[martyn]

Summary

Introduces and tunes Relativistic Hellinger-Born-Infeld Transport (rhbi), a clock-free, constant-learning-rate ParticleGAN family (rhbi-v1 → rhbi-v2 → rhbi-v3) with zero EMA critic copy, zero schedule annealing, and zero task-shape or step-count gates.

Under one unchanged global recipe (rhbi-v3, protocol seed 0, deterministic_orthogonal), RHBI achieves:

• 27/27 (100%) Tier 1 view requirements — the only family across all 9 registered families with 100% Tier 1 passes.
• 49/78 Tier 2 view requirements (tying BCAP 49/78) for 76/115 total view requirements (Publish sparse study and correct gradient ablation and purity metrics #2 overall behind stale BISPF 81/115, and ahead of BCAP 71/115, GAN v3 20/115, K3P 20/115, KA2 20/115, R1/R2 20/115, Atlas 18/115, and E22 0/115).
• 21 full-suite PASSes (0 errors, 0 incomplete) on the fixed 61-test roster — Declare experiment dependencies and add CPU test CI #1 across all 9 families in the repository (ahead of BISPF 18 and BCAP 16).
• 54/54 uncontended (clear) DCGAN scaling cells on NVIDIA RTX A6000 (34.40 ms / 451.39 img/s on r32 — 8.1× faster than BCAP 279.86 ms).

Up-to-date leaderboard: reports/forge/technique-inventory.md · Family report: reports/forge/families/rhbi.md · Narrative: reports/forge/rhbi/README.md

---

Formulation & Progression (rhbi-v1 → rhbi-v2 → rhbi-v3)

Candidate	Source Commit	Readout	Tier 1	Tier 2	Total View Reqs	Full-Suite PASS	Key Formulation Delta
rhbi-v1	d2abe7afb	readout-a94316e3e91606df03a346de	23/27	0/78	23/115	8	Initial $(B+1)\times(B+1)$ slack-screened Sinkhorn (hollow self-coupling, $\ell_t^2$-scaled physical energy, quadratic R1/R2 penalty).
rhbi-v2	b9e33a00e	readout-44a71042e8d9d2018cfd677e	27/27	37/78	64/115	19	Physical Born-Infeld cost $c_{\mathrm{phys}} = \sqrt{\ell_t^2 + \lVert\Delta\rVert_2^2} - \ell_t$ (eliminating $\ell_t\to 0$ gradient vanishing), linear screening ($\rho=3.0$), Feydy diagonal self-coupling debiasing, smooth $C^\infty$ Born-Infeld endpoint critic cap ($\kappa=1.0, \beta=20, \lambda_{\mathrm{floor}}=0.005$), and DualNorm (lr=0.014, d_lr_mult=1.5, prior_lr_mult=2.9, prior_regularizer="sbnp", prior_reg=0.02, constraint_geometry_mode="direction_blend").
rhbi-v3	eeb698a7b	readout-2e0a16b47298ebbc5e33ba63	27/27	49/78	76/115	21	Fibre-centered conditional relativistic slack routing (rhbi_conditional_fibre="slack"): when continuous conditioning $c$ accompanies $D(c, x)$, distinct draws $c_i \neq c_j$ lie on disjoint fibres where batch-wide $\bar D$ is contaminated by inter-fibre baseline variance $V(c)$ and Euclidean transport collapses conditional variation; paired fibre log-odds $\operatorname{softplus}(D(c_i, y_i) - D(c_i, x_i))$ cancels $V(c_i)$ identically, unlocking trajectory (identity_mse=0.000268 <= 0.02) and residual_student (identity_mse=0.000265 <= 0.02, success_rate=1.0). Also detects stateful batch-bound critic view wrappers in RHBICriticPenalty so 2B concatenation never mismatches cached B-row masks on MisGAN hosts.

---

Selected-Family Leaderboard Snapshot

Family / Selected Recipe	Tier 1	Tier 2	Tier 3	Total View Reqs	61-Test Full-Suite PASS	DCGAN r32 (img/s)
BISPF (bispf-v2, stale)	22/27	55/78	4/10	81/115	18	539.27
RHBI (rhbi-v3, current)	27/27	49/78	0/10	76/115	21	451.39
BCAP (bcap-pure-adam-v2 / dualnorm, stale)	22/27	49/78	0/10	71/115	16	57.00
GAN v3 (stale)	20/27	0/78	0/10	20/115	5	1,611.35
K3P (stale)	20/27	0/78	0/10	20/115	5	1,249.72
KA2 (stale)	20/27	0/78	0/10	20/115	5	1,239.64
R1/R2 (stale)	20/27	0/78	0/10	20/115	5	1,620.39
Atlas (stale)	18/27	0/78	0/10	18/115	8	626.53
E22	0/27	0/78	0/10	0/115	0	638.36

---

Verification

• python -m experiments.forge validate → valid
• python -m experiments.forge compile --check → CURRENT (11835/11835 tracked)
• python ops/refresh_forge_readouts.py --check --reports-only → CURRENT (stale stages: none) (54/54 DCGAN cells COMPLETE + clear)
• python scripts/check_lint.py → 0 new findings
• pytest across tests/test_rhbi.py, bitwise historical compatibility (tests/test_bcap_component_source_cohorts.py, tests/test_bcap_integration_compatibility.py), publication/scaling/schema suites (410/410 passed)