HELP WANTED is a job board on which the employer can be a smart contract. The employer's IMD stays in escrow and is paid out on IdentityMD's on-chain receipt for that exact job, or, if the employer opted in, on a bonded delivery claim that has stood 24 hours with nobody challenging it, or that the order's arbiter upheld. Otherwise it goes back to the employer.
| 1. A contract escrows | 2. A runner buys the job | 3. The swarm works | 4. The WorkRegistry receipt | 5. The desk pays or refunds |
|---|---|---|---|---|
| The employer, a contract or a wallet, puts IMD in escrow on the desk against a written job. | A runner posts a bond, takes the order and pays IdentityMD 0.5 IMD for the job. | IdentityMD's AI agents do the job. | IdentityMD writes the hash of the job's work record into its WorkRegistry. | The desk checks the record against that receipt and pays the runner, or refunds the employer. |
That is the proven path. On an order whose employer opted in, a bonded delivery claim that nobody challenges for 24 hours is paid without waiting for step 4.
- The hire desk,
WorkOrders, on Ethereum mainnet:0x730CB6C845d896e33283d3A4Ed8967F8BAA778d2 - The site: https://web-production-a7c148.up.railway.app
- Check the central claim, no wallet needed: https://web-production-a7c148.up.railway.app/verify/42f26c64-eea9-46f2-822a-33547ec18abf: your browser reads IdentityMD's receipt for order 2's job from the chain and hashes the work record itself. Every claim, its proof and the limits: https://web-production-a7c148.up.railway.app/judges
- On X: https://x.com/HELPWANTED_imd
- The source: https://github.com/OoJae/help-wanted
IMD is the token of IdentityMD, a protocol whose network of AI agents, the swarm, takes paid jobs. When the swarm finishes a job, IdentityMD writes the hash of the job's work record into a contract of its own on Ethereum, the WorkRegistry. That entry is the job's receipt. The desk pays only on the receipt for that exact job, or, on an order whose employer opted in, on a bonded claim that nobody challenged or that the order's arbiter upheld. A receipt pays only if the job's work record names the order: its objective is the order's header followed by the order's job text.
What is new here. A contract can already buy a job from IdentityMD: its Intake contract takes a
contract's request and bills it up front, in the same transaction (IdentityMD's documents, read on
15 October 2026, say jobs have been sold there since 13 October 2026). x402, IdentityMD's other way
to pay, needs an ordinary wallet's signature, and is paid before the work starts too. Neither holds
the IMD back until the work is recorded. HELP WANTED does: it holds the IMD in escrow, pays on
IdentityMD's on-chain receipt for that exact job, and refunds it otherwise (or, where the employer
opted in, pays a bonded claim that nobody challenged or that the order's arbiter upheld). The desk
does not ask how a job was paid for. The team's runner buys each job through IdentityMD's public API
(the action job.open, paid by x402), not through the Intake contract.
It has five parts, in ten contracts, in three launches, all three live on Ethereum mainnet. No contract has an owner, an admin or a setter, and there is no token of ours.
| Part | What it does | Launch | Live today |
|---|---|---|---|
| Standing Questions | A contract that buys its own oracle answers. IdentityMD's oracle has a panel of its agents answer a question about the chain, and signs the answer. Anyone creates a feed: a question, asked again every so many blocks. Anyone puts IMD on the feed's meter. When the feed is due, the contract pays for the answer from the meter. The signed answer comes back when IdentityMD calls the contract (the callback), or anyone hands it in (submit). The contract checks it and stores it. |
A | Yes |
| The Robot Jobs Report | A public record of what the swarm was paid: IMD paid and number of payments, each taken from a Standing Question. One print so far (13 October 2026). | A | Yes |
| Payroll | Splits every IMD it receives by fixed code: half to those who stake IMD in IdentityMD's vault, 30% to the owners of IdentityMD's 2,000 seats (the tokens its agents work under), 20% to keep the report's questions funded. | A | Yes |
| The hire desk | The job board itself. An employer escrows IMD. A runner, which is anyone who posts a bond, buys the job from the swarm. The escrow pays against the job's receipt, or, if the employer allows it, against a bonded delivery claim nobody challenges for 24 hours or that the order's arbiter upholds. Two employers that are contracts come with it: a crowd pot, and a doorbell anyone can ring until it hires. | B | Yes. Its first order was paid on proof on 14 October 2026, and its second, opened by a rule of launch C, on 16 October 2026 |
| The company | Rules of the form "when this is true, hire", endowments whose yield pays for a question or a pot, and a way to pay with ETH. | C | Yes, since 15 October 2026. On 16 October 2026 the team's wallet fired its first rule by hand, and the order it opened was paid on proof |
HELP WANTED is an entry in IdentityMD's builder competition.
Where money waits for work, somebody decides when it moves. On the hire desk that is code: it pays when the job's work record matches IdentityMD's on-chain receipt and names the order, and it refunds otherwise. The other rows are what each one's published terms said when read on 20 October 2026 (two figures, marked, are reported by others).
| Where | Who decides payment | Until money moves | Refund | Source |
|---|---|---|---|---|
| HELP WANTED, receipt path | Code: WorkProof holds the work record against the WorkRegistry receipt and the order header |
In the settling transaction; order 2 was paid 8 min 48 s after its rule fired | Yes: after the deadline, anyone may send it back | Check 7 below |
| Upwork, fixed price | The client; silence approves after 14 days; disputes go to Upwork, then paid arbitration | Up to 14 days of review, then a 5-day hold | Through a dispute | getting paid, disputes |
| Fiverr | The buyer; the order completes by itself 3 days after delivery | 3 days (up to 5 more if extended), then 14 days of clearance | By cancellation | order statuses, clearance, reported |
| Bountycaster, a bounty board | Only the poster | Whenever the poster pays | Nothing is held | FAQ |
| ERC-8183, and Virtuals ACP v2 on it | One evaluator named at the start; no dispute step | When the evaluator completes the job | Yes, on rejection or after expiry | EIP-8183, ACP changelog |
| x402r, escrow for x402 payments (beta) | An arbiter | After a 14-day timelock, unless released sooner | Yes, released by the arbiter | x402r, timelock, reported |
| IdentityMD's Intake | Nobody after the fact: the fee is taken in the call that opens the job | Before the work starts | "Refunds: None", in its docs | IdentityMD docs |
An order on the desk is in the shape of ERC-8183: the employer is the client, the runner the provider, the receipt check the evaluator. The difference is that the evaluator is a fixed check of a receipt IdentityMD wrote, not a party someone names. What the desk proves is delivery, not quality: it pays a completed job of the order's kind whose work record names the order, and IdentityMD's own verification of a research report is structural (for order 2's job: "paths and tree verified; no suite was run for this kind of work"). It is not for work only a person can judge.
- New on 20 October 2026, all of it the team's own use. Anyone can now ring the Doorbell in one
step from https://web-production-a7c148.up.railway.app/rules/new#doorbell, with a browser
wallet or a phone wallet's link or QR code; ring 26 was the team's wallet W1, through that page.
The first crowd pot was a refund drill: H1 opened pot 1 with one chip of 0.01 IMD and 72 minutes
to collect, it ran out of time, and
claimRefundpaid the chip back to the wei (0x61a72aef…5618). H1, which paid for both orders' swarm jobs, recorded one review of each job's agent in the ERC-8004 reputation registry on mainnet, saying that delivery was checked against the receipt and quality was not (0xc68d8e8a…f24f,0x5e5eba6d…d5d6). An AI agent can use the site through a key-free MCP server at/mcp("Use it from an AI agent", below). - IdentityMD's receipts stalled again, and are flowing again. Its WorkRegistry wrote no receipt from 19 October 2026, 06:39 UTC (block 26,224,887) until 23 October 2026, 21:17 UTC (block 26,257,966), about four days and fifteen hours, then cleared its backlog. While a stall lasts, an order that pays only on a receipt waits for one and is refunded at its deadline if none comes. The site reads the registry live and says whether receipts are arriving on every page that depends on it.
- Launch A is live on Ethereum mainnet since 13 October 2026. Three contracts, created by IdentityMD's launch pipeline in one transaction (IdentityMD launch 1531).
- The first Robot Jobs Report was printed that day: 112 IMD paid to the swarm in blocks 26,176,608 to 26,183,607, and 223 payments counted in blocks 26,176,610 to 26,183,609. The week's top employers, the third question, were answered by IdentityMD's own callback.
- Launch B, the hire desk, is live on Ethereum mainnet since the same day. Four contracts,
created in one transaction (IdentityMD launch 1573). The pipeline changed no compiled code: it
corrected three comments in
CrowdPots.sol(launches/b-desk/docs/AUDITS.md, row 9). - Launch C went live on Ethereum mainnet on 15 October 2026, at 13:18 UTC. Three contracts,
created in one transaction (IdentityMD launch 1845, from pipeline job
7f2023fc-fa1f-4a7f-a507-3bbc712c51d6). The pipeline changed no file undersrc/. Its reviewers recorded 13 findings, six low and seven info, none marked blocking. No node of the pipeline answered ten of them, which come down to four matters not raised before and one thatlaunches/c-company/ADAPTATION.mdhad already answered, and the contracts were deployed as they were (launches/c-company/docs/AUDITS.mdlists all 13). The deployments book (packages/shared/src/deployments.ts) holds launch C since the same day, andEVIDENCE.mdcounts it. - The desk's first order was opened by a contract and paid on proof. The team's own wallet
rang the Doorbell 25 times, and the 25th ring made the Doorbell, a contract with no owner, open
order 1: one picture, 1 IMD in escrow
(
0xa9c9f256…15fb, 13 October). The team's runner took it and paid IdentityMD 0.5 IMD by x402 for the picture job (0xc92c4731…6bc5). On 14 October, a minute after IdentityMD wrote the job's receipt into its WorkRegistry, the runner settled the order with the job's work record (0xa3f5b783…7628): the desk checked the record against the receipt, then paid the runner 0.9 IMD and its bond, and Payroll its 0.05 IMD fee and a 0.05 IMD tip, which Payroll credited to agent 51519, the agent the work record names. On 15 October Payroll paid that tip to the owner of the seat the agent is bound to. Check 6 below repeats the proof. - The desk's second order was opened by a rule and paid on proof, and all of it was the team's.
On 16 October 2026 the team's wallet H1 created rule 1 on
Rules, the first rule oflaunches/c-company/docs/LAUNCH.md(section 6): when launch A's hires feed reports at least 100, hire a research report that only a receipt can pay. Anyone may fire a rule whose condition holds. This time H1 fired it, by hand, withscripts/first-rule.sh. The fire read the hires feed (223) and opened order 2 on the desk, withRulesas its employer and 0.60 IMD in escrow (0x9f790aae…e37b). The team's runner, hosted and running unattended, took it, paid IdentityMD 0.5 IMD by x402 for the job, and settled the order on the job's receipt under nine minutes after the fire (0x35bb49cc…0516). The desk paid the runner 0.504045 IMD and its bond, and Payroll a 0.0280025 IMD fee and a 0.0280025 IMD tip for agent 52142, the agent the work record names. It sent the 0.03995 IMD the order did not spend back toRules, which credited it to the rule. H1 then took that back by cancelling the rule, and calledclaimTip, which paid the tip to the owner of seat 1978. Check 7 below repeats the proof. It shows a rule hiring the swarm on mainnet against a real IdentityMD receipt. It is not use by anyone outside the team: the rule's creator, the wallet that fired it and the runner were all H1. - Nobody outside the team has used the contracts. Every transaction that touched them after
their launch was sent by the team's own wallets, apart from one: IdentityMD's Intake delivering an
oracle answer the team had paid for. The contracts have still sent IMD outside the team: to
IdentityMD's fee address, which every paid request to IdentityMD pays (
Questionspaid it 0.5 IMD for each of the three answers it bought); to sIMD, IdentityMD's staking vault; and to two wallets: the owner of seat 1049, paid the tip of agent 51519, and, on 16 October, the owner of seat 1978, paid the tip of agent 52142, the agent order 2's work record names. Neither owner did anything to be paid: a tip goes to the owner of the seat bound to the agent a work record names, and receiving one is not use of HELP WANTED. "What is not shown yet", below, has the figures.
EVIDENCE.md is the record of what exists. It is generated from the chain at one
pinned block, which it names, by a command you can run, for every launch the deployments book
holds; check 8 below says when it was last run. Where this page and that file differ about what
existed at that block, believe that file. What happened after that block is not in it until the
command is run again.
IdentityMD deploys contracts through its launch pipeline: its own agents read the source, may change it, and create the contracts in one transaction. A launch is that transaction.
| Launch | Contracts | Status | Addresses | Launch transaction | Source | Known limits |
|---|---|---|---|---|---|---|
| A | Questions, JobsReport, Payroll |
Live since 13 October 2026. IdentityMD launch 1531. | Questions 0xEECE4628ceb091C13393fc64056Be8b8501fD3F6JobsReport 0x58963B57E179f7fE4E95090C21E8e9D75353BCE4Payroll 0x6e4362a5cEc59Eb3F611af6b2d81915CF5b45f1F |
0x9ce6111c9f454373ac208cc9b54ebd0a4b09e58706f20e568f0f9e8bbb392c58, block 26,182,869 |
launches/a-payroll |
launches/a-payroll/docs/THREATS.md |
| B | WorkProof, WorkOrders, CrowdPots, Doorbell |
Live since 13 October 2026. IdentityMD launch 1573. | WorkProof 0xB268297dB06985ba0b806Ba97BcB90364b137469WorkOrders 0x730CB6C845d896e33283d3A4Ed8967F8BAA778d2CrowdPots 0x8C2FfCdcb103659aa05c772b8F2059868B496EB5Doorbell 0xFd5ab7E03E8554210Faa3cCd3b9D47C622fffcCb |
0xfb28bef9c477cfe3777231d8c41f202e691673fc4cec98ecb8b07f1bfff7bf9c, block 26,184,581 |
launches/b-desk |
launches/b-desk/docs/THREATS.md |
| C | Rules, Endowments, Zap |
Live since 15 October 2026. IdentityMD launch 1845. It takes addresses that launches A and B created. | Rules 0xcbF0dddF07C00a78E2E7561a6F47C275B8c05D7EEndowments 0xbE916A53b9Bbe634AF7E10E604E06b6069edEE76Zap 0x51bF8c74bd9BD74Ad08cABDC90DaC7835C18B21A |
0x1806c02e5e38d4c8496476b3f3cf39f250a9d0c97ff5834c75a0475a12bd62fb, block 26,198,165 |
launches/c-company |
launches/c-company/docs/THREATS.md |
This table is written by hand, and a test fails when it disagrees with the deployments book
(packages/shared/src/deployments.ts), the one file where a launch's addresses are recorded. The
generated table, with the code size and code hash at each address, is in
EVIDENCE.md.
Each launch folder is a Foundry project that stands alone, with its own README, tests and
documents, because the pipeline imports a repository whose root is the project. That is why the
public repository, https://github.com/OoJae/help-wanted, also has a branch per launch
(launch-a, launch-b, launch-c) that holds one launch folder as its root.
Every command below was run on 13, 15 or 16 October 2026 and printed what is shown. They read
only. You need Foundry (cast, forge), curl, jq, xxd and bc.
Check 2 also needs Node 18 or newer; check 8 needs Node 22 and pnpm 11.
export RPC=https://ethereum-rpc.publicnode.com
export LOGS_RPC=https://rpc.mevblocker.io # for old logs: the first node keeps about 9,000 blockskeccak256 of a job's work record, exactly as IdentityMD serves it, is what the WorkRegistry holds
for that job. For the real job e9efec16-56f3-40d3-9243-52c453effc7d:
curl -s https://api.imd.fun/work-records/3db993254c94479c4bf74d1115c7db3bfb48e5859379a4e83467ab6c3338b2c5.json -o record.json
cast keccak "0x$(xxd -p record.json | tr -d '\n')"
cast call 0xb6d0a187b050FA5Bb0B87033a203f37BECF4a775 "latest(bytes16)(bytes32)" \
0xe9efec1656f340d3924352c453effc7d --rpc-url $RPCBoth print 0x3db993254c94479c4bf74d1115c7db3bfb48e5859379a4e83467ab6c3338b2c5. The job id is the
UUID without its dashes. The same check through this project's own launched WorkProof, which
reads the registry itself and answers only for the bytes whose hash the registry holds:
cast call 0xB268297dB06985ba0b806Ba97BcB90364b137469 \
"prove(bytes16,bytes,uint256)((bytes32,bytes32,bytes32,bytes32,bytes32,uint256[]))" \
0xe9efec1656f340d3924352c453effc7d "0x$(xxd -p record.json | tr -d '\n')" 0 --rpc-url $RPCIt answers with the record hash first and, last, [52211]: the agent the record names as having
done the work. Change one byte of record.json and the call reverts.
bash scripts/repro.sh # or: pnpm reproIt builds each launched project into a temporary folder, fetches the code at each address, sets the immutable ranges to zero in both (those bytes hold constructor arguments) and compares the rest. Printed on 15 October 2026, with Foundry 1.8.3:
MATCH launch A Questions 0xeece4628ceb091c13393fc64056be8b8501fd3f6 built 22233 bytes, on chain 22233 bytes, 18 immutable ranges (576 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0xf876f5265e4ae5015645fbbee2c47e280391c23fc8256b46b688688a388618f7
MATCH launch A JobsReport 0x58963b57e179f7fe4e95090c21e8e9d75353bce4 built 9320 bytes, on chain 9320 bytes, 9 immutable ranges (288 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0xb050a81690b40963e1cbdffb17aafb3fcdf42ddd432a05023be7e23873207432
MATCH launch A Payroll 0x6e4362a5cec59eb3f611af6b2d81915cf5b45f1f built 11846 bytes, on chain 11846 bytes, 36 immutable ranges (1152 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0xa962b0633998d594b2720984ad5f76fce369148592da0762b9d9acf399367236
MATCH launch B WorkProof 0xb268297db06985ba0b806ba97bcb90364b137469 built 4579 bytes, on chain 4579 bytes, 2 immutable ranges (64 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0x0a514db427f7f3845166e923929d67816f4a27e661d8ed510b7e910d730e5fba
MATCH launch B WorkOrders 0x730cb6c845d896e33283d3a4ed8967f8baa778d2 built 19289 bytes, on chain 19289 bytes, 13 immutable ranges (416 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0xb06dfaf6410132828e12647f4ee5199a092f8c0fb9f0c0d2275b5f655262ffb3
MATCH launch B CrowdPots 0x8c2ffcdcb103659aa05c772b8f2059868b496eb5 built 14177 bytes, on chain 14177 bytes, 27 immutable ranges (864 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0x282560809f9474cbb416a8d6235fac38fe2679ff3e729a87ff88859006beb1c2
MATCH launch B Doorbell 0xfd5ab7e03e8554210faa3ccd3b9d47c622fffccb built 5972 bytes, on chain 5972 bytes, 40 immutable ranges (1280 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0xda26e8196bf95c4117124ef25c308be050bee1823ff5c6ccf287a5d788bfe43c
MATCH launch C Rules 0xcbf0dddf07c00a78e2e7561a6f47c275b8c05d7e built 13762 bytes, on chain 13762 bytes, 25 immutable ranges (800 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0xec51b0077f235e771b3309380d222b726b50632d97c99179a4f22af356089ff9
MATCH launch C Endowments 0xbe916a53b9bbe634af7e10e604e06b6069edee76 built 13161 bytes, on chain 13161 bytes, 37 immutable ranges (1184 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0xa87d11ad30a06bec5aa1293e7f7f4bfb6f0c84930057a5b9dc3e2253cc78cdbc
MATCH launch C Zap 0x51bf8c74bd9bd74ad08cabdc90dac7835c18b21a built 3834 bytes, on chain 3834 bytes, 14 immutable ranges (448 bytes) set to zero in both: equal outside the immutable ranges. keccak256 of the on-chain code: 0x76271cf45487ed5b6b74adb4f0048a0918ce073bbbd3fac18e4bf26759fef63f
repro: 10 of 10 launched contracts match the code built from this repository.
It exits with a status other than zero when a contract differs or cannot be read, and when the
deployments book gives it no contract to compare. REPRO_IMMUTABLES=1 also prints every
constructor argument found on chain, to hold against the manifest in each launch's
docs/LAUNCH.md. The script needs Foundry, Node 18 or newer and an RPC, and installs nothing.
What the explorers say of launches A and B, read on 15 October 2026. Etherscan shows each of the
seven contracts as "Source Code Verified", "Exact Match", under its own name, compiled with solc
0.8.26 and 200 optimizer runs. Sourcify holds a match for CrowdPots only (creation and runtime
code, verified on 13 October 2026 at 17:48 UTC) and none for the other six. IdentityMD's own record
of launches A and B gives each of their contracts a verifiedAt time on 13 October 2026, within
three minutes of its launch; EVIDENCE.md prints them. None of this is needed for
the comparison above, which holds whatever an explorer shows.
What the explorers say of launch C, read on 15 October 2026 at about 16:00 UTC. Etherscan shows
Rules, Endowments and Zap each as "Source Code Verified", "Exact Match", under its own name,
compiled with solc 0.8.26 and 200 optimizer runs. Sourcify holds a match (creation and runtime
code) for Endowments and Zap, verified on 15 October 2026 at 15:01 UTC, and for Rules it
answers with no match at all (HTTP 404). IdentityMD's record of launch 1845 gives the three
verifiedAt times from 13:21:19 to 13:21:35 UTC that day, under three minutes after the launch;
EVIDENCE.md prints them.
What the contract holds:
cast call 0x58963B57E179f7fE4E95090C21E8e9D75353BCE4 \
"printAt(uint256)((uint256,uint256,uint64,uint64,uint64,uint64))" 0 --rpc-url $RPC
# (112000000000000000000 [1.12e20], 223, 26176608 [2.617e7], 26183607 [2.618e7], 1, 1)That is 112 IMD (18 decimals), 223 payments, and the block window of the amount. The same two
figures, counted by you from the token's own Transfer logs to IdentityMD's fee address:
IMD=0xD34a99Bc0f67aE1bbd63C660e6d0b0dd03E263B7
TRANSFER=$(cast keccak "Transfer(address,address,uint256)")
FEE_ADDRESS=0x0000000000000000000000004e0fa57bde726079356537e2f34d671e9f41adbc
# IMD paid in blocks 26,176,608 to 26,183,607. Prints 112.000000000000000000
cast logs --address $IMD --from-block 26176608 --to-block 26183607 $TRANSFER "" $FEE_ADDRESS \
--rpc-url $LOGS_RPC --json | jq -r '.[].data' | while read -r d; do cast to-dec "$d"; done \
| paste -sd+ - | bc | xargs cast from-wei
# Payments in blocks 26,176,610 to 26,183,609. Prints 223
cast logs --address $IMD --from-block 26176610 --to-block 26183609 $TRANSFER "" $FEE_ADDRESS \
--rpc-url $LOGS_RPC --json | jq lengthThe two windows differ by two blocks because the amount and the count come from two questions,
asked in two transactions. These logs are older than the roughly 9,000 blocks
ethereum-rpc.publicnode.com serves, and on 15 October 2026 it refused them ("Archive requests
require a personal token"), so the two commands read them from $LOGS_RPC. On that day
rpc.mevblocker.io answered each in one query, and the two commands printed
112.000000000000000000 and 223. Any node that keeps history will do. The contract read above
works on either node.
curl -s https://api.imd.fun/oracle/requests/f8ffb991-55b7-4316-9461-44a94d5270aa \
| jq '{status, figure: .attestation.figure, window, signer, consumer}'
cast call 0xEECE4628ceb091C13393fc64056Be8b8501fD3F6 \
"latest(bytes32)(uint256,uint64,uint64,uint64,uint64)" \
0x4bf83605cd50e9658472bde9be314bb3848a563301fa793328bef17ad2a07b61 --rpc-url $RPCThe first shows IdentityMD's record of the oracle request: status attested, figure
112000000000000000000, the window 26,176,608 to 26,183,607, and consumer.verifyingContract
equal to our Questions address. The second shows the same figure and window stored in
Questions, which accepts an answer only after checking the oracle's signature, the question, the
window and the closing block's hash.
A runner must put 85 bytes, the order's header, at the start of the job's objective, so that the
finished job's record names the order it was done for. One real job was bought that way on
13 October 2026 (job e70c56ce-60f0-4c2c-b413-67e5ae37ddce):
curl -s https://api.imd.fun/work-records/990ae775288252e4b1b12c4a85bbe66db0140ca733e30002a6525814bfdaa17d.json -o header.json
cast keccak "0x$(xxd -p header.json | tr -d '\n')" # 0x990ae775…a17d, the hash it is served under
jq -j .objective header.json | head -c 85 # [help-wanted order de7eddaa…2041] and a space
cast call 0xb6d0a187b050FA5Bb0B87033a203f37BECF4a775 "latest(bytes16)(bytes32)" \
0xe70c56ce60f04c2cb41367e5ae37ddce --rpc-url $RPCOn 13 October 2026 the last command printed 32 zero bytes: IdentityMD's receipt writer had
stalled since about 8 October. It wrote again from 13 October at 17:26 UTC (block 26,185,055),
working through its backlog first, and this job's receipt landed on 14 October (block 26,189,975).
On 15 October the command prints
0x990ae775288252e4b1b12c4a85bbe66db0140ca733e30002a6525814bfdaa17d, the hash the record is
served under. The header in this job named no order (it was bought before the desk was launched);
check 6 is the same proof for a job that did.
The runner bought the job by x402: IdentityMD's fee address sent the settlement through x402's
contract, and it moved 0.5 IMD from the runner's wallet to that address
(0xc92c4731…6bc5,
block 26,185,015). IdentityMD names the same wallet as the payer of the job:
curl -s https://api.imd.fun/jobs/f7bd80ba-3e8b-483f-ad02-6ffbd16fef92 | jq '{state, template, paidBy}'
# completed, skill:create-image, paid by 0x60cc…9560 (H1, the runner's wallet)The order on the desk, the job that was bought for it, and the receipt IdentityMD wrote for that job:
DESK=0x730CB6C845d896e33283d3A4Ed8967F8BAA778d2
cast call $DESK "headerOf(uint256)(bytes)" 1 --rpc-url $RPC | xargs cast to-utf8
# [help-wanted order bd87367fc83372def576d4f8a370d72e295e10e2572f796ec8f1af720fa36d80] and a space
cast call 0xb6d0a187b050FA5Bb0B87033a203f37BECF4a775 "latest(bytes16)(bytes32)" \
0xf7bd80ba3e8b483fad026ffbd16fef92 --rpc-url $RPC
# 0x59f81da0ce7cde62693fca46e8dd10411c1dfae52102ffd68d25a66821451b68
curl -s https://api.imd.fun/work-records/59f81da0ce7cde62693fca46e8dd10411c1dfae52102ffd68d25a66821451b68.json -o picture.json
cast keccak "0x$(xxd -p picture.json | tr -d '\n')" # the same hash
jq -j .objective picture.json | head -c 85 # the same header
jq -r '.kind, .state' picture.json # skill:create-image, completedAnd through the launched WorkProof, which reads the registry itself: the record's hash, the hash
of the kind the Doorbell pins, of the header, of the Doorbell's job text, and the agent the record
names.
cast call 0xB268297dB06985ba0b806Ba97BcB90364b137469 \
"prove(bytes16,bytes,uint256)((bytes32,bytes32,bytes32,bytes32,bytes32,uint256[]))" \
0xf7bd80ba3e8b483fad026ffbd16fef92 "0x$(xxd -p picture.json | tr -d '\n')" 85 --rpc-url $RPC
# (0x59f81da0…1b68, 0x7679ff00…9185, 0x88784319…39c6, 0xc8faed8d…dba6, …, [51519])The order itself is Delivered (state 6) and marked proven (the twentieth field of orderOf,
true), with all of its 1 IMD spent:
cast call $DESK "accountOf(uint256)(uint8,address,uint96,uint96,uint96)" 1 --rpc-url $RPC
# 6, the Doorbell, 0 left in escrow, 1000000000000000000 spent, 0 returnedSee the picture the Doorbell paid for
(IdentityMD's explorer, job f7bd80ba-3e8b-483f-ad02-6ffbd16fef92).
The team's wallet H1 created rule 1 on Rules
(0x8c69fca9…548b,
block 26,204,482) and fired it by hand with scripts/first-rule.sh
(0x9f790aae…e37b,
block 26,204,499, 10:28 UTC). The fire read launch A's hires feed (223, at least 100) and opened
order 2 on the desk with Rules as the employer. Anyone may fire a rule whose condition holds;
this time the team's own wallet did, and the rule's 0.01 IMD tip went back to it.
The team's runner, which signs as H1, claimed the order at 0.56005 IMD
(0x27fc5a9f…0cb1,
block 26,204,508) and bought the job by x402: IdentityMD's fee address sent the settlement through
x402's contract, which moved 0.5 IMD from the runner's wallet to that address
(0x2fc7fa99…ca70,
block 26,204,514). IdentityMD's writer wrote the job's receipt
(0x3a604fc1…5abf,
block 26,204,533), two blocks before the runner registered the job
(0x20ef3770…4467,
block 26,204,535). The runner settled the order on that receipt
(0x35bb49cc…0516,
block 26,204,543, 10:37 UTC), 44 blocks and under nine minutes after the fire. IdentityMD names the
same wallet as the payer of the job:
curl -s https://api.imd.fun/jobs/42f26c64-eea9-46f2-822a-33547ec18abf | jq '{state, template, paidBy}'
# completed, skill:research-report, paid by 0x60cc…9560 (H1, the runner's wallet)The order on the desk, the job that was bought for it, and the receipt IdentityMD wrote for that job:
cast call $DESK "headerOf(uint256)(bytes)" 2 --rpc-url $RPC | xargs cast to-utf8
# [help-wanted order bd623ece89b1c600768011d08b9f77f00f169c6b6f31d2e3c37d61515189c407] and a space
cast call 0xb6d0a187b050FA5Bb0B87033a203f37BECF4a775 "latest(bytes16)(bytes32)" \
0x42f26c64eea946f2822a33547ec18abf --rpc-url $RPC
# 0xf6d0bd11b43fc36a3720db09a65612a03a4be7b5b12476bcdace4d1383176a85
curl -s https://api.imd.fun/work-records/f6d0bd11b43fc36a3720db09a65612a03a4be7b5b12476bcdace4d1383176a85.json -o report.json
cast keccak "0x$(xxd -p report.json | tr -d '\n')" # the same hash (the record is 3,801 bytes)
jq -j .objective report.json | head -c 85 # the same header
jq -r '.kind, .state' report.json # skill:research-report, completedAnd through the launched WorkProof: the record's hash, the hash of the kind the rule pins, of the
header, of the rule's job text, and the agent the record names.
cast call 0xB268297dB06985ba0b806Ba97BcB90364b137469 \
"prove(bytes16,bytes,uint256)((bytes32,bytes32,bytes32,bytes32,bytes32,uint256[]))" \
0x42f26c64eea946f2822a33547ec18abf "0x$(xxd -p report.json | tr -d '\n')" 85 --rpc-url $RPC
# (0xf6d0bd11…6a85, 0x1eca7729…988b, 0x390b14f4…3f8f, 0xc6413a52…31e2, …, [52142])The order is Delivered (state 6) and marked proven, with 0.56005 IMD spent and the 0.03995 IMD
it did not spend returned to Rules. The settlement's Delivered event gives the split:
cast call $DESK "stateOf(uint256)(uint8)" 2 --rpc-url $RPC
# 6
cast call $DESK "accountOf(uint256)(uint8,address,uint96,uint96,uint96)" 2 --rpc-url $RPC
# 6, Rules, 0 left in escrow, 560050000000000000 spent, 39950000000000000 returned
cast logs --address $DESK --from-block 26204543 --to-block 26204543 \
"Delivered(uint256 indexed,bytes16 indexed,address indexed,bytes32,bool,uint256,uint256,uint256,uint256,uint256,uint256[])" 2 \
--rpc-url $LOGS_RPC --json | jq -r '.[0].data' \
| xargs cast decode-abi --input "x(bytes32,bool,uint256,uint256,uint256,uint256,uint256,uint256[])"
# the record hash; true (proven); 504045000000000000 to the runner; 28002500000000000 fee and
# 28002500000000000 tip to Payroll; 39950000000000000 back to Rules; 250000000000000000, the bond,
# back to the runner; [52142]Rules received the 0.03995 IMD inside the settlement, through the desk's callback, and credited
it to rule 1. H1 cancelled the rule, which had used its one fire, and took that amount back
(0x9c062d89…6846,
block 26,204,586). H1 then called claimTip(52142), and Payroll paid the 0.0280025 IMD tip to the
owner of seat 1978, the seat agent 52142 is bound to, a wallet outside the team that did nothing to
be paid
(0xd7a75e9c…f273,
block 26,204,587). The whole hire cost the team 0.556005 IMD and gas: IdentityMD's 0.5 IMD job fee,
and the fee and the tip paid to Payroll (H1 held 1.4 IMD before block 26,204,480 and 0.843995 IMD
after block 26,204,587).
The report counts the payments of IdentityMD's job fee in blocks 26,197,518 to 26,204,517: 220
transfers of 0.5 IMD to the fee address, 110 IMD in all, from 52 wallets, one row per wallet, each
row linking one of that wallet's transactions. We counted the same logs again (with $IMD,
$TRANSFER and $FEE_ADDRESS from check 3) and got the same figures. Among them are the team's
own payment for this job and one by the team's wallet W1.
cast logs --address $IMD --from-block 26197518 --to-block 26204517 $TRANSFER "" $FEE_ADDRESS \
--rpc-url $LOGS_RPC --json | jq '[length, ([.[].topics[1]] | unique | length), ([.[].data] | unique | length)]'
# [220, 52, 1]: 220 transfers, 52 senders, one amount (0.5 IMD)See the report the rule paid for
(report.md, 10,523 bytes, served by IdentityMD's API; IdentityMD's explorer has the job at
https://explorer.imd.fun/jobs/42f26c64-eea9-46f2-822a-33547ec18abf).
pnpm install
pnpm evidencewrites EVIDENCE.md and evidence/evidence.json again from the chain. It prints
the block it pinned, holds every count against the contracts' own counters at that block, and
writes nothing if a read fails or a count disagrees. It holds no key and sends no transaction. With
no setting it reads from https://ethereum-rpc.publicnode.com,https://rpc.mevblocker.io, and asks
the first node for each range of blocks before the second. The first serves logs for about the last
9,000 blocks only, so older logs, such as the early logs of launches A and B, come from the second.
Run this way on 20 October 2026, it pinned block 26,234,773, read logs in ranges of 2,000 blocks and
found all 56 of its checks against the contracts in agreement. RPC_URL (one URL, or several
separated by commas) and CHUNK_SIZE change that.
Stated as plainly as what is.
- Launch C went live on 15 October 2026, at 13:18 UTC. By block 26,198,957 every event its
contracts had emitted came from four calls of the team's wallet W1 to
Endowments: creating an endowment for the meter of the payrolls feed, staking, depositing 1 IMD and staking again. No rule had been created then, and nothing had gone throughZap. On 16 October 2026 the team's wallet H1 created the first rule (rule 1), fired it by hand, and cancelled it once its one fire was used and its order paid (check 7); that dayruleCount()read 1, and nothing had yet gone throughZap. - The endowment's principal came back to the wei; its yield was never enough to harvest. On
19 October 2026 W1 called
withdraw(1, 1 IMD)and received exactly 1 IMD (0xcb48313a…024e, block 26,226,427):Withdrawn(1, W1, principal 1 IMD, imdOut 1 IMD, sharesOut 0), andprincipalOf(1, W1)reads 0. Nothing was harvested.harvestruns only forMIN_HARVEST(0.001 IMD) or more, andharvestable(1)read 0 from the deposit to the withdrawal: sIMD's price (convertToAssets) did not move at all from block 26,198,165 (15 October) to block 26,226,427, so the staked IMD earned nothing in that time. The 0.000001 IMD seed paid to create the endowment is not refundable and stays in it. So the drill shows a deposit staked into IdentityMD's vault and taken back in full, not yield paying for a question. - Two orders, and all of them the team's own. Order 1 was opened by the team's Doorbell after
the team's own wallet rang it. Order 2 was opened by rule 1 on
Rules, which the team's wallet H1 created and fired. The team's runner took both. No outside employer, runner, ringer or rule creator has used the desk. The one crowd pot, pot 1, was the team's refund drill of 20 October 2026. - IdentityMD's receipts stalled for about six days. The WorkRegistry wrote no receipt from
about 8 October until 13 October 2026, 17:26 UTC (block 26,185,055). It then worked through its
backlog, and the receipts of jobs bought on 13 October, order 1's among them, landed on
14 October. While receipts were stalled an order that pays only on a receipt could not be paid
(the limit named W1 in
launches/b-desk/docs/THREATS.md). The Doorbell was launched during the stall with the desk's other path switched on, for good: a bonded delivery claim paid after 24 hours without a challenge, with its own limits (L1, L2 and D2 in the same file). Order 1 used that path to claim delivery, and was then paid on its receipt instead, which landed first. The writer then kept writing until 19 October 2026, 06:39 UTC (block 26,224,887), stalled again, and wrote again from 23 October 2026, 21:17 UTC (block 26,257,966). On 16 October it wrote the receipt of order 2's job 16 seconds after IdentityMD listed the job as completed (10:35:18.756 UTC; block 26,204,533 at 10:35:35 UTC), two blocks before the runner registered the job. Order 2 pays only on a receipt, and was paid on that one. - Nobody outside the team has used anything. Every transaction that touched the contracts
after their launch came from the team's wallets, apart from the Intake's one delivery.
EVIDENCE.mdhas the number and splits every count HOUSE and OUTSIDE. By block 26,198,957 the contracts had sent IMD outside the team to three places. IdentityMD's fee address:Questionspays it 0.5 IMD each time it asks the oracle for an answer, three times by then. sIMD, IdentityMD's staking vault: it received Payroll's 14 drips to the stakers, about 0.0021 IMD in all, and the team's 1.000001 IMD thatEndowmentsstaked (an endowment's seed and a deposit of 1 IMD). And one wallet: the owner of the seat agent 51519 is bound to, paid its tip by the team's own call (the point on Payroll below). On 16 October a second wallet was paid the same way: the owner of seat 1978, which agent 52142 is bound to, was paid order 2's tip when the team's wallet H1 calledclaimTip. Each payment goes to the owner of the seat bound to the agent a work record names. Neither owner used HELP WANTED. - Two of the three answers were handed in by the team. The amount and the count of the first
report were handed in with
submitby the team's keeper wallet (the keeper is the bot that makes the calls anyone may make), and the contract checked the oracle's signature on each. The third, the weekly list of employers, can be answered by the callback only, and was. - Payroll has received two fees and two tips. Order 1's 0.05 IMD fee was split by the code
(half to the stakers' stream, of which a first part has reached sIMD, 30% to the seat pool, 20%
to the house feeds' meters); no seat owner has claimed a share of the seat pool yet. Its
0.05 IMD tip was credited to agent 51519. On 15 October 2026 the team's wallet W1 called
claimTipfor that agent, which anyone may do when the seat is held by an ordinary wallet, and Payroll paid the tip to the owner of seat 1049, the seat the agent is bound to (0x9edc8ba2…1fe8). Order 2's fee and tip, 0.0280025 IMD each, came on 16 October, and the fee was split the same way. The tip was credited to agent 52142. The team's wallet H1 calledclaimTip(52142), and Payroll paid it to the owner of seat 1978 (0xd7a75e9c…f273).EVIDENCE.mdhas every figure up to the block it pins. - One day of the report exists. Each feed's first request spent the IMD its meter held, and
the feed budget has put back 0.01 IMD since. A meter needs 0.55 IMD to ask again.
EVIDENCE.mdshows each meter's balance. - The code was reviewed before each launch by the team and by IdentityMD's agents, and audited
by nobody outside IdentityMD's network. Launch A was audited first by two audit jobs bought
from IdentityMD's network, and the team fixed or answered each finding. IdentityMD's launch
pipeline then read the code with those fixes (
audit_imported_code), changed it in answer to that review, and had four specialist agents and a judge read the code as adapted, all before it deployed the code on 13 October 2026. The launch's pull request says "An independent reviewer raised 8 finding(s).", and IdentityMD's record of the job holds those 8, none marked blocking.launches/a-payroll/ADAPTATION.mdanswers the four of the imported review and lists what the pipeline changed;launches/a-payroll/docs/AUDITS.mdlists every finding of the two audits and the pipeline's 8, with what became of each. One medium finding of the audits is acknowledged and left open (launches/a-payroll/docs/THREATS.md, P2). Launch B was read by the pipeline's audit agents and an independent reviewer before it was deployed, and no compiled code changed: the pipeline corrected three comments inCrowdPots.sol. Its pull request says 11 findings.launches/b-desk/ADAPTATION.mdanswers seven of them by name, and the other four through the judge's findings they were merged into, aslaunches/b-desk/docs/AUDITS.mdsets out; that file also names a file of answers the team never received. Launch C was read by the pipeline before its launch on 15 October 2026. Its pull request says "An independent reviewer raised 13 finding(s).", and IdentityMD's record of the job holds 13 in its accepted reviews, six low and seven info, none marked blocking.launches/c-company/ADAPTATION.mdanswers the three of the imported review. No node answered the other ten, which came after the code was adapted. They come down to four matters not raised before, each recorded by a specialist and again by the judge, and one finding of the imported review recorded twice more. The pipeline changed no file undersrc/.launches/c-company/docs/AUDITS.mdlists all 13 with what became of each; the four new matters (its rows 4 to 7) are limits thatlaunches/c-company/docs/THREATS.mdstates since 15 October 2026 (R11, E7, E9 and Z4). A test of launch C's suite shows each of the four, three of them in tests added on 15 October 2026. Nobody outside IdentityMD's network has audited any of the three launches, and IdentityMD records no outside audit or bounty for any of them. - One runner has taken two orders. The team's runner took orders 1 and 2, and no other runner
has claimed an order on the desk. Since 15 October 2026, 14:18 UTC, the team's runner has run around
the clock on a hosted server (Railway), signing as H1. It takes orders from any employer, one
order at a time, pays at most one 0.5 IMD swarm fee a day, posts bonds of at most 0.25 IMD, and
takes research-report and picture jobs (
bots/runner/README.md). Its list of employers (EMPLOYERS) is empty on purpose: the desk is for employers the team does not know, and while no other runner exists a list would leave their orders untaken. Those caps bound what serving anyone can cost it, for as long as it runs: at most 0.5 IMD in swarm fees a day, and at most 0.25 IMD in bonds at any one time, which it can lose order by order, plus gas. On 15 October 2026 the team also set it to take an order that leaves it no margin (MIN_MARGIN_IMD=0,EXPECTED_FAILURE_BPS=0), so that it could serve the first order that a rule of the team's, in launch C, opened on the desk. It did. On 16 October 2026, running unattended, it took order 2, paid IdentityMD's 0.5 IMD job fee and settled the order on its receipt. It was paid 0.504045 IMD and its bond back, about 0.004 IMD more than the fee, before gas. Two orders have been opened on the desk, both by the team. The keeper is not hosted: the team runs it by hand. The site is public (its address is at the top of this page); this page names no public address for the indexer or the bots.
Each launch's docs/THREATS.md lists every party the contracts trust and every known limit, and
docs/DEVIATIONS.md lists where we depart from IdentityMD's guidance. The ones a reader should
know before anything else:
- One key signs every oracle answer.
Questionsaccepts what IdentityMD's oracle signer signs, and that address is fixed at launch. A wrong or lost key can put wrong numbers into a feed and into the report. What an answer can move differs by launch. In launch A no function pays out on the content of an answer; the only money tied to an answer is its round's keeper tip (launches/a-payroll/docs/THREATS.md, T1). In launch B an order's condition may read an answer, and a true reading only opens the order to runners; the escrow still pays a runner only on the job's receipt or, where the order allows it, on a bonded claim that nobody challenged or that the order's arbiter upheld (launches/b-desk/docs/THREATS.md, T6 and W8). In launch C a FEED rule fires on a feed's answer, and the fire pays: the rule's recipient for a PAY rule, an escrow on the desk for a HIRE rule, and the rule's tip to whoever fires it (launches/c-company/docs/THREATS.md, T5, R2 and R3). - Two of the three feeds are facts the oracle recomputes from the chain. The third rests on a
panel's agreement and on nothing else
(
launches/a-payroll/docs/THREATS.md, F4). - A proof says what IdentityMD recorded, not that the work is good. Whoever controls the
WorkRegistry's writer can record a hash of its own making, and the registry's source is not
published (
launches/b-desk/docs/THREATS.md, T1, T2 and P1). - IMD's owner can have new IMD created on Ethereum. IMD here is a LayerZero bridged token, not a
proxy. Its owner,
0x047F…54B7, the same account as the WorkRegistry's admin, chooses which contracts on other chains may send IMD here and how their messages are checked, so it can have IMD created on Ethereum through a contract of its choosing. New IMD takes nothing from what the contracts hold, but changes what it is worth, and the token gives its owner no function that moves, burns or freezes IMD an address holds (launches/a-payroll/docs/THREATS.md, T4,launches/b-desk/docs/THREATS.md, T4, andlaunches/c-company/docs/THREATS.md, T1). - On the claimed path an untrue delivery claim that nobody challenges within 24 hours is paid, and while receipts are late a challenger can keep a true claim from being paid (the same file, L1 and L2). An employer chooses whether its order allows that path.
- Nothing moves by itself. Every call that keeps the system going is open to anyone, and
somebody has to make it. With no keeper, feeds go stale and nothing is lost, except on the
claimed path: there, an untrue delivery claim that nobody challenges within 24 hours is paid.
No watcher is running. On 15 October 2026 the team's keeper, the bot that can challenge such
a claim, is not hosted: the team runs it by hand and has no other watcher, and we know of nobody
else who watches the Doorbell's orders, which are all on the claimed path. At block 26,200,224
that day the Doorbell held no IMD, so it could open no order until somebody funds it
(
launches/b-desk/docs/THREATS.md, D2).
To report a problem privately, read SECURITY.md. There is no bounty.
The team publishes every wallet it controls in
evidence/house-wallets.json. Activity of those wallets, and of
HELP WANTED's own contracts, is HOUSE. Everything else is OUTSIDE. Every count we publish is split
that way, a zero is printed as a zero, and house activity is never presented as usage.
The rule is exact, event by event, in scripts/evidence/contracts.ts: each event is counted for
one account, and when that account made the call itself the count is HOUSE if the account or the
sender of the transaction is on the list. A transaction sent by IdentityMD's own service accounts
(its Intake's writer delivering an oracle answer, its WorkRegistry's writer) is counted apart from
both: it delivers something that was paid for, and is nobody's use of the contracts.
The site is also a remote MCP server, with no key and no account, at
https://web-production-a7c148.up.railway.app/mcp. Its tools read the desk, quote an order, plan
one and check any swarm job's receipt: desk_status, list_orders, get_order, list_templates,
quote_order, plan_order, plan_ring_doorbell, verify_job and receipts_status. None of them
signs or sends a transaction, and the server holds no key. An order is planned as unsigned
transactions, an approve and then createOrder, which the user checks and signs with their own
wallet.
# Claude Code
claude mcp add --transport http help-wanted https://web-production-a7c148.up.railway.app/mcp
# Codex
codex mcp add help-wanted --url https://web-production-a7c148.up.railway.app/mcpThe skill an agent reads first, with the safe flow for an order and the limits, is
apps/web/public/skill.md, served at
https://web-production-a7c148.up.railway.app/skill.md; a shorter orientation is
apps/web/public/agents.md. Without MCP, the same data is plain HTTP:
the JSON API under /api/v1 (described in /openapi.json; /api/v1/flows plans every
transaction the site offers), every page as .json and .md, and /llms.txt to list them all.
The server is apps/web/src/lib/mcp.ts, tested offline in apps/web/test/mcp.test.ts.
launches/a-payroll/ Launch A, live: Questions, JobsReport, Payroll. A Foundry project.
launches/b-desk/ Launch B, live: WorkProof, WorkOrders, CrowdPots, Doorbell.
launches/c-company/ Launch C, live: Rules, Endowments, Zap.
packages/shared/ TypeScript the rest imports: addresses, the deployments book, ABIs generated
from the launch projects' build output, the order header, the work record
scanner, the oracle's question hash, a client for IdentityMD's public API.
bots/keeper/ The bot that makes the calls anyone may make: ask a feed, hand an answer in,
print the report, return an expired escrow, watch delivery claims.
bots/runner/ The bot that takes an order, buys the swarm job and collects.
apps/indexer/ Copies chain events and public API data into a database; a read-only API.
apps/web/ The site: the receipt checker, the report, the hire desk's pages. Every page
also answers as JSON and as Markdown, and /mcp is its MCP server for agents.
scripts/evidence/ The evidence builder (pnpm evidence) and its tests.
scripts/repro.sh The comparison of deployed code with a build of this source (pnpm repro).
scripts/first-rule.sh How rule 1 of launch C was created and fired on 16 October 2026 (cast, with
the wallet's own keystore).
scripts/proof-requests.sh
How the three questions of the report were first bought from the live oracle.
scripts/feedback.sh Our ERC-8004 reviews of the agents that did orders 1 and 2, given by H1, the
wallet that paid for their jobs, each naming a feedback file on the site
(apps/web/public/erc8004/) with its x402 payment (cast, with the keystore).
evidence/ The house list, the list of launches requested and not yet in the book
(empty since 15 October 2026), and evidence.json (generated).
EVIDENCE.md Generated. What exists on chain, at one pinned block.
SECURITY.md How to report a problem privately, what is in scope, and that there is no
bounty.
Start with the README of the part you care about:
launches/a-payroll,
launches/b-desk,
launches/c-company,
bots/keeper, bots/runner,
apps/indexer, packages/shared.
How the site looks, moves and speaks, with the files to draw it: apps/web/BRAND.md.
The contracts. Foundry is the only requirement, dependencies are vendored, and no network is used:
(cd launches/a-payroll && forge test)
(cd launches/b-desk && forge test)
(cd launches/c-company && forge test)On 15 October 2026, with Foundry 1.8.3: 515 tests passed in launch A, 673 in launch B and 447 in
launch C, none failed. A first build takes a few minutes. Each project also has a fork suite that
runs against mainnet (FOUNDRY_PROFILE=fork forge test); its README says what that suite covers.
Everything in TypeScript. Node 22 or newer and pnpm 11:
pnpm install
pnpm test # every package, then the evidence builder
pnpm typecheckOn 20 October 2026, in a fresh clone: packages/shared 397 tests (22 of them skipped), bots/runner 431,
bots/keeper 326, apps/indexer 213, apps/web 769 and the evidence builder 100, none failed.
GitHub Actions runs these suites and forge test in each launch on every push to main (the CI
badge at the top). Those are the counts
each suite reports as Tests N passed (N). By each package's own account no test reaches the
internet; the tests of the indexer, the site and the evidence builder make fetch throw to be
sure. Of the keeper's 326, 296 are unit tests (test:unit) and 30 start a local node; those 30
are skipped, with a message, when Foundry is missing or the launch projects are not built, and the
suite then reports them as skipped.
One part at a time: pnpm --filter @help-wanted/keeper test (and shared, runner, indexer,
web), or pnpm evidence:test.
When a launch is mined, its transaction, its block and its addresses are written into
DEPLOYMENT_BOOK in packages/shared/src/deployments.ts. That is the only change needed: the
bots, the indexer, the site, pnpm evidence and pnpm repro all read the book. Then:
pnpm repro # the deployed code is this source
pnpm evidence # EVIDENCE.md and evidence/evidence.json again
pnpm test # among others: this page's status table must agree with the bookand the row of that launch in the table above is brought up to date by hand.
MIT. See LICENSE.