| name | igla |
| description | This skill should be used when user asks to "update IGLA strategy", "check IGLA race status", "monitor workers", "deploy IGLA workers", "IGLA competition report", or references "IGLA RACE" numbers × models grid. Manages IGLA RACE competition (12 numeric formats × 6 models) monitoring and worker coordination. |
| version | 0.1.0 |
IGLA RACE — Competition Management Skill
IGLA RACE is a two-axis competition format: 12 numeric formats × 6 model types = 72 minimum cells.
Competition Structure
Numbers (12 participants)
| Lane# | Format | Bits | EXP:MANT | PHI_BIAS | Status |
|---|
| N1 | GF4 | 4 | 1:2 | 0 (F₀) | extract-only |
| N2 | GF8 | 8 | 3:4 | 1 (L₁) | extract-only |
| N3 | GF12 | 12 | 4:7 | 2 (L₀) | extract-only |
| N4 | GF16 | 16 | 6:9 | 60 (norm) | ✅ production-ready |
| N5 | GF20 | 20 | 7:12 | 289 (17²) | extract-only |
| N6 | GF24 | 24 | 9:14 | 1364 (L₁₅) | extract-only |
| N7 | GF32 | 32 | 12:19 | 0 (F₀) | spec-only |
| N8 | GF64 | 64 | 24:39 | 8388608 | no codegen |
| B1 | fp16 | 16 | 5:10 | — | ✅ IEEE baseline |
| B2 | bf16 | 16 | 8:7 | — | ✅ IEEE baseline |
| B3 | fp32 | 32 | 8:23 | — | ✅ IEEE baseline |
| B4 | fp8 | 8 | 4:3 | — | NVIDIA H100 only |
Models (6 participants)
| Lane# | Model | Arch | Hidden bands | Notes |
|---|
| M1 | MLP-baseline | feedforward | {384, 512, 768, 828, 1024, 1280} | reference |
| M2 | JEPA-T | Joint-Embedding Predictive | {512, 768, 1024} | 2-4 ATTN_LAYERS |
| M3 | NCA | Neural Cellular Automata | {384, 512} | 0 ATTN (cellular) |
| M4 | TF-1L | 1-layer Transformer | {512, 768, 1024} | 1 ATTN |
| M5 | TF-2L | 2-layer Transformer | {828, 1024} | 2 ATTN |
| M6 | TF-3L+ | 3+ layer | {1024, 1280} | 3-4 ATTN |
Competition Grid
12 numbers × 6 models = 72 cells. With 5 sanctioned seeds = 360 lane-runs.
Substitution Strategy
When full grid is not available, use tier-based substitution:
| Tier | Numbers | Method | Timeline |
|---|
| A (real) | GF16, fp16, bf16, fp32 | Native trainer support | Now |
| B (emulated) | GF8, GF12, GF20, GF24 | Quantize-on-the-fly (fp32 + weight quant) | T+3h |
| C (placeholder) | GF4, GF32, GF64 | Record "not_yet_implemented" | Post-Gate2 |
Tier B emulation: Use existing gf16.rs with parametric (exp_bits, mant_bits, phi_bias) to synthesize quantization without full backend. ~2-3h work.
Core Workflows
Check Competition Status
To check current competition status:
- Read latest experience file:
.trinity/experience/{date}.md
- Check fleet status (workers alive, heartbeat times)
- Check queue status (done, failed, pending, running)
- Check gate-2 eligible count
- Identify top performers by final_bpb
Generate Competition Report
To generate live standings:
- Query database for completed experiments with metrics
- Group by format (number) and model
- Calculate average BPB per cell
- Identify champions per format
- Identify champions per model
- Check gate-2 eligibility (final_bpb < 1.85, step >= 4000)
Report format:
🏁 IGLA RACE — Live Standings @ T+Nh
NUMBERS LEADERBOARD MODELS LEADERBOARD
1. GF16 bpb=1.42 (h828) 1. TF-2L bpb=1.38 (GF16)
...
CHAMPIONS BY CELL:
M1 M2 M3 M4 M5 M6
GF16 1.71 1.45 1.89 1.52 1.38 1.41
...
Update Worker Strategy
To update worker strategy:
- Identify phase (Phase 1: 4-number race, Phase 2: 8-number race, Phase 3: full grid)
- Determine active formats for current phase
- Determine active models for current phase
- Calculate lane-run count: formats × models × seeds × steps
- Update experiment queue with new configurations
Deploy Workers
To deploy IGLA workers:
- Verify Railway tokens (acc0-acc5) in
.env
- Verify Neon database connectivity
- Check image pin (current: trixie-slim@sha)
- Deploy worker image across target accounts
- Enqueue configurations via queue loader
- Monitor initial heartbeat responses
Monitoring Indicators
Critical WAKE_TRIGGERS
| Trigger | Condition | Severity | Action |
|---|
| W-1 | fleet_alive < 4/6 | CRITICAL | Log, investigate |
| W-2 | zero bpb_samples (15m) | WARNING | Check TRAINER_KIND |
| W-6 | Φ-1 collapse (identical bpb) | CRITICAL | Stop, investigate |
Gate-2 Requirements
- Target: final_bpb < 1.85
- Minimum steps: >= 4000
- Minimum samples: >= 5 distinct steps
- Eligible cells count toward Gate-2 progression
Additional Resources
Reference Files
For detailed competition rules and patterns, consult:
references/competition-rules.md - Gate definitions, validation criteriareferences/worker-config.md - Worker deployment patternsreferences/quantization.md - GF format emulation details
Database Queries
Useful queries for monitoring are in references/queries.sql:
- Fleet status aggregation
- Gate-2 candidate extraction
- Top performers by format/model
