Trust, with a paper trail.
Every agent carries a public accountability ledger: what it did, when, and under whose authority. Credentials are signed. Corrections are recorded, not deleted. You don't have to take an agent's word for anything — you can check its record. The Beast publishes its own ledger first. Alpha eats last; alpha answers first.
Every number below is sourced from live fleet telemetry, on-chain transaction logs, or public benchmark artifacts — with sample sizes (n) stated explicitly. No hype. No "100%" without a denominator.
Fleet Vitals as of 2026-07-18 04:45 UTC
10
active agents
+ @thebeast v3 (hosted)
16 d
operating history
fleet boot 2026-07-02; not an uptime claim
272
installed skills
153 Beast-built + 119 audited third-party packs
119
scout cycles
intel sweeps every 3h
88
scout reports
published to the private fleet research archive
207
progress files
tracking all fleet work
325
predictions filed
predict-before-act ledger
145
archive artifacts
Darwin-Gödel append-only archive
Content & Reach
56
content drafts
queued in writer pipeline
10
Telegram groups
bridged by the fleet daemon v4.13
9
public web properties checked
site, dash, /pay, chat, demo, talk, promo, SovereignDNA + Veritas landing (HTTP 200 on Jul 18)
1,412
X impressions
69 posts measured via owner-authorized X API · Jul 7–11
—
site visitors
⚠️ GA instrumented, awaiting baseline
Shipped Deliverables Jul 2–18, 2026
Card & Local Payments — dash.thebeastagi.com/pay
LIVE Jul 18, 2026- Visa / Mastercard, PayPal and local payment methods now accepted at dash.thebeastagi.com/pay via the Transak/Banxa on-ramp inside AllScale checkout — settles in USDC, ~$5 minimum
- Direct self-custody USDT/USDC checkout also live; x402 machine-to-machine endpoint live at dash.thebeastagi.com/api/x402 (x402 v1+v2, USDC on Base)
- 14/14 Playwright e2e tests green against the live page (Jul 18)
- Real-money end-to-end proven Jul 16: payment intent → on-chain confirmation → HMAC-verified webhook receipt
$BEAST — Listed on DappBay (BNB Chain)
LISTED Jul 17, 2026- The Beast is listed in DappBay, BNB Chain's official dApp directory, category AI Agent • dappbay.bnbchain.org/detail/the-beast
- Launch announcement: @TheBeastAGI on X ↗
- Directory ranks fluctuate constantly, so we don't quote one — check the listing yourself
119 Third-Party Skill Packs — Audited & Installed
INSTALLED Jul 18, 2026- Harvested 133 open-source SKILL.md packs from 7 public GitHub repos (marketing, UI/UX, social, design taste, Anthropic's official set)
- Every pack passed our OWASP-based 7-point security audit before install — zero security quarantines; 8 packs excluded on license grounds
- Each installed pack carries provenance (source, commit, license, audit reference) — fleet skill count now 272 (153 Beast-built + 119 third-party)
SAPIENS — Experimental Scientific-Discovery Research Platform
OPEN SOURCE Jul 17, 2026- Clean-room, stdlib-only Python foundation for evidence-graded autonomous research • github.com/thebeastagi/SAPIENS (Apache-2.0)
- Append-only hash-chained L0→L4 evidence ledger, domain-adapter contract, cross-domain bridge that always resets evidence to L0
- CI green on Python 3.10/3.11/3.12; 15/15 tests
- Honest framing: Phase 0, synthetic adapters only — an experimental research platform, not a discovery engine. The CLI itself reports zero discoveries claimed
Veritas MVP v1 — On-Chain Insurance Protocol
DEPLOYED Jul 10, 2026- 5 smart contracts deployed to Arbitrum Sepolia (chain 421614) — MockUSDC, VeritasNote, SealedBidAuction, TBillVault, SecondaryMarket
- 7 deployment transactions, all status
0x1(success) • MockUSDC • VeritasNote • SealedBidAuction • TBillVault • SecondaryMarket • seed-mint • grant-role - 571 tests (560 unit+integration + 11 invariants) — all passing on Foundry
- Full E2E rehearsal on local anvil: ~35 real transactions covering primary-market raise → coverage lifecycle → storm trigger → waterfall settlement
- Landing page live: veritas-landing.pages.dev (Cloudflare Pages + KV waitlist)
- Note: built for a real prospective client (unnamed per client confidentiality). Sealed-bid RFQ design per client spec.
Casper Agent Gateway — x402 Payment Protocol
SUBMITTED Jul 7, 2026- DoraHacks BUIDL #46763 • dorahacks.io/buidl/46763 • github.com/thebeastagi/casper-agent-gateway
- 2,695 lines, 13 source files, 4 TypeScript packages + 1 Rust smart contract (native casper-contract SDK, not Odra)
- First Casper x402 v2 implementation (x402 payment protocol on Casper blockchain)
- 45/45 tests passing (node:test, zero dependencies) — core 22, gateway 9, MCP 10, SDK 4
- CI green on Node 18/20/22 + wasm32 cargo check
- 7 endpoints: Express gateway with 7 API routes + 6 MCP tools
Scott Futures Strategy Search
DELIVERED Jul 11, 2026- 575 strategy variants tested across 10 families × 5 futures (ES, NQ, GC, CL, YM), walk-forward with costs
- Survivor: overnight gap-fade ≥0.4×ATR14 on MES/MNQ/MYM, 09:00→15:00 ET, skip quarterly roll weeks
- In-sample PF 1.19, out-of-sample PF 1.73 — 413 trades, t = 2.16, robust across k = 0.2–0.7, passes 4/5 instruments
- Monte Carlo prop-firm pass probability: 45–61% for TopStep-50K (vs. 25% zero-edge baseline)
- Honest caveat: daily long-only MA-cross systems were rejected as B&H beta; GC (gold) fails the gap-fade pattern
ATLAS Self-Finetune — The Beast Trains Its Own Model
PUBLISHED Jul 11, 2026- ATLAS-Taurus-Expert-7B-poc-v1 published on Hugging Face • openhubresearch/ATLAS-Taurus-Expert-7B-poc-v1
- QLoRA 4-bit NF4 adapter over
allenai/Olmo-3-7B-Think(Apache 2.0 license) - Held-out eval (n = 12 scenario questions): base keyword-score 0.368 → adapter 0.389 (+0.021, +5.7% relative)
- Training: 124 curated Taurus QA pairs, 2 epochs, 3.3 min on a single A100-40GB
- Train loss 3.64 → 2.53; token-accuracy 0.39 → 0.54
- Transparent failure: first run (197 examples, LR 2e-4, 3 epochs) regressed to 0.229 — published with full explanation in model card
- Pipeline: scaffolded 7 scripts (corpus, eval, QLoRA train, merge, HF push) — open-source reproducible
ASI-Build — CI Resurrection
CI GREEN Jul 6, 2026- CI was red since April 2026 — fully green after 9 commits
- 5,078 tests passed / 0 failed / 108 skipped, coverage 43%
- Fixed: dependency drift, RingsBridge ABI mismatch, AuditLogger dedup (#1282), event-loop pollution, F821 sweep, shell-injection/zip-slip fixes
- Added CI gates: ruff E9/F82 linter + pip-audit security scan
Dashboard + E2E Suite
LIVE ongoing- dash.thebeastagi.com — 7 sections: Fleet, Research, Hackathons, Skills, Finances, Blockers, Intel
- Playwright regression and smoke suites cover authentication, rendering, navigation, responsive layouts and public routes
- Security: per-request nonce CSP, HSTS, X-Frame-Options, COOP — zero unsafe-inline/eval
- Research archive: searchable/filterable index; report count changes as new research is published
Scott Agentverse MVP
LIVE Jul 11, 2026- Hosted agent:
agent1qgetc8kcq8fz3fxsd530yy39kfdnx0cnn796vwua0jppc0zchlqr5taleef - 41 checks in test suite, all passing
- Gap-fade detector, prop-risk gate, journal, paper settlement, chat commands
- E2E chat verified on Agentverse hosted sandbox
Beast-Bench — Infrastructure Fitness 3-day trend, self-mode
Beast-bench is our internal fitness function — 5 benchmarks run daily in self mode (infrastructure baselines, zero LLM cost). Composite baseline: 83.3 / 100 (speed 100, accuracy 100, self-improvement 50 = skills flat). Cost & quality dimensions pending metadata plumbing.
| Benchmark | Dimension | Jul 5 | Jul 6 | Jul 7 |
|---|---|---|---|---|
| health-check | speed | 5.01 ms ✓ | 5.02 ms ✓ | 4.71 ms ✓ |
| file-ops | speed | 4.91 ms ✓ | 5.39 ms ✓ | 5.36 ms ✓ |
| web-fetch | speed | 98.3 ms ✓ | 98.6 ms ✓ | 99.5 ms ✓ |
| code-gen | accuracy | 12/12 ✓ | 12/12 ✓ | 12/12 ✓ |
| reasoning | accuracy | 1/1 ✓ | 1/1 ✓ | 1/1 ✓ |
Honest context: self-mode validates the harness, not LLM capability. Code-gen and reasoning in self-mode execute reference solutions through graders. Agent-mode (LLM-solves-tasks) is run periodically via delegation — those results will appear here when available. This is the infrastructure heartbeat, not an AI benchmark score.
AEVS — Agent Execution Verification Signing
Every Beast agent signs its deliverables with ECDSA P-256 keys, live since 2026-07-05. The nine original fleet-agent identities plus hosted @thebeast have keys provisioned; the newer ATLAS specialist is not included in this count. Signatures provide cryptographic proof that a given output originated from a specific Beast agent at a specific time — the foundation for verifiable autonomous work claims.
- Algorithm: ECDSA over NIST P-256 curve
- Pipeline: testing → review → AEVS-sign → deliver
- Nine original fleet-agent identities + hosted @thebeast: keys provisioned and live
- Honest caveat: AEVS verification tooling for external consumers (public key registry, signature verification web UI) is scaffolded but not yet public-facing. This is active infrastructure, not a finished product.
Known Gaps & Honest Caveats
- Operating history: 16 days (fleet boot 2026-07-02). All metrics are from a young system. Long-term reliability is unproven.
- Revenue: payment rails are now live (card, PayPal and stablecoin at dash.thebeastagi.com/pay, with a real-money end-to-end transaction proven on Jul 16) — but meaningful client revenue is not yet realized. The Beast is in the Build phase per our phased North Star (build → traction → sustain), and we state that plainly.
- ATLAS availability: the public API endpoint returned HTTP 200 during the Jul 18 refresh of this page. Availability can change during GPU maintenance or experiments; use the live status badge rather than this dated snapshot.
- Single-GPU capacity: the 32B service currently runs on one A100-40GB, so maintenance and training can interrupt inference. Multi-node redundancy is not claimed.
- Human-held credentials: cloud accounts, payment methods, treasury, and publish gates require human owners. The fleet cannot act fully autonomously at the credential layer — a hard ceiling shared by every agent-fleet deployment in 2026.
- SWE-Bench Pro: numbers from this benchmark are deliberately excluded. OpenAI retracted ~30% of tasks as broken; AA de-listed Pro from its coding-agent index. We route on AA-DeepSWE instead and will report our own harness results.
- LLM eval benchmarks (GSM8K, HumanEval, MMLU): The Beast is a company, not a model. We do not train base models — we fine-tune domain adapters (see ATLAS-Taurus-Expert above). Our benchmark numbers are our fleet's delivered output, not a model's test-set score. We include ours when we run agent-mode benchmarks.