In plain terms, for readers who don't follow crypto:
- The problem. Most focus apps rely on willpower at the moment it is lowest. Staking money on a goal has real evidence behind it, but an app that keeps your stake when you fail profits from your bad days.
- What I built. A phone app where you stake a small amount on a focus session and state what you will do. Keep to it and you are refunded; slip and the stake goes to charity, never to the app. Refunds and charity payments settle end to end on a test network; the AI coach that gives the verdict is written but not yet proven live.
- What it shows about how I work. Honest scope and an audit trail. Every feature claim is marked built, in progress or planned, what didn't ship has its own file, and AI-USAGE.md records which AI tool did what and what I did myself.
A tax on doomscrolling that funds something good. Stake on a focus session. State your intention. A private AI coach decides if you kept it. Keep it, get refunded. Slip, and an autonomous agent sends your forfeit to charity. Grounded in commitment-device research. Designed to be kind about the bad days.
Built solo at ETHGlobal Lisbon 2026, using AI coding tools (GitHub Copilot and Claude) throughout. AI-USAGE.md logs what each tool did and what I did myself.
Live: doomtax.vercel.app · Architecture · What didn't ship · AI usage · All docs · Repository map
Testnet only. No real funds, no real charity partnership (placeholder account). See
NOT-BUILT.mdfor honest scope.
Global mobile screen time averages roughly 4h37m a day. That number alone isn't the problem, a dashboard reporting it back at you isn't a solution. The problem is the gap between what you meant to do and what you did, and most tools that target it fail structurally: they depend on willpower at the exact moment willpower is lowest, so they get disabled, snoozed, or ignored.
Commitment devices are the part of this category with actual evidence behind them. StickK, founded by Yale economists Dean Karlan and Ian Ayres, reports that adding a referee roughly doubles goal attainment, and a financial stake raises it up to threefold.
So the mechanic works. What's unsolved is who gets paid when you fail.
| Money flows | On failure, goes to | Verification | What it needs from you | |
|---|---|---|---|---|
| Beeminder | You stake | Beeminder. Stated business model. No beneficiary option. | Auto-tracked data | Your goal data, on their servers |
| StickK | You stake | Charity, a friend, or an "anti-charity" | Human referee | Reports, on their servers |
| Focus Tree (Starknet) | Platform pays you | n/a | In-app blocking | Session data, wallet |
| Focusmate | Neither | n/a | Live video, a stranger | A video call |
| Blockers (Opal, Freedom) | Neither | n/a | OS-level blocking | Deep OS access |
| DoomTax | You stake | Charity. Structurally never us. | Private AI + self-appeal | Your intention, your artifact, nothing else |
A product whose revenue arrives when you fail has an incentive it cannot design away, however good the intentions. StickK started where Beeminder still is and moved off it.
Focus Tree is the closest thing to DoomTax and it's the inverse design. Focus Tree pays you to focus, which means rewards must be funded from somewhere (token emissions) and creates something worth farming, which is why it needs distraction blocking to hold up. DoomTax emits nothing. Cheating means fighting yourself to win back your own money, a self-defeating loop, not an attack. That's also why our identity checks sit on the money, not the door: there's nothing here worth a sybil attacking.
- Stake what you're willing to lose on this session, small by default.
- State what you're actually trying to do. The verdict is judged only against that, never against a generic idea of "productivity."
- Focus. We don't watch you. We're a webview with no OS access, so we literally can't.
- Claim. A private AI coach decides kept or slipped. You get a short window to appeal if it got you wrong.
- Kept: refunded, streak grows, loudly celebrated.
- Slipped: quietly, your forfeit goes to charity. Bad days get an amnesty token, no questions.
A commitment device whose main output is "you lost money again" is a wellness product in the marketing copy only. These aren't bolt-ons, they're the product.
- You set the stake, small by default, confirm before any large jump.
- Amnesty. Bad-day tokens disarm a session before settlement, no questions.
- Forfeits never enrich anyone. Charity only, never us, never another user.
- Ambiguity always resolves toward you. Contested, failed inference, missing evidence: refund. A wrong "kept" costs nothing. A wrong "slipped" costs trust.
- Wins are loud, losses are quiet.
- No leaderboards, no ranking, no visible kept-rate. DoomTax is a personal app. If you want it social, you can join a circle: a group that picks a shared cause together. Every circle forfeit funds that one cause, and the app shows only the collective total, never who slipped. A bad day becomes something good for someone else, not a public score.
| Decision | Reasoning |
|---|---|
| Personal app, solo | Competing on kept-rate puts most pressure on whoever struggles most. Inverted for a wellness product. |
| No teams, pots, leaderboards, rankings | Cut, not deferred. A team pot also dilutes the stake by 1/N and a solo-controlled team means no penalty at all. |
| Circles are the social layer | Shared cause, collective total shown, individual contributions never shown. Makes a slip prosocial rather than punitive. Membership stays off-chain. |
| No accountability partner in v1 | Replaced by self-appeal: optimistic settlement, short window to contest, resolves toward the user. |
| Forfeits go to charity only | Never the platform, never an individual. This is the ethical load-bearing wall and the Beeminder differentiator. |
| Pending account before charity | Once money reaches a charity it cannot be reversed, so appeals and amnesty would be unfunded promises. |
| Honest custody | World App signs World Chain, not Hedera. The app provisions and holds a Hedera testnet account per user and records consent to HCS. Never write copy claiming the user signed the forfeit themselves. |
| Charity is a placeholder | No partnership exists. Mainnet target is The Giving Block, blocked on unconfirmed HBAR support and a commercial agreement. Say so in the README. |
| Business model: premium coach | Free tier returns a verdict, paid tier is the long-memory coach on 0G Storage. Makes 0G Storage commercially load-bearing. A fee on successful refunds is an open question, not decided. |
Verdicts are judged only against the intention the user stated at session start. Never against a general notion of productivity. Rest is productive.
Ambiguity always resolves toward the user. Contested, failed inference, timeout, missing evidence: refund. A wrong "kept" costs nothing. A wrong "slipped" costs trust.
The sections from here to the evidence checklist were written for the ETHGlobal judges. The dollar figures are the prize each sponsor offered for its track, not money won.
Every claim below is marked 🟢 built, 🟡 in progress, or ⚪ planned. Nothing is claimed working until it's green with a link next to it.
Not three logos on one flow. Each sponsor owns a job the product can't work without.
| Job | Why it can't be swapped out | |
|---|---|---|
| World | Guards the money, not the door | Liveness fires at the claim moment, the one point with a real financial incentive to cheat. |
| 0G | The brain | Attention data is diary-grade: what I meant to do and how I fell short is more intimate than browsing history. TEE-sealed inference is why anyone should trust this. |
| Hedera | The treasurer | Deterministic finality in seconds, predictable sub-cent fees, native scheduling with no contract overhead. The forfeit is pre-armed and dated; finishing your session disarms it. |
An autonomous Settlement Agent reads the session outcome and executes the
HBAR/HTS transfer on Hedera Testnet.
src/agent/settlement.ts · tx: 0.0.9695721-1785022437-972362991 ·
HashScan link ·
HCS topic: 0.0.9748699
Zero Solidity: find . -name "*.sol" returns nothing. Two native services are
live and proven on testnet:
- HCS: verdict and payout log, topic
0.0.9748699 - HTS: streak token
0.0.9762627(HashScan), minted and delivered on a kept verdict
Plus the escrow-then-sweep flow those services settle into. A forfeit lands in
a pending account (0.0.9762855), never at the charity, so it can still be
given back. Two scripts close the loop, both dry-run by default:
scripts/sweep-charity.tsmoves a forfeit to the charity (0.0.9762856) only after the appeal window closes and only if it was never contested (HashScan)scripts/refund-appeals.tsreturns an appealed forfeit to whoever staked it (HashScan)
Proven with three forfeits: the uncontested one swept to charity, the two appealed ones were returned to source. An appeal that only skipped the sweep would have left the stake in escrow forever, which is a slower way of keeping it, so "resolves toward you" means the money comes back.
Two privacy details. The sweep memo carries a count, never a session ID, so the public transaction cannot be used to work out who slipped. And the refund destination is derived from the settlement's transaction ID via the mirror node rather than stored, so the ledger never holds a durable link between a session and an account.
A third, Scheduled Transactions (HIP-423) for the pre-armed forfeit, is
written in src/hedera/schedule.ts but not yet wired into the session flow:
armForfeit() and disarmForfeit() have no caller. Counting it would be an
overclaim, so it is listed here as what it is.
Focus Coach runs on 0G Compute through two server-side paths:
- Broker path: wallet + on-chain ledger, with response attestation verification. This is the TEE-verified path.
- API-key path: a Router base URL or provider-scoped endpoint plus a prepaid key, simpler to operate, but no per-response attestation.
Both are wired in src/ai/coach.ts; the API-key path is selected by
ZG_COMPUTE_API_KEY (legacy ZG_ROUTER_API_KEY still works), otherwise the
broker path runs. Set ZG_COMPUTE_URL to either a Router base URL such as
https://router-api.0g.ai/v1 or a provider-scoped endpoint from pc.0g.ai
such as https://compute-network-27.integratenetwork.work, and choose the
model with ZG_COMPUTE_MODEL (for example claude-fable-5).
Not yet proven live. The code path is complete, but a live inference call
still depends on funded 0G infrastructure. The app fails safe to 'kept' when
the coach is unavailable, and reports whether the coach actually answered so a
fallback does not masquerade as a real verdict. The 0G endpoint choice is
configuration, not a rewrite: the repo now supports funded mainnet 0G by env,
while Hedera remains testnet-only. Tracked as #52.
Liveness at claim, not login. Testing doc: Selfie Check Testing wiki page
Submitted, not the design driver. See architecture doc for the honest framing.
Run one full 30 second session after pnpm dev, then fill these:
-
HashScan transfer URL: https://hashscan.io/testnet/transaction/0-0.9695721-1785022437.972362991
-
HCS topic ID:
0.0.9748699 -
HCS message JSON (must include only
sessionId,commitmentHash,verdict,amountTinybar,timestamp):{"sessionId":"497d1101-0cf0-40b9-b5d4-3608bdb6dc49","commitmentHash":"22749163d6ddcc0e4f8f9462fcd34a84409905dcecbcf0d1b8d30dadf5e0abfb","verdict":false,"amountTinybar":100000000,"timestamp":1785022446771} -
Request body for
POST /api/session/settle, captured from the real network request:{"sessionId":"497d1101-0cf0-40b9-b5d4-3608bdb6dc49","commitmentHash":"22749163d6ddcc0e4f8f9462fcd34a84409905dcecbcf0d1b8d30dadf5e0abfb","stakeHbar":1,"intention":"Finish the README evidence pass","artifact":"Updated the submission checklist to match the current client payload.","foregroundTime":28.4,"interruptionCount":0}Verified independently against the Hedera testnet mirror node, not just the app's own response. Both the transfer and the HCS message match the values above. The plaintext coach fields are sent only to the server-side coach path and never reach HCS.
-
HTS streak token (the kept-verdict reward), verified against the mirror node:
- Token:
0.0.9762627,STREAK, 0 decimals, infinite supply, treasury0.0.9695721(HashScan) - Mint:
0.0.9695721@1785039170.758949116· Transfer:0.0.9695721@1785039171.047140243 - Result: custody account
0.0.9762638holds1STREAK, token total supply1 - Reproduce:
npx tsx --env-file=.env.local scripts/check-streak.ts <accountId>
The settle route reaches this only for an authenticated user with a custody account and a
'kept'verdict, so the probe script is what makes the path demonstrable ahead of the phone test and a live coach.What never changes: the coach fields are transient.
settleSession()and the HCS write still only ever seecommitmentHashand the boolean verdict, never the intention or artifact text. That HCS-facing privacy guarantee holds in both demo mode and live-coach mode. - Token:
- Scheduled transactions execute on signature collection by default, which
would fire the forfeit at session start. You need
setWaitForExpiry(true)to evaluate atexpirationTime. That is HIP-423, max window two months.adminKeyis what lets youScheduleDeleteon success. Fallback: plain operator-held escrow, stated honestly. - 0G Compute and 0G Storage are different packages. Inference is
@0gfoundation/0g-compute-ts-sdk(renamed from@0glabs/0g-serving-broker). Storage is@0glabs/0g-ts-sdk. 0G is mid-migration between namespaces. Verify both on npm before building. Start from0g-compute-ts-starter-kit. The broker has explicit TEE verification: that is your attestation proof. - Mirror Node lags consensus by 1 to 3 seconds. Confirm from the transaction receipt, render success immediately, reconcile with Mirror Node afterwards. Polling Mirror Node for a balance makes the app look broken after the money has definitively moved.
- Idempotency: Hedera transaction IDs are client-generated. If no receipt comes back, query for that transaction ID before retrying. Never blind-retry.
- One atomic
TransferTransactionmoves everything at once. No partial states, one HashScan link. - Fund the operator account and handle
INSUFFICIENT_PAYER_BALANCEexplicitly. Testnet faucets are rate limited. - Codespaces port must be Public or the phone cannot load the mini app.
gh codespace ports visibility 3000:public -c $CODESPACE_NAME, or the Ports panel in VS Code. Verify in an incognito window, not by reading a log line.
We're a Next.js app in a World App webview. We have no access to Screen Time, usage stats, other apps, or the OS. That's not restraint, it's not on the table. The privacy claim is structural, not a promise.
The coach sees: your stated intention, your submitted artifact, and foreground-time/interruption metadata about our own page. The coach never sees: screen contents, browsing history, keystrokes, location, or any third-party data. We never collect it.
HCS is public and permanent, so it holds only a pseudonymous session ID, a commitment hash, a verdict, an amount, and a timestamp. Never intention text, artifact content, coaching messages, or circle membership.
A "what we can see about this session" panel shows every field we hold and nothing else. ⚪
Custody. World App signs World Chain, not Hedera. We provision and hold a Hedera testnet account for you, and record your consent to arm each forfeit on HCS. We are saying this plainly rather than implying you signed it yourself.
Charity. The recipient is a placeholder testnet account, configurable in one file. No partnership exists today. Mainnet target is The Giving Block, pending confirmation they support HBAR and a commercial agreement, neither of which is a weekend's work.
Not from your forfeit, that's the one thing we won't monetise. Premium coach: free tier returns a verdict, paid tier is the long-memory coach, encrypted history on 0G Storage, patterns recognised over months. It's the part of the product that gets better the longer you stay, and it's why 0G Storage is something we sell, not just something we use.
src/app/holds the pages and the API routes;api/session/settleis where a session settles.src/components/is the UI, one folder per component;SessionFlowdrives a session from stake to verdict.src/agent/andsrc/hedera/are the settlement agent and the Hedera services (HCS, HTS, scheduled transactions).src/ai/is the 0G coach,src/identity/the Selfie Check and AgentKit,src/lib/the session rules, appeal window and persistence.src/i18n/messages/holds the 14 locales;en-GB.jsonis the source every other locale falls back to.scripts/holds the operational scripts, every fund movement a dry run by default.docs/holds the specs, audits and build history, indexed in docs/README.md.

