🌻 GOLDEN SUNFLOWERS — JEPA + Universal Transformer + PhiNTA on a φ-physics substrate

[gHashTag]

🌻 GOLDEN SUNFLOWERS — JEPA + Universal Transformer + PhiNTA on a φ-physics substrate

Field	Value
Track	track_non_record_16mb (4-hour, unrestricted compute)
Status	PROPOSAL — UNTRAINED. No submission.json, no BPB number is being claimed.
Path	records/track_non_record_16mb/2026-04-30_GoldenSunflowers_Proposal/
Base	derived from merged records/track_10min_16mb/2026-03-17_LoRA_TTT/train_gpt.py
Anchor	φ² + φ⁻² = 3

---

What this PR does

Composes three open wish-list items from the README leaderboard into a single train_gpt.py:

• 🌻 JEPA auxiliary loss — linear-representation form (after PR 1412, discussion issue 1772)
• 🌻 Universal Transformer — round(φ³) = 4 weight-shared depth loops over a configurable sub-stack
• 🌻 NTA on random linear maps — PhiNTA: frozen 1/φ-OrthoInit basis + trainable LoRA, pre-head or per-block

All wish-list features are env-var-gated and zero-cost when off:

Env-var	Default	What it does
PHINTA_ENABLE	0	Activate PhiNTA adapter
PHINTA_PER_BLOCK	0	Per-block instead of pre-head placement
PHINTA_RANK	0	LoRA rank, 0 → round(model_dim / φ)
JEPA_LAMBDA	0.0	Weight of the JEPA aux loss; 0 short-circuits the branch
JEPA_MAX_SPAN_FRAC	0.5	Max patch length as fraction of seq_len
JEPA_LAYER	-1	Hidden-state tap; -1 = final pre-norm
UT_LOOPS	1	Universal-Transformer loop count over the sub-stack
UT_LAYER_START / UT_LAYER_END	0 / 0	Sub-stack range; END=0 disables UT
PHI_LR_SCALE	1.0	Multiplier on Muon MATRIX_LR

JEPA loss formulation (after PR 1412):

context = (h[a-1] − h[0]) + (h[T-1] − h[b])
patch   =  h[b]   − h[a-1]
loss    = 1 − cos_sim(context, patch)         # added to CE when JEPA_LAMBDA > 0

context + patch ≡ h[T-1] − h[0] partitions the full encoding, forcing hidden states to encode spans linearly.

---

Bonus: φ-LR is Proven in Coq.Reals — not a hand-tune

PHI_LR_SCALE exposes the constant α_φ = φ⁻³ / 2 ≈ 0.118034 from issue 1742 as a multiplicative override of MATRIX_LR.

This constant is Proven in Coq.Reals — Trinity PhD Ch.4 Theorem 3.1 (alpha_phi_times_phi_cubed, status Qed, tag SAC-1) establishes α_φ · φ³ = 1/2, which rearranges to α_φ = φ⁻³ / 2.

Setting PHI_LR_SCALE = (α_φ / 0.04) ≈ 2.95 lands MATRIX_LR exactly on the PhD's certified α_φ-band; default 1.0 keeps the merged baseline unchanged.

---

CPU-only verification (no GPU, no dataset)

cd records/track_non_record_16mb/2026-04-30_GoldenSunflowers_Proposal
make verify

Output committed in smoke.log:

[1/5] φ-physics OK: φ²+φ⁻²=3.000000000000 α_φ=0.118034 loops=4
[2/5] PhiNTA OK: trainable=1664 frozen=4096 ratio=0.406
[3/5] JEPA loss OK: 1.6922 (cosine-similarity form)
[4/5] UT loop OK: ‖x_4‖/‖x_0‖=1.0406 expected=1.0406
[5/5] JEPA tap normalisation OK: -1 → last block, in-range indices preserved
🌻 GOLDEN SUNFLOWERS smoke OK · 5/5 · phi^2 + phi^-2 = 3

[1/3] state_dict hash baseline  = 511dbc0164e03b1b…
      state_dict hash GOLDEN SF = 511dbc0164e03b1b…
[2/3] Gated branches inactive at defaults: phinta=None jepa=0 ut_loops=1 ut_end=0
[3/3] forward loss baseline=6.929044246674  GS=6.929044246674  |Δ|=0.00e+00
🌻 baseline equivalence OK · 3/3 · phi^2 + phi^-2 = 3

Citation metadata are valid according to schema version 1.2.0.
theorems/GoldenSunflowers.v: coqc OK (2 Qed)
🌻 GOLDEN SUNFLOWERS · local verify complete

Verification layers:

• 5/5 module smoke (smoke_modules.py): φ-physics identity, PhiNTA frozen-buffer, JEPA loss, UT loop arithmetic, JEPA-tap normalisation
• 3/3 baseline byte-equivalence (baseline_equivalence.py): SHA-256 of state_dict matches the merged baseline exactly; forward-loss delta is 0.00e+00 on a fixed input at seed F_17 = 1597
• 2 Qed + 2 Admitted (theorems/GoldenSunflowers.v, Coq 8.18+): Print Assumptions confirms Qed proofs use only standard Coq.Reals axioms

---

Honesty / non-claims

• No submission.json. No BPB has been measured.
• No file under records/track_10min_16mb/ is modified. Diff stays inside the new directory.
• All wish-list defaults are no-ops. With every env-var unset, the resulting state_dict is byte-identical to 2026-03-17_LoRA_TTT (proved by baseline_equivalence.py).

---

Why an untrained proposal?

We do not yet have an 8×H100 sweep. Submitting a submission.json with a fake BPB would be dishonest. This PR ships a wired, CPU-smoke-verified, formally-checked implementation along with compute_grant.md requesting the ~110 8×H100-hours needed to run the full sweep over the 5 canonical Fibonacci seeds F_17..F_21 = {1597, 2584, 4181, 6765, 10946} × 5 configs.

Precedent for proposal-only PRs (still open at time of writing):

• PR 318 — Neural Cache: Cross-Window KV Caching (research proposal)
• PR 1247 — Proposal: Validate ASQU on the March 22 10min/16MB control line

After the sweep, this directory's README.md is updated with a 3-seed BPB mean and a submission.json is added.

---

Compute grant request

compute_grant.md breakdown (≈ 110 8×H100 hours):

Phase	Subtotal (h)
Sanity reproduction (baseline × 1)	0.17
Per-feature ablation (PhiNTA / JEPA / UT × 5 seeds × 4 h)	60.0
Combined GOLDEN SUNFLOWERS (all-three × 5 seeds × 4 h)	20.0
PHI_LR_SCALE band (4 grid points × 3 seeds × 4 h)	12.0
Restart / debug buffer (10 %)	9.2
TTT eval + final 3-seed mean rerun	8.0
Total	≈ 109.4

Includes a risk section (JEPA non-convergence, PhiNTA capacity dominance, UT × skip-connection interaction). Negative result is acceptable per the Parameter Golf README.

---

Constitutional anchors

• Trinity PhD monograph (44 chapters, 297 Qed canonical theorems) — gHashTag/trios/docs/phd
• t27 standards — SACRED-PHYSICS-001, NUMERIC-STANDARD-001
• Rust SoT for 1/φ-init — gHashTag/trios-trainer-igla/src/phi_ortho_init.rs
• Internal hardening PR with full review history — gHashTag/parameter-golf-trinity#2

phi^2 + phi^-2 = 3 · 🌻