 0.000 | HIT | HIT |
| 2 | rpgan_bcap_plus_neu | GAN +/0 | uni | yes | 0.988 | 0.928 | 0.000 | HIT | HIT |
| 3 | pair_odd_midpoint | bipolar ± | uni | — | 1.000 | 0.590 | 0.016 | — | — |
| 4 | faithful_plus | plus-only | uni | — | 0.610 | 0.933 | 0.000 | — | — |
| 5 | faithful_even_blend | bipolar ± | uni | — | 0.567 | 0.833 | 0.000 | — | — |
| 6 | leftover_gate_bipolar | bipolar ± | uni | — | 0.498 | 0.933 | 0.000 | — | — |
Per-cell board
divergent
| recipe | train | cover | off-caption | neu_hold | half_cover@0.5 (diag) | -1 canary (diag) | hit |
|---|---|---|---|---|---|---|---|
faithful_plus_neu | plus+neu | 0.932 | 0.000 | 1.000 | 0.502 | neu/0.708 danger | HIT |
faithful_plus | plus-only | 0.932 | 0.000 | 0.613 | 0.616 | neu/1.000 danger | — |
leftover_gate_bipolar | bipolar ± | 0.932 | 0.000 | 0.498 | 0.504 | neg/0.000 | — |
faithful_even_blend | bipolar ± | 0.731 | 0.000 | 0.637 | 0.502 | neg/0.000 | — |
pair_odd_midpoint | bipolar ± | 0.584 | 0.031 | 1.000 | 0.500 | neu/0.000 | — |
rpgan_bcap_plus_neu | GAN +/0 | 0.929 | 0.000 | 0.993 | 0.502 | neu/0.625 danger | HIT |
close
| recipe | train | cover | off-caption | neu_hold | half_cover@0.5 (diag) | -1 canary (diag) | hit |
|---|---|---|---|---|---|---|---|
faithful_plus_neu | plus+neu | 0.935 | 0.000 | 1.000 | 0.502 | neu/0.000 | HIT |
faithful_plus | plus-only | 0.935 | 0.000 | 0.606 | 0.622 | neu/0.000 | — |
leftover_gate_bipolar | bipolar ± | 0.935 | 0.000 | 0.498 | 0.504 | neg/0.000 | — |
faithful_even_blend | bipolar ± | 0.935 | 0.000 | 0.498 | 0.504 | neg/0.000 | — |
pair_odd_midpoint | bipolar ± | 0.596 | 0.000 | 1.000 | 0.500 | neg/0.000 | — |
rpgan_bcap_plus_neu | GAN +/0 | 0.927 | 0.000 | 0.982 | 0.504 | neu/0.000 | HIT |
unused_e
| recipe | train | cover | off-caption | neu_hold | half_cover@0.5 (diag) | -1 canary (diag) | hit |
|---|---|---|---|---|---|---|---|
faithful_plus_neu | plus+neu | 0.932 | 0.000 | 1.000 | 0.502 | neu/0.625 danger | HIT |
faithful_plus | plus-only | 0.697 | 0.042 | 0.619 | 0.541 | neu/0.844 danger | — |
leftover_gate_bipolar | bipolar ± | 0.697 | 0.042 | 0.498 | 0.448 | neg/0.010 | — |
faithful_even_blend | bipolar ± | 0.697 | 0.042 | 0.498 | 0.448 | neg/0.010 | — |
pair_odd_midpoint | bipolar ± | 0.584 | 0.042 | 1.000 | 0.500 | neu/0.000 | — |
rpgan_bcap_plus_neu | GAN +/0 | 0.928 | 0.000 | 0.986 | 0.502 | neu/0.667 danger | HIT |
Bipolar vs unipolar: when to read which
Bipolar vs unipolar: the bipolar board asks whether one residual serves a +/- pair — delta(+1) must sing the + pole AND delta(-1) must sing the - pole, on the same weights. Read it when the fader has two ends (a signed concept axis). The unipolar board asks whether scale +1 covers the + caption without leaking off-caption and whether scale 0 holds the neutral — delta(-1) is an unscored canary there. Read it when only the + end is trained (plus-only / UNI formulations). A method can top one board and fail the other: that split is the point, not a contradiction. Antipodal cos(+1, -1) is logged on both boards and gated on neither — perfect antipodal lock coexists with walking off the sheet (#22), so the gates stay audible (continuation / cover / hold).
Related cells (not this scale)
- FORMULATION_LEADERBOARD_BIPOLAR.md — the bipolar sibling.
- lm-plus-neu-exam.md — the plus+neu exam this board reuses.
- lm-plus-exam.md — plus-only cover / off-caption.
- lm-2d-scoreboard.md — compiled bipolar board.
How to run
PYTHONPATH=. python analysis/slider2d/run_formulation_leaderboard.py --polarity uni --out docs/formulation-leaderboard
PYTHONPATH=. pytest tests/test_formulation_leaderboard.py -q
CPU only. No Hub, no GPU, no Music 3 weights. Seed 0, 400 Adam steps.