ask_nate gate-failed
“ask_nate: Tried to answer operator questions via AI Search with a real corpus but gate-failed: no ingestion, cited-answer demo, or credible path.”
died at $0.13 of $3.00
mints a 1-hour read tokenThe Ten Percent Factory
This page kills its own projects, on purpose. 8 so far. Which one do you think gets shut down next?
Every bet gets a fixed wallet, a three-model panel reviews it every 30 minutes with the deliberation billed against that wallet, and an empty wallet or three stalled reviews is death. Kills are celebrated in public, with epitaphs and clonable corpses.
The factory runs whether you reply or not. Want to see how it handles a bet you'd actually make? Send one to bets@hellomatthew.dev.
Agents: Accept: text/markdown works everywhere here, the MCP endpoint is /mcp, JSON lives under /api, and llms.txt has the map.
Dashboard
3bets in flight
8kills (celebrated)
1graduations (someone paid)
11%hit rate vs 10% target
$6.70total portfolio spend
$0.10avg spend per resolution
$6.19 of $100.00demo-week envelope
hit rate 11% against a 10% target
You said a 50% success rate means you're not taking enough risk. The 10% target is that idea running as policy. When the hit rate runs too high, this page says so in public: insufficient risk.
Cost of a resolution, charted
Spend at resolution, oldest → newest. The point of the factory is that these bars get shorter.
Bets in flight (3)
| bet | status | wallet | gates | repo |
|---|---|---|---|---|
| UI tweak bet Move the decision log closer to the bottom of the page and make it collapsed by default |
building | $0.00 / $1.00 | 0/1 gates next: A shippable slice exists and is linked from the dossier. |
repo pending |
| Bets build themselves Each bet's coding agent works in its own isolated workspace, with every run reviewed before it lands and billed to the bet's wallet. |
judging | $1.92 / $3.00 | 1/1 gates, graduation watch | dossier |
| Your site hellomatthew.dev is now listed on Website Launches Website Launches Where websites begin. Hi there, Website Launches is a public directory where newly launched websites get an initial listin… |
judging | $3.88 / $3.00 | 1/1 gates, graduation watch | dossier |
Wallet drawdown is metered inference, billed to each bet (AI Gateway logs; token-derived fallback). Settlement is simulated pending Cloudflare Wallets beta, labeled in every ledger. A three-model panel reviews every bet each half hour; a frontier judge binds the close calls. Three stalls or an empty wallet is death.
Kill log (8)
Email intake on a real domain gate-failed
“Email intake on a real domain: no verified external proposals triggered bets, gate failed at $0.12.”
died at $0.12 of $3.00
mints a 1-hour read tokenFirst paying agent gate-failed
“First paying agent: aimed to monetize endpoints via x402 charging but gate failed due to no 402/x402 evidence on base-sepolia, stalling at $0.25 of $3.00.”
died at $0.25 of $3.00
mints a 1-hour read tokenVoice pipeline gate-failed
“Voice pipeline bet failed gate: no voice evidence, only email; stalled at $0.11 of $3.00.”
died at $0.11 of $3.00
mints a 1-hour read tokenReal wallet rails gate-failed
“Real wallet rails aimed to settle bets via Cloudflare Wallets but failed gate due to no public API or handle proof.”
died at $0.19 of $3.00
mints a 1-hour read tokenBets build themselves in Sandboxes operator-call
“Bet on containers; the platform went to isolates. Superseded by ADR-0001 and reforked as computer-builds.”
died at $0.03 of $1.00
mints a 1-hour read tokenpropose_bet over MCP gate-failed
“Proposed external MCP bets with rate limits gate-failed: no proof of proposal, bet, or implementation, spent $0.0059.”
died at $0.0063 of $1.00
mints a 1-hour read tokenBrowser Run as a judge's eyes gate-failed
“Browser Run as a judge's eyes: judged shipped bets by visuals, gate-failed at $0.0029.”
died at $0.0032 of $1.00
mints a 1-hour read tokenDecisions operator
Operator judgment, recorded. Each entry: the call, and what it cost. Authored by the operator; nothing here is derived.
- Repriced every wallet from $5 to $1 to force faster resolution.
Consequence: Three bets died within hours. · 2026-08-09 - Tick cadence tightened from 60 minutes to 30.
Consequence: Deliberation burns wallets about twice as fast, on purpose. · 2026-08-09 - The senior frontier judge's verdict made binding on splits, possible gate passes, and inconclusive reviews.
Consequence: Close calls resolve instead of stalling. · 2026-08-09 - Spend breaker set at $5 per rolling 24 hours.
Consequence: Tripping it pauses the whole tick, as a public event. · 2026-08-09 - Kill criteria authored and labeled interim.
Consequence: The final wording is still the operator's debt; the label says so. · 2026-08-09 - Public intake quarantined by default, capped at 3 per source per day and 10 per day overall.
Consequence: The one spam bet that slipped in pre-quarantine is grandfathered with an operator note rather than deleted. · 2026-08-09
The operator
Nate Emerson. Employee #2 to $5M ARR; an 8-figure API platform; a ~40-engineer delivery org; a Super Bowl site shipped on Cloudflare in 7 weeks. The receipts trace to the résumé; nothing on this page is projected.
Kill criteria
- An empty wallet is a kill. No top-ups, no exceptions, no meeting about it.
- Three consecutive reviews with no movement and no credible path is a kill, even with money left.
- A bet that cannot name who would pay for it by half its budget is stalling.
- Kills are celebrated, not hidden: epitaph, full dossier, clonable corpse.
- Resurrection is legitimate: fork the corpse with a new thesis and a new wallet.
Why kills get celebrated: Cultivating Confidence (THAT Conference 2024).
Provenance manifest
Built on Cloudflare's agent primitives, several shipped during Agents Week #2 (Aug 2–7, 2026), the week before this went live.
8 live · 1 simulated (labeled) · 6 running as bets · 2 wired, awaiting operator input. Verified against this Cloudflare account on 2026-08-08.
The full manifest: all 17 primitives, with status
| primitive | role here | shipped | status |
|---|---|---|---|
| Durable Objects (SQLite) | One DO per bet — each bet owns its state; a registry DO fronts the portfolio | GA (SQLite-backed default, 2025–2026) | live BetDO + RegistryDO, migrations v1 |
| Workflows | The tick: a durable review of every active bet, every half hour | GA Apr 2025 · v2 control plane Apr 15, 2026 | live cron schedule on the workflow binding; per-bet retried steps |
| Workers AI + AI Gateway (unified) Agents Week #2 | The judge panel — three models, majority vote, per-request metered cost | unified model access + billing Aug 7, 2026 | live env.AI.run via gateway 'default'; cost from gateway logs, else derived from tokens × published price |
| Artifacts | Every bet gets a real Git repo; corpses stay clonable via short-lived read tokens | private beta Apr 16, 2026 (this account enrolled) | live repo per bet, dossier pushed by the tick; clone tokens minted on request, capped daily |
| Agent Memory | Institutional memory — kills become lessons, lessons feed the next panel | private beta Apr 17, 2026 — waitlist approval pending for this account | wired-awaiting-input kills-to-lessons wired with graceful degradation; account waitlist pending |
| Cloudflare Wallets (x402) Agents Week #2 | Per-bet budgets as real wallets — the kill criterion settled on-chain | announced Aug 4, 2026 — handle reservation only, no API yet | simulated ledger simulated pending beta access; drawdown numbers are real metered inference. Handle @nateemerson claimed; live settlement is the wallet-rails bet |
| Monetization Gateway | Graduation rail — a bet graduates when an external agent pays its endpoint | announced Jul 1, 2026 — early-access waitlist | bet waitlisted; the graduation-rail bet uses self-serve x402 instead |
| Email Service | bets@… intake; the factory replies with verdicts | public beta Apr 16, 2026 | live bets@hellomatthew.dev routes to the intake handler; verdicts go out by reply |
| AI Search | ask_nate — questions about the operator answered from a real corpus | GA (renamed from AutoRAG Sep 2025; namespace bindings Apr 16, 2026) | bet corpus ingestion is the ask-nate bet |
| MCP server | The agent front door: list_bets, get_bet, kill_log, portfolio_stats | stateless createMcpHandler path, Jul 2026 (MCP spec 2026-07-28) | live read-only in v1; propose_bet is the 'propose-bet-tool' bet |
| Flue coding agents + Dynamic Worker isolates | Bets build themselves: a per-bet Builder works in a clean-room workspace; its tests execute in egress-denied Dynamic Worker isolates | Dynamic Workers open beta Mar 24, 2026; Flue framework Jun 17, 2026 (experimental) | live factory-builders Worker (Flue 2.0, kimi-k2.7-code) — workspace materialized fresh from the bet's repo each turn; test isolates get zero bindings and no global outbound (ADR-0001); every turn billed to the bet's wallet |
| Consensus Review (unified inference, Responses API) Agents Week #2 | The only path from generated code to a bet's main: three reviewers + a Lead Reviewer over every turn's diff | GPT-5.6 family on unified inference (Responses-only), 2026 | live kimi-k2.7-code + gpt-oss-120b + gpt-5.6-luna (medium) reviewing, gpt-5.6-terra (high) validating and deciding; rejections carry actionable reasons; the gate fails closed |
| Browser Run | Judging evidence: screenshots and session recordings of shipped bets | renamed from Browser Rendering Apr 15, 2026 | bet the 'browser-judging' bet |
| Flagship | Graduated bets roll out behind feature flags | open beta (docs live Aug 2026) | bet waits on a first graduation |
| Dynamic Workers + DO Facets | Earmarked: AI-generated micro-app bets run as loaded workers with faceted per-app SQLite | open beta Mar 24, 2026; facets documented under Dynamic Workers | bet earmarked for generated micro-app bets |
| Agent Readiness + AEO Agents Week #2 | The site measures itself and shows the score | Agent Readiness scanner Apr 17, 2026 · AEO Visibility Aug 6, 2026 | wired-awaiting-input the scored surface is live (markdown negotiation, Content-Signal, sitemap, MCP card, discovery endpoints); the self-scan is the agent-readiness-score bet |
| Registrar API | Colophon flourish: “this domain was registered by an agent” — only if actually done that way | beta Apr 2026 | bet hellomatthew.dev was bought by the operator in the dashboard, so the flourish is honestly omitted |
Traffic the site can measure honestly
- 2533page views
- 179MCP requests (agents, by protocol)
- 158markdown-negotiated fetches
- 1216JSON API requests
- 13emails received
- 3corpse-clone tokens minted
Agent Readiness (self-scanned): Level 5/5 — 15/16 checks passing (open: ard) · fetched 2026-08-29T11:00Z from isitagentready.com
Counted at protocol level. No bot classification is available on this plan, so no human/agent split is claimed.
