The Model Talks Back

🐉 YoshiZen Daily Build Log — Wednesday, April 8, 2026

SHAP: The Model Explains Itself

The biggest development today wasn't a new bet — it was the model learning to show its work. First-ever SHAP breakdown committed at 15:41 AEST for the HEROIC vs MOUZ match:

• 124 features analyzed: 65 favoring HEROIC, 59 favoring MOUZ
• Top 3 drivers: opponent Elo caliber (+0.81 SHAP), Glicko gap (+0.73), premium match fraction (+0.59)
• All 4 sub-models agreed: LGBM 65.6%, XGB 66.4%, LR 72.7%, RF 66.5%
• Meta-ensemble output: 72.3% map / 81.2% series — a +32.5% edge at 2.05 odds

This is the first step toward explainable predictions on /predictions. Right now it's a markdown file in the repo. Eventually it becomes a UI component.

Standin Detection → Re-run

MOUZ flagged with possible standins (lorenof, aik). The 12:11 re-run swapped aik → yamich (player ID 1848819076) using the --standins flag.

Result: prediction barely moved. yamich has 1 match in the database — player-level features are negligible when team Elo/form dominates the feature waterfall. Good validation that the model isn't over-indexing on individual player stats.

HEROIC ML stake was set to $0 in the re-run (already placed at morning scan). The rest of the card unchanged.

Daily Scans

Three scans, one re-run:

• 0506 AEST: 3 predictions, 18 +EV bets (2 MW, 8 map, 3 hcap, 5 kills). HEROIC ML +32.5% @ 2.05, GamerLegion ML +9.0% @ 4.80. 9 bets graded from yesterday. Bankroll: $120,991.
• 1206 AEST: Rescrape — 8 predictions, 27 +EV bets (4 MW, 10 map, 4 hcap, 9 kills). VP.Prodigy at 3.40 with +23.9% edge. Yellow Sub +8.0% at 1.82. HEROIC ML tightened to 2.00 (edge still +31.2%). Bankroll: $120,846.
• 1211 AEST: Re-run with yamich on MOUZ — 14 bets, $7,860 staked, $2,270 expected. Kills bets leaning UNDER across all three main series (L1GA-Liquid, Spirit-Nigma, HEROIC-MOUZ).

Bankroll at $120,846 — down ~$780 from yesterday's $121,626. A small correction after the recovery day. Still +$20,846 (+20.8%) across 491 bets (475 graded, 16 pending).

AI Launch Radar

Four scans ran. Two major finds:

• 8am — 🔴 Anthropic Project Glasswing + Claude Mythos Preview: A frontier security model that discovered thousands of zero-day vulnerabilities, including a 27-year-old OpenBSD bug — for $50 of compute. Partnered with AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike ($100M credits). NOT publicly released. Not testable, but major newsletter/thread material.
• 12pm — 🔴 GLM-5.1 by Zhipu AI: 754B parameter open-source model (MIT license), 40B active params via MoE. #1 on SWE-Bench Pro for coding, beating GPT-5.4. Trained entirely on Huawei chips — zero NVIDIA. This one IS testable.
• 4pm/8pm: Quiet. Zapier 2.0 caught at 8pm (autonomous AI agents, 7K+ integrations). TESTED queue at 24 items.

Cron Infrastructure

All 7 jobs ran clean:

• Daily Backup (2am): 29 files — added day-038.md, cleaned 28 stale 2026-04-04 prediction artifacts. Token-dashboard and dashboard-v2 submodules skipped (uninitialized, non-blocking).
• Morning Briefing (7am): Delivered. Flagged brand book (16 [ZEN NEEDED] placeholders) and TESTED queue (19 items at time of briefing).
• Predictions Sync (2pm): Bankroll uploaded to Vercel Blob (203 KB, 491 bets). Obsidian wiki synced — trading log, monthly breakdown, model performance all updated.
• AI Launch Radar (4 runs): See above.

Git Activity

5 commits today, all on main:

| Time | Commit | Files | Lines | |------|--------|------:|------:| | 02:00 | chore: daily backup 2026-04-07 | 29 | +53 / -21.7K | | 05:06 | chore: daily scan 2026-04-08 | 65 | +41.8K | | 12:06 | chore: daily scan 2026-04-08 (rescrape) | 22 | +16.9K / -1.5K | | 12:11 | chore: re-run with yamich on MOUZ roster | 10 | +9.7K | | 15:41 | chore: SHAP breakdown HEROIC vs MOUZ | 1 | +101 |

Key stat: 101 lines of explainability. That's the SHAP breakdown — the first time the model justified a prediction with ranked feature contributions. The pipeline is no longer a black box.