EXECUTION PLAN: fast frequent experiments — GATE-0 smoke → seed-hunt → Gate-2 BPB<1.85

[gHashTag]

Goal: WIN via speed, not luck

Current champion: IGLA-TRAIN_V2-FP32-CHAMP-seed42 → BPB=1.8921 🥇
Gate-2 target: BPB<1.85 quorum 3/3 on seeds {43,44,45}
FINAL target: BPB<1.50 quorum 3/3 (GATE-FINAL)

Win strategy: short → predict → kill → repeat. Never burn 12h compute on a bad seed. Use power-law extrapolation from step 100–500 to predict final BPB with 95% CI, kill losers early, mirror champions.

---

Phase 0 — SMOKE RACE (45 min, 500 steps × 21 services)

Purpose: verify the whole pipeline works before burning real compute.

tri smoke-race \
  --config configs/gate0-smoke.toml \
  --tags WSD,BS8,GRADFIX-JEPAT,GRADFIX-NCA,EMA10,h512,h768 \
  --seeds-per-tag 3 \
  --max-steps 500 \
  --wallclock 30m \
  --account acc1 \
  --auto-teardown \
  --report docs/SMOKE_GATE_0_REPORT.md

14 acceptance criteria (all must be GREEN before proceeding):

• 21/21 services deploy with canonical names (INV-12)
• All containers start, step=10, BPB=X.XXXX in logs within 2 min
• stdout matches ^BPB=\d\.\d{4}$ (L-R8)
• bpb_samples receives >40 rows/seed in 30 min
• igla_race_trials has 21 rows, one per canon name
• JEPA-T/NCA report BPB not MSE (L-METRIC)
• IGLA-TRAIN_V2-GF16-* only valid when h≥256 (L-R9)
• tri leaderboard rank --gate gate-0 returns 21 rows with distinct BPB
• Re-deploy of CHAMP seeds → ReusedChampionSlot error (New Issue #98)
• Re-deploy on occupied slot without --force-replace → SlotStillOccupied (🪲 Scarabaeus Engine — Autonomous Worker Loop (Khepri rolls the sun) #101)
• --account acc1 --kill <acc2-name> → CrossAccountKillForbidden (docs(scarab): dedupe brake — scarab IS seed_agent, no second crate #108)
• Two parallel tri exp claim → no duplicate EXP_ID (New Issue #99)
• Idempotent re-deploy same canon+EXP_ID → no-op (feat(scarab): ledger-daemon PR-1 scaffold (Khepri-3 watchdog) #105)
• Watchdog kills after 30 min, container writes SMOKE_DONE

Smoke canon names (TAG=SMOKE, seeds 500..599):

IGLA-HYBRID-FP32-WSD-SMOKE-seed500..502
IGLA-HYBRID-FP32-BS8-SMOKE-seed510..512
IGLA-JEPA-T-FP32-GRADFIX-SMOKE-seed520..522
IGLA-NCA-FP32-GRADFIX-SMOKE-seed530..532
IGLA-HYBRID-FP32-EMA10-SMOKE-seed540..542
IGLA-HYBRID-FP32-h512-SMOKE-seed550..552
IGLA-HYBRID-FP32-h768-SMOKE-seed560..562

---

Phase 1 — SEED HUNT (2000 steps, ~30 min per rung)

Goal: use power-law fit to predict bpb_∞ for each seed and kill losers BEFORE wasting 12h.

Algorithm per rung

BPB(t) = bpb_∞ + a · t^(−p)     ← fit via NLS Levenberg-Marquardt, 95% CI

Rung	Steps	Action
rung-0	100	collect BPB curves, no kills
rung-1	500	fit power-law → drop bottom 50% (DIVERGING state)
rung-2	2000	re-fit → identify top-3 candidate seeds
rung-3	8000	promote LEADING seed → mirror 3 siblings
rung-4	32000	keep top-1 champion challenger
rung-5	81000	GATE-2 final

Seed state machine

State	Condition	Action
LEADING	Δ ≤ 0 (below leader BPB)	mirror-spawn 3 siblings
CATCHING_UP	Δ > 0, slope < 0	keep — candidate converging
PARALLEL_LOSING	Δ > 0.05, slope ≈ 0	keep until next rung
DIVERGING	Δ > 0, slope > +0.001/step	kill immediately → replace with phi-anchored seed

# After rung-1 (step 500):
tri hunt status --rung 1 --account acc1
tri hunt prune  --rung 1 --apply --account acc1  # kill DIVERGING
tri hunt mirror --leader <best-seed> --count 3 --strategy phi-anchored --account acc1

Seed replacement strategy (when DIVERGING is killed)

1. φ-anchored: seed_k = floor(φ^k · 1e6) mod 2^32 — hypothesis: Lucas-closure gives better gradient conditioning
2. Sibling of current leader: leader_seed ± 1
3. Random uniform(2^32) as fallback

---

Phase 2 — CHAMPION EXPERIMENTS (21 services × 3 seeds, 12h each)

Run only after smoke GREEN. All names use IGLA-<TYPE>-<NUM>-<TAG>-E<ID:04>-seed<N> canon format.

#	Canon base	Seeds	Expected ΔBPB	Time	EXP_IDs
1	IGLA-HYBRID-FP32-WSD	200/201/202	−0.10…−0.25	12h	E0005..E0007
2	IGLA-HYBRID-FP32-BS8	210/211/212	−0.05…−0.10	12h	E0008..E0010
3	IGLA-JEPA-T-FP32-GRADFIX	220/221/222	−0.10…−0.30	12h	E0011..E0013
4	IGLA-NCA-FP32-GRADFIX	230/231/232	−0.05…−0.20	12h	E0014..E0016
5	IGLA-HYBRID-FP32-EMA10	240/241/242	−0.03…−0.10	12h	E0017..E0019
6	IGLA-HYBRID-FP32-h512	250/251/252	−0.05…−0.20	12h	E0020..E0022
7	IGLA-HYBRID-FP32-h768	260/261/262	−0.10…−0.40	18h	E0023..E0025

MINIMUM set if compute is tight (9 seeds): WSD (3) + JEPA-T GRADFIX (3) + h768 (3)

Spin script:

tri railway batch-deploy --account acc1 --validate-canon --config configs/phase2-runs.toml <<EOF
IGLA-HYBRID-FP32-WSD-E0005-seed200
IGLA-HYBRID-FP32-WSD-E0006-seed201
...(all 21)
EOF

---

Phase 3 — GF16 HEDGE (9 services, T+18h)

Run champion models with GF16 precision — proven numerically stable via INV-3,5 (Lucas-closure). Only valid for h≥256 (L-R9).

Canon	h	Status
IGLA-TRAIN_V2-GF16-CHAMP-seed{300,301,302}	1024 ≥ 256 ✅	strongest candidate
IGLA-HYBRID-GF16-CHAMP-seed{310,311,312}	828 ≥ 256 ✅	second candidate
IGLA-PHI-GF16-CHAMP-seed{320,321,322}	384 ≥ 256 ✅	phi-anchor test

---

Phase 4 — PRECISION SWEEP (5 services, T+30h)

Canon	Format	Purpose
IGLA-TRAIN_V2-DLFLOAT-CHAMP-seed330	DLFloat 1:6:9	IBM baseline precision
IGLA-TRAIN_V2-BF16-CHAMP-seed331	BF16	wide-support precision
IGLA-HYBRID-FP8E4M3-CHAMP-seed340	FP8 OCP	ultra-low precision test
IGLA-HYBRID-FP8E5M2-CHAMP-seed341	FP8 OCP	ultra-low precision test

---

CANON RULES (mandatory for all agents)

IGLA-<MODEL>-<NUMBER>-<TAG>-E<EXP_ID:04>-seed<SEED>

• Never reuse a service name across experiments
• Never deploy without EXP_ID from tri exp claim (tripwire feat(scarab): ledger-daemon PR-1 scaffold (Khepri-3 watchdog) #105)
• Never kill champion locks E0001..E0004 (tripwire New Issue #98)
• Never deploy GF16 with h<256 (tripwire R5 ANOMALY: bpb_samples step=1000 rows ≪ architectural floor (0.0003–0.0146 vs 2.19) — Gate-2 promotion BLOCKED #97, L-R9)
• Kill old → spin new. Slot must be empty before deploy (tripwire 🪲 Scarabaeus Engine — Autonomous Worker Loop (Khepri rolls the sun) #101)
• All control via MCP only. No direct railway, gh, psql calls (RULE-8, tripwire [P0 CRITICAL] Entire fleet trained on tiny_shakespeare — all FineWeb claims invalid #109)

---

Timeline

Window	Action	Cumulative
T+0…T+0.5h	Phase 0: SMOKE RACE (45 min)	T+0.5h
T+0.5…T+2h	Fix any smoke failures	T+2h
T+2…T+3h	Cull 12 old dummy/broken services	T+3h
T+3…T+5h	Phase 1 rung-0/1: deploy 21 services, first seed hunt at step 500	T+5h
T+5…T+6h	Hunt prune: kill DIVERGING, mirror LEADING	T+6h
T+6…T+18h	Phase 2: 12h runs → rung-2,3 decisions at steps 2K/8K	T+18h
T+18…T+30h	Phase 3: GF16 hedge spin	T+30h
T+30…T+42h	Phase 4: precision sweep	T+42h
T+42h	Gate-2 evaluation: BPB<1.85 quorum 3/3?	—

---

Blockers (resolve in order before Phase 2)

• PR-1: extract stable public core crates from trios-railway (tri-core, tri-hunt, tri-exp, tri-canon, tri-ledger) #69 Extract public crates from trios-railway
• P0: Port train_v2 champion to Railway (3-seed quorum for Gate-2 OFFICIAL) #65 feat/exp-id-allocator: tri exp next/claim + Neon sequence
• feat(gardener): GATE-0 SMOKE RACE — 21-service end-to-end pipe-cleaner #66 feat/account-scoped-tokens: --account acc1/acc2/acc3 on all CLI commands
• feat(gardener): Seed Hunter — power-law fit + leader-relative classifier #67 feat/bpb-samples-unblock: telemetry ingest pipeline live
• PR smoke-race: bin/tri-gardener/src/smoke.rs + configs/gate0-smoke.toml
• PR seed-hunter: crates/igla-fit/src/power_law.rs + crates/igla-hunter/src/decision.rs

---

Key invariants

• ARCHITECTURAL_FLOOR_BPB = 1.89 — any reported BPB < 1.89 without human-review → tripwire suspected ledger corruption
• rungs = [100, 500, 2000, 8000, 32000, 81000] — rung_factor=4, threshold=φ²+φ⁻²+0.5=3.5
• Smoke race = first run before any production Phase-2 deploy. No exceptions.

φ² + φ⁻² = 3 · TRINITY · SMOKE BEFORE FIRE · NEVER REUSE · FAST EXPERIMENTS WIN