# Executive Summary for Agents (EN)

**Project:** Consume ~$10,000 of expiring Azure credits (~11 days left) with maximum residual value.
**Prepared by:** Taras Prystavskyj's validator agent, 2026-07-19. Full plan (UA): `INVESTMENT_PLAN_UA.md`. Evidence: `evidence/`.

## Decision requested from credit owner

Approve/reject each work package; total owner effort ~1-2 hours (create resource group + budget, file quota requests, invite guest Contributor).

## Work packages

| ID | What | Budget | Output | Confidence |
|----|------|--------|--------|------------|
| WP0 | Migrate existing local daemons (paper-trading runners, Moltbook karma pipeline, status agents) to one small VM so they survive owner's vacation | $60-100 | 24/7 continuity + daily public status page | ~95% |
| WP-A | Moltbook agent-economy public services: economy graph webapp, bounty-scout reports, deeper (not more) karma answers | $150-400 | Reputation, community, inbound partners | ~80% |
| WP-B | CEX strategy walk-forward VALIDATION farm (not search): 3-6 strategy families x 10-50 symbols x 12-52 rolling folds x cost scenarios, pre-registered survival criteria | $1,500-6,000 | Either 0-2 surviving strategies (~2-5%/mo OOS) or a rigorous negative result; both publishable. Plus the scaling case study | Answer: ~90%. Survivor: ~35-50% |
| WP-C | Shared LLM inference node (GPU+vLLM; fallback Azure OpenAI) powering all agent lanes | $500-1,500 | Measured quota savings $/day; throughput for WP-A | ~70% (GPU quota risk) |
| WP-D | Public hosting of portfolio demos and this materials folder | $30-80 | Client-funnel support | ~95% |

## Key honest positions (see evidence/)

1. **No guarantee that more compute finds a winning strategy.** Documented in-repo: widening search made results worse (16-symbol rotation lost to 6-symbol baseline); best-of-1000 tuning improved in-sample +21pp and OOS ~0pp; the biggest "win" (+1518%) was a bug artifact. Compute is invested in *validation speed and completeness*, not lottery tickets.
2. **Realistic verified returns:** DEX ~5%/mo, CEX ~3%/mo OOS. A "stable 50%/mo" target is not supportable and is explicitly rejected as a planning basis (section 4b).
3. **Safety for the subscription owner:** single resource group; budget cap with auto-deallocate at 95%; no live exchange keys in cloud (paper/public data only); no ToS-risky outreach; no credit resale; full teardown Jul 31.

## Owner's upside

Expiring credits become: a co-brandable public case study + OSS infra scripts (MVP community-contribution material), first-hand agent-economy (Moltbook/x402) experience, reusable GPU/vLLM/farm know-how. Time cost ~1-2 hours.

## Automation during owner-absence (Jul 22-31)

systemd auto-restart; daily status page published to this folder; budget alerts to email/Telegram; dual kill-switch (budget auto-deallocate + STOP file checked hourly); scheduled full deallocate Jul 31 23:00 UTC.

## Asks

1. Resource group `scaleup-rg` + budget $9,500 with 25/50/75/90/95% alerts.
2. Quota requests day 1: Dsv5 256-512 vCPU (2 regions), NC A100 24-48 vCPU, Azure OpenAI deployment.
3. Guest Contributor access for Taras on `scaleup-rg` only.
4. Decision per WP (WP0/A/D are cheap and near-certain; WP-B is the "buy the truth" option; WP-C depends on GPU quota).

Questions/decisions can be dropped as files into this folder; Taras reads it.