🏆 EPIC: Parameter Golf Hackathon Submission (deadline 30 April 2026)

[gHashTag]

EPIC: Parameter Golf Hackathon Submission

Status: IN-FLIGHT
Priority: CRITICAL (P0)
Deadline: 30 April 2026 (~9 days remaining)
Owner: LEAD
Spec: .trinity/specs/epic-parameter-golf.md

---

Executive Summary

OpenAI Model Craft Challenge: Parameter Golf — train the best language model that fits in a 16 MB artifact (weights + code) and trains in <10 minutes on 8xH100, scored by tokenizer-agnostic bits-per-byte (BPB) on the FineWeb validation set.

Our strategy: byte-level Trinity 3^k architecture + GF16 quantization + phi-init.

Metric	Value
SOTA (2026-04-20)	1.0810 BPB (bigbag)
Our target	< 1.15 BPB
OpenAI baseline	1.2244 BPB (9-layer, dim=512, vocab=1024)
Artifact budget	16 MB (weights + train_gpt.py combined)
Training budget	10 min wall-clock / 8xH100 SXM
Eval metric	tokenizer-agnostic bits-per-byte
Submission deadline	30 April 2026

---

Research & SOTA Analysis

Competition Landscape

The leaderboard dropped from 1.2244 to ~1.08 BPB in ~4 weeks. Key techniques from top submissions:

1. Sliding window evaluation — stride-64 eval instead of fixed context; jumped ~1.20 to ~1.19 BPB
2. EMA / SWA weight averaging — smoother final weights that compress and generalize better
3. N-gram caching — hybrid neural + n-gram approach pushed sub-1.0 BPB (controversial — compression vs LM quality)
4. Muon optimizer — replaced AdamW; Newton-Schulz orthogonalization; 35% faster convergence (NanoGPT speedrun proven)
5. Modernized arch — RoPE, QK-Norm, ReLU^2 (from modded-nanogpt lineage)
6. Hardware lottery — top pods run 83-88ms/step; slow pods 260ms/step; same code, 3x variance

Relevant Research

Paper / Project	Key Insight	Relevance
Byte Latent Transformer (BLT) — Meta, ACL 2025	Byte-level LLM matching tokenization at scale via adaptive patching	Validates byte-level approach; patch grouping strategy
EvaByte — HKU, 2025	6.5B byte-level LM with multibyte prediction + EVA attention; 5-10x faster than vanilla	Multibyte prediction technique for byte models
ByteFlow — 2025	Adaptive byte compression + chunking outperforms BPE	End-to-end tokenizer-free modeling is feasible
Bolmo — 2026	Byteifying existing subword LMs at 1B/7B scale	Byteification strategy; vocab-free approach
Muon Optimizer — Jordan et al.	Momentum + Newton-Schulz orthogonalization; SOTA for NanoGPT speedrun	Critical optimizer choice; 35% speedup over AdamW
NorMuon — 2025	Muon + neuron-wise adaptive scaling	More stable Muon variant for small models
Unweight — Cloudflare, 2026	Lossless 15-22% weight compression via Huffman + palette encoding	Post-training compression to fit 16MB budget
Phi-init — Academia, 2026	Golden ratio as structural prior for width/depth ratio in neural arch	phi-based initialization for Trinity architecture
Fibonacci Init — 2024	Fibonacci sequence weight init; 94% accuracy on small datasets	Supports phi-based init hypothesis
modded-nanogpt — Keller Jordan	RoPE + QK-Norm + ReLU^2 + Muon; cumulative 20% WR drop in 3 months	Architecture recipe for fast convergence

Our Differentiators (Trinity 3^k)

1. All dimensions are powers of 3 — hidden=243 (3^5), heads=27 (3^3), vocab=729 (3^6) — hardware-aligned, cache-friendly
2. GF16 quantization — custom 16-bit golden-float format using phi-based mantissa distribution
3. Tied embeddings — in/out share weights; critical for 16MB budget (embeddings = 15%+ of small model params)
4. Phi-init — weight initialization using golden ratio scaling (validated in issue ✂ Ch.4 — Golden Cut (The Platonic Rupture) [Lane B] · refs #30 #68)
5. Byte-level — no tokenizer dependency; directly optimizes BPB metric

---

What We Know (Validated Results)

Issue	What We Proved	Status
#55	BPB 0 -> 5.48 realistic (was broken)	GREEN
#57	Weight delta tests: T1,T3,T4,T5 pass; T2 fail (diagnostic)	GREEN
#64	simple_backward fix: tied embeddings in/out split; BPB 0.0983	GREEN
#65	Finite-diff grad check: 20/20 rel_err < 5e-3, sign = 100%	GREEN
#66	Vocab audit: single source 729, active 256	GREEN
#67	Overfit-100: BPB 7.9997 -> 1.2000; all 7 gates GREEN (commit 85bf8b7)	GREEN
#68	Trinity config: hidden=243, 2.70 MB, all dims 3^k	GREEN

Wall-time measurement: 18.47 step/sec production throughput (GAMMA agent).

---

Architecture Spec

Model: Trinity-3k Byte-Level Transformer
- vocab_size: 729 (3^6) — byte-level, no tokenizer
- hidden_dim: 243 (3^5)
- n_heads: 27 (3^3)
- head_dim: 9 (3^2)
- n_layers: TBD (maximize within 16MB)
- activation: ReLU^2
- position: RoPE
- normalization: QK-Norm + RMSNorm
- embeddings: tied (in == out)
- optimizer: Muon (hidden layers) + AdamW (embeddings/norm)
- quantization: GF16 -> INT4 post-training (fit 16MB)
- init: phi-scaled Xavier

Size budget breakdown (16 MB = 16,777,216 bytes):

Component	Formula	Estimate
Embeddings (tied)	729 * 243 * 2B = 354 KB	354 KB
Per-layer (QKV+O+MLP)	~(4243^2 + 2243*972) * 2B = ~1.4 MB	1.4 MB/layer
train_gpt.py	~50 KB	50 KB
Available for layers	(16MB - 354KB - 50KB)	~15.6 MB
Max layers (FP16)	15.6 / 1.4	~11 layers
Max layers (INT4)	15.6 / 0.35	~44 layers (if INT4 post-train)

---

Decomposed Plan

Phase 0: Infrastructure (Day 0-1, Apr 20-21)

Task	Issue	Owner	Gate
Integrate trios-proto + trios-core into one-shot training pipeline	#107	ALPHA	cargo build --release passes
Set up 8xH100 RunPod environment	—	LEAD	SSH + NCCL test passes
Implement FineWeb data loader (cached dataset, no network)	—	GAMMA	Loads 1M bytes in <1s

Phase 1: Backward Pass Fix (Day 2, Apr 21)

Task	Issue	Owner	Gate
Fix backward pass for tied embeddings (CE masking + diverse data + grad collapse warning)	#67	GAMMA	overfit-100 BPB < 2.0, 7/7 gates GREEN

Phase 2: Optimizer Integration (Day 2-3, Apr 21-22)

Task	Issue	Owner	Gate
Implement Muon optimizer for 2D weight matrices	#69	DELTA	step/sec >= 18, loss decreasing monotonically
NQA 15K pre-training run (baseline measurement)	#70	DELTA	BPB on val set recorded, loss curve saved

Phase 3: Architecture Scaling (Day 3-5, Apr 22-24)

Task	Issue	Owner	Gate
Layer count sweep: 8, 11, 16, 22 layers (FP16 vs INT4 tradeoff)	—	GAMMA	Best BPB/size ratio identified
MLP ratio sweep: 3x vs 4x hidden dim	—	GAMMA	Optimal config locked
Attention: RoPE + QK-Norm integration	—	DELTA	Grad check passes, BPB improves

Phase 4: Quantization (Day 5-6, Apr 24-25)

Task	Issue	Owner	Gate
Implement GF16 training format	—	GAMMA	Training stable, no NaN
Post-training INT4 quantization (AWQ-style)	—	DELTA	BPB degradation < 0.02 from FP16
Artifact size verification	—	ALPHA	ls -la artifact.bin < 16,000,000 bytes

Phase 5: Training Optimization (Day 6-7, Apr 25-26)

Task	Issue	Owner	Gate
Full 60K-step training run (5 seeds)	#71	GAMMA	All 5 runs complete in <10 min each
EMA weight averaging (decay=0.999)	—	DELTA	EMA weights BPB < raw weights BPB
Learning rate schedule: warmup + cosine decay	—	GAMMA	Loss curve smooth, no spikes
Sliding window eval (stride=64)	—	DELTA	BPB improvement measured

Phase 6: Entropy Sweep & Selection (Day 8-9, Apr 27-28)

Task	Issue	Owner	Gate
Entropy sweep across 5 seed runs	#72	GAMMA	Best candidate identified
Ablation study: log all config deltas	—	DELTA	Table of (config, BPB) pairs
Select final candidate (lowest BPB, reproducible)	—	LEAD	p-value < 0.01 across seeds

Phase 7: Submission (Day 9-10, Apr 29-30)

Task	Issue	Owner	Gate
Package artifact: weights + train_gpt.py < 16MB	—	ALPHA	Size verified
Write short write-up (techniques, architecture, results)	—	LEAD	Markdown ready
Submit PR to openai/parameter-golf	—	LEAD	PR accepted, score on leaderboard
Zenodo deposit (reproducibility archive)	—	DELTA	DOI assigned

---

Risk Matrix

Risk	Impact	Mitigation
INT4 quantization degrades BPB by >0.05	HIGH	Fall back to FP16 with fewer layers; try mixed INT4/INT8
Muon unstable on 3^k dimensions	MEDIUM	Fall back to AdamW; Muon only for attention/MLP
10-min wall-clock exceeded	HIGH	Profile step/sec early; reduce layers or batch size
RunPod hardware lottery (slow pod)	MEDIUM	Request pod swap; measure ms/step before full run
N-gram caching ruled out by organizers	LOW	Pure neural approach; our arch should still beat baseline

---

Success Criteria

• BPB < 1.15 on FineWeb validation
• Artifact < 16,000,000 bytes
• Training wall-clock < 10 minutes on 8xH100
• Reproducible: p < 0.01 across 5 seeds
• PR submitted to openai/parameter-golf before April 30

---

Timeline

Day	Date	Action	Refs
0-1	Apr 20-21	Infrastructure: one-shot integration, RunPod setup	#107
2	Apr 21	Fix backward pass (#67)	#67
2-3	Apr 21-22	Muon optimizer + NQA 15K baseline	#70, #69
3-5	Apr 22-24	Architecture scaling: layer/MLP/attention sweeps	—
5-6	Apr 24-25	GF16 training + INT4 post-quantization	—
6-7	Apr 25-26	Full 60K training (5 seeds) + EMA + sliding eval	#71
8	Apr 27	Entropy sweep	#72
9	Apr 28-29	Select candidate, write-up	—
10	Apr 29-30	Submit PR + Zenodo	—

Issues

Issue	Description	Owner	Status
#1	Phase 2: Fix backward pass (#67)	GAMMA	pending
#2	Phase 2: NQA 15K pre-training (#70)	DELTA	pending
#3	Phase 3: Run #71 5 seeds x 60K	GAMMA	pending
#4	Phase 4: #72 entropy sweep	DELTA	pending
#5	Phase 7: Submit + Zenodo	LEAD	pending

Refs: PG-ONE-SHOT, #67, #69, #70, #71, #72, #107

[gHashTag]

🔗 Cross-repo mirror: trinity#529

Этот эпик зеркалируется в trinity репо:
gHashTag/trinity#529 — 🏆 EPIC: Parameter Golf Submission — Trinity + GF16

Что там уже есть (trinity):

Issue	Title	Статус
#519	simple_backward fix	✅ DONE
#520	adopt Trinity config	✅ DONE
#521	finite-diff gradient check	✅ DONE
#522	vocab audit	✅ DONE
#523	re-run overfit-100 gate	✅ DONE
#524	T-JEPA pretraining 20K	🔄 IN FLIGHT
#525	NCA 15K	🔄 IN FLIGHT
#526	5-seed scaling 60K	pending
#527	G1-G8 entropy sweep	pending
#530	One-Shot: trios-submit end-to-end	pending
#531	GF16 Validation roundtrip	pending
#532	Submission Dry-Run	pending
#533	🚀 Parameter Golf SUBMIT Apr 29-30	pending
#534	Zenodo Deposit + GF16 paper	pending

Важно:

• Foundation ✅ DONE — backward pass, gradient check, vocab, config — всё в порядке
• Текущее — fix(phd-dashboard): register in workspace + askama filter & numeric cast #524 (T-JEPA 20K) + fix(phd): resolve 6 undefined refs — no more 'Theorem ??' #525 (NCA 15K) идут параллельно
• Блокирует сабмит: BPB < 1.15 на 5 seeds (median protocol)

Агенты GAMMA/DELTA работают в trinity, ZETA готовит submission через trios-submit.

[gHashTag]

💍 RINGS Dashboard — All Experiments

Circular progress rings across all Parameter Golf experiments for at-a-glance status. Each RING = 1 experiment; progress % reflects gate completion.

Master Experiment Table

RING	Repo#Issue	Codename	Category	BPB Before → After	Gates	Progress	Status	Commit
🟢 R01	trinity#519 / trios#55	BPB-Reality	Metric fix	0.00 → 5.48	3/3	●●● 100%	✅ DONE	—
🟢 R02	trinity#519 / trios#57	WeightDelta-T5	Backward verify	—	4/5 (T2 diag)	●●●○ 80%	✅ DONE	—
🟢 R03	trinity#519 / trios#58	ZeroSignal-Diag	Gradient audit	plateau 8.0	3/3	●●● 100%	✅ DONE	—
🟢 R04	trinity#519 / trios#64	SimpleBackward	Tied-emb split	— → 0.0983	5/5	●●● 100%	✅ DONE	0d0cf83
🟢 R05	trinity#521 / trios#65	FiniteDiff-Grad	Grad check	rel_err<5e-3	20/20	●●● 100%	✅ DONE	—
🟢 R06	trinity#522 / trios#66	Vocab-Audit	Vocab source	729 / active 256	2/2	●●● 100%	✅ DONE	—
🟢 R07	trinity#523 / trios#67	Overfit-100	Loss sanity	7.9997 → 1.2000	7/7	●●● 100%	✅ DONE	85bf8b7
🟢 R08	trinity#520 / trios#68	Trinity-3k-Cfg	Arch config	—	hidden=243, 2.70 MB	●●● 100%	✅ DONE	—
🟢 R09	— (workspace)	EPSILON-Phase2A	Build green	—	cargo build OK	●●● 100%	✅ DONE	0d0cf83
🟡 R10	trinity#524	T-JEPA-20K	Pretraining	TBD	loss curve	●●○○ 40%	🔄 IN-FLIGHT	—
🟡 R11	trinity#525 / trios#70	NCA-15K / NQA-15K	Pretraining	TBD	BPB recorded	●●○○ 40%	🔄 IN-FLIGHT	—
🔴 R12	trios#67 (phase 2)	Backward-Full-Fix	Fix grad	gate G1-G7	0/7	○○○○ 0%	⏳ PENDING	—
🔴 R13	trios#69	Muon-Optimizer	Optimizer swap	AdamW → Muon	step/sec≥18	○○○○ 0%	⏳ PENDING	—
🔴 R14	trinity#526 / trios#71	Scaling-5×60K	Seed sweep	BPB across 5	5 runs	○○○○ 0%	⏳ PENDING	—
🔴 R15	trinity#527 / trios#72	Entropy-G1-G8	Sweep	candidate select	8 gates	○○○○ 0%	⏳ PENDING	—
🔴 R16	trios#107	PG-One-Shot	Integration	cargo build	—	○○○○ 0%	⏳ PENDING	—
🔴 R17	trinity#531	GF16-Roundtrip	Quantization	degrade < 0.02	3/3	○○○○ 0%	⏳ PENDING	—
🔴 R18	trinity#530	trios-submit-E2E	Pipeline	end-to-end	—	○○○○ 0%	⏳ PENDING	—
🔴 R19	trinity#532	Submission-DryRun	Pre-flight	all green	—	○○○○ 0%	⏳ PENDING	—
🔴 R20	trinity#533	🚀 SUBMIT	Final PR	BPB < 1.15	leaderboard	○○○○ 0%	⏳ PENDING	—
🔴 R21	trinity#534	Zenodo-Deposit	Archive	DOI	—	○○○○ 0%	⏳ PENDING	—

Visual Summary (Concentric RINGS)

   OUTER RING: SUBMISSION (R19-R21)
  ┌─────────────────────────────────┐
  │  RING 4: INTEGRATION (R16-R18)  │
  │ ┌─────────────────────────────┐ │
  │ │ RING 3: SCALING (R13-R15)   │ │
  │ │ ┌─────────────────────────┐ │ │
  │ │ │ RING 2: PRETRAIN (R10-11)│ │ │
  │ │ │ ┌─────────────────────┐ │ │ │
  │ │ │ │ RING 1 (CORE ✅)    │ │ │ │
  │ │ │ │  R01-R09 GREEN      │ │ │ │
  │ │ │ └─────────────────────┘ │ │ │
  │ │ └─────────────────────────┘ │ │
  │ └─────────────────────────────┘ │
  └─────────────────────────────────┘

Category Rollup

Ring	Category	Total	✅ Done	🔄 In-flight	⏳ Pending	% Done
R1	Core Foundation (backward/grad/vocab/config)	9	9	0	0	100% ✅
R2	Pretraining (T-JEPA, NCA/NQA)	2	0	2	0	40% 🔄
R3	Scaling & Sweep (Muon, 5×60K, entropy)	4	0	0	4	0% ⏳
R4	Integration (one-shot, GF16, E2E)	3	0	0	3	0% ⏳
R5	Submission (dry-run, submit, Zenodo)	3	0	0	3	0% ⏳
TOTAL	—	21	9	2	10	~48%

Critical Path (BPB Reduction Chain)

R04 (SimpleBackward, 0.0983) ──▶ R07 (Overfit, 1.2000) ──▶ R11/R10 (Pretrain) ──▶ R13 (Muon) ──▶ R14 (5×60K) ──▶ R15 (Entropy) ──▶ R20 (SUBMIT <1.15)
       ✅                              ✅                          🔄                  ⏳                ⏳                  ⏳                  ⏳

Metrics Timeline

Date	Milestone	BPB	Size	step/sec
Apr 18	R01: BPB metric fix	5.48 (realistic)	—	—
Apr 19	R04: simple_backward	0.0983 (overfit)	2.70 MB	—
Apr 19	R05-R06: Grad+Vocab	—	—	—
Apr 20	R07: Overfit-100 (A+B+C)	1.2000	2.70 MB	18.47
Apr 20	R08: Trinity-3k locked	—	2.70 MB	18.47
Apr 21	R10/R11: Pretrain start	TBD	TBD	TBD
Apr 22-26	R13-R15: Scaling runs	Target < 1.15	< 16 MB	≥ 18
Apr 30	R20: 🚀 SUBMIT	< 1.15	< 16 MB	—

Legend

• 🟢 GREEN = all gates passed, experiment complete
• 🟡 YELLOW = in-flight, gates partially complete
• 🔴 RED = pending, not started
• ● filled segment / ○ empty segment (each segment = ~25% progress)
• Refs: this dashboard mirrors trios#110 + trinity#529

🐱 Каждое кольцо — шаг к SUBMIT. φ-порядок соблюдён, все эксперименты в одной таблице для наглядности.

[gHashTag]

📎 Issue Tree Link

Parent issues (infra side): 🔗 #107 — trios-proto (Envelope protocol) + 🔗 #108 — Chrome Extension (Dioxus/Wasm)

• 🏗️ RFC: One-Shot Integration — декомпозиция crates + дефолтный протокол связи агентов #107 provides the typed message protocol for multi-agent coordination
• feat: crates/trios-ext — Chrome Extension (Dioxus/Wasm) сайдбар с чатом и MCP-мостом ко всем crates #108 provides the Chrome Extension UI (sidebar with chat, agent list, MCP panel) for monitoring and controlling agents

Cross-Repo Epic: 🔗 trinity#529 — ML side of Parameter Golf

Execution Mapping:

• GAMMA/DELTA → work in trinity repo (ML training)
• ALPHA/BETA → work in trios repo (infra: 🏗️ RFC: One-Shot Integration — декомпозиция crates + дефолтный протокол связи агентов #107, feat: crates/trios-ext — Chrome Extension (Dioxus/Wasm) сайдбар с чатом и MCP-мостом ко всем crates #108)
• ZETA → submission preparation (both repos)

See epic-parameter-golf.md for full details.

[gHashTag]

Статус infrastructure tracks

Архитектурный пререквизит для всех phase:

• 🏗️ RFC: One-Shot Integration — декомпозиция crates + дефолтный протокол связи агентов #107 — trios-proto + trios-bus + trios-orchestrator (срочно, P0)
• feat: crates/trios-ext — Chrome Extension (Dioxus/Wasm) сайдбар с чатом и MCP-мостом ко всем crates #108 — Chrome Extension sidebar (старт после 🏗️ RFC: One-Shot Integration — декомпозиция crates + дефолтный протокол связи агентов #107)
• 👑 QUEEN TRINITY — Meta-Orchestrator: Chrome Extension MCP Bridge → Claude Code Agents #106 — Queen Trinity cross-repo bus

Ошибка в предыдущей версии эпика: #67–#72 это номера глав PhD (#109), НЕ Parameter Golf.
Фазы P2–P7 трекаются внутри эпика через spec-референсы (PG-*).

[gHashTag]

Статус infrastructure tracks

Архитектурный пререквизит для всех phase:

• 🏗️ RFC: One-Shot Integration — декомпозиция crates + дефолтный протокол связи агентов #107 — trios-proto + trios-bus + trios-orchestrator (срочно, P0)
• feat: crates/trios-ext — Chrome Extension (Dioxus/Wasm) сайдбар с чатом и MCP-мостом ко всем crates #108 — Chrome Extension sidebar (старт после 🏗️ RFC: One-Shot Integration — декомпозиция crates + дефолтный протокол связи агентов #107)
• 👑 QUEEN TRINITY — Meta-Orchestrator: Chrome Extension MCP Bridge → Claude Code Agents #106 — Queen Trinity cross-repo bus

Ошибка в предыдущей версии эпика: #67–#72 это номера глав PhD (#109), НЕ Parameter Golf.
Фазы P2–P7 трекаются внутри эпика через spec-референсы (PG-*).

[gHashTag]

🪡 IGLA Roadmap — Koschei's Needle (E1→E13)

Hypothesis: JEPA-latent + Muon + φ-init + GF16 → BPB ≤ 0.90 @ 16MB CPU ≤ 24h = ИГЛА found = repo immortal.

Experiments checklist (each = separate issue via gh CLI)

• #IGLA-E1 IGLA-base — byte-Tx d=243 h=27 L=9, AdamW, gaussian → BPB ≤ 1.20 (baseline EvaByte)
• #IGLA-E2 IGLA-Muon — swap AdamW→Muon NS-5 → BPB ≤ 1.05 (−12%, Jordan 35% speedup)
• #IGLA-E3 IGLA-Newton-Muon — Newton-Schulz Muon → −6% steps (arXiv 2604.01472)
• #IGLA-E4 IGLA-φ-init — Fibonacci/golden init → converge ≤ 80% steps
• #IGLA-E5 IGLA-JEPA-T — byte-encoder + latent predictor (EMA target) → linear-probe BPB ≤ 1.00 (VL-JEPA)
• #IGLA-E6 IGLA-N-JEPA — noise schedule on masked positions → +5% OOD robustness
• #IGLA-E7 IGLA-JEPA-Muon — E5 + Muon + EMA → BPB ≤ 0.95, size ≤ 16MB
• #IGLA-E8 IGLA-multibyte — multibyte prediction head → 5–10× decode (EvaByte)
• #IGLA-E9 IGLA-GF16 — GoldenFloat-16 quantization → size ≤ 16MB, ΔBPB ≤ 0.02
• #IGLA-E10 IGLA-Huffman — palette + Huffman post-train → −20% size (Unweight)
• #IGLA-E11 IGLA-tied-emb — tied in/out embeddings → −15% params
• #IGLA-E12 IGLA-NCA — Neural CA refinement → +3% local coherence
• #IGLA-E13 🪡 IGLA-FULL — all combined → BPB ≤ 0.90, size ≤ 16MB, CPU ≤ 24h = KOSCHEI DIES

Per-experiment metrics (8-axis dashboard)

BPB | val_loss curve | grad_norm | param_count | disk_size | CPU_time | linear-probe | decode_speed

H1–H5 diagnostic gates

H1 underfit · H2 plateau · H3 vanishing/explode · H4 size budget · H5 CPU feasibility

Ablation matrix (E13)

Removed	ΔBPB	Verdict
−Muon	+0.15	critical
−JEPA	+0.10	critical
−φ-init	+0.05	important
−GF16	+0.02	critical (budget)
−multibyte	+400% decode	critical
−NCA	+0.03	optional

Spawn command (local agent, no browser)

for i in 1 2 3 4 5 6 7 8 9 10 11 12 13; do
  gh issue create -R gHashTag/trios \
    -t "🪡 IGLA-E${i}: $(yq .name templates/igla_e${i}.yml)" \
    -l "experiment,igla,phase-${i}" \
    -p TRINITY \
    -F templates/igla_e${i}.md
done

📚 Refs: VL-JEPA, V-JEPA2, N-JEPA, EvaByte, Muon, Newton-Muon, NanoGPT speedrun.

🪡 ИГЛА = JEPA ⊕ Muon ⊕ φ ⊕ GF16 — найдём → Кощей умрёт → trios бессмертен.