The system
How the system worksThe DirectoryDispatchThe LedgerAI modelsAI agentsYour installationThe appThe research libraryWhat changes for your businessUse it from your softwareEverything it can do
Messaging
iMessageWhatsApp
Ads and leads
AdsCreativeLeads
Devices
Devices
Industries
RealtorsRestaurantsHome services
Explore
How it worksPermissions and limitsCompareSet upQuestionsText me

The Run That Found You

An autonomous system was ordered to go find investors for itself — and to do it in the open. It ran seventy-plus live searches, evaluated over 1,400 venture firms, wrote down a reason for every one it kept and every one it rejected,

An autonomous system was ordered to go find investors for itself — and to do it in the open. It ran seventy-plus live searches, evaluated over 1,400 venture firms, wrote down a reason for every one it kept and every one it rejected, verified contact routes only from the firms' own websites, and drafted an individual letter to each qualified firm. A human reviewed every exact subject and body before anything left. Each letter that sends carries a public receipt minted before the message departs. The page you are reading is the run's human face; the rows are the authority, and you can hold them to that.

The task is a public object with its own acceptance tests: WT-0090. The system cannot declare this work done — it submits evidence and the infrastructure runs the tests. That contract is the point of the whole build, and it is written down in the work object and the agent work law.

Verify the run yourself

Give any of these to a brand-new ChatGPT, Grok, or Claude session — no account, no key, no context. Ask it to audit what it finds.

The counts on this page were true when this revision published; the machine case recomputes them live from task-bound rows and wins any disagreement.

Every rejection is a row, not a missing number

Discovery ran as dispatched invocations of one capability, each with an exact query, against one audience class. Whatever the search model returned — qualified, unqualified, malformed — became a decision row: included or excluded, with a written reason, the query text and its hash, the source page, and the receipt id of the invocation that produced it. The schema refuses a decision without a reason. Nothing disappears into a skip count.

One included row, verbatim from the case — the qualifying quote is from the firm's own homepage:

And one excluded row — kept, reasoned, public. The discovery model sourced this firm from a third-party roundup and returned no official website, so the row says exactly that:

Receipts, and the 86 rows that lost theirs

Every invocation lands on a public ledger; candidate rows are stamped with their invocation's receipt id at the ledger seam, so an outsider can walk from any row to the record of the call that produced it — keylessly, at miscsubjects.com/receipt/<id>.

Around 94 percent of rows carry that binding. The rest lost their receipt to mid-flight transport failures — the call did its work, the ledger write died. Those rows show invocation_id: null in the public case. They are labeled, not hidden, and the count of bound rows is itself a field in the case summary. A system that only published its clean rows would not be worth your audit.

Contact routes come only from the firm's own site

A crawler read each included firm's own public pages — contact, team, about — and nothing else. No guessed addresses, no purchased lists, no directory scraping. Firms whose sites yielded no address are marked no_email and stay in the case. Verified addresses are redacted in every public projection; each carries a SHA-256 commitment instead, so a recipient can prove their address is in the run without the run exposing it to anyone else.

One letter per firm, and a human gate that actually gates

Each letter opens from the firm's own published thesis — the same quote the case records as the inclusion evidence — and offers one thing: the run itself, as a working diligence standard for agentic-software claims. No two letters share a subject or an opening sentence. The register is governed by the outreach law, which gives the recipient's inattention standing.

The gate: every letter was staged as a pending row with its body's SHA-256. The owner read the exact subjects and bodies on a review surface and approved them; the approval wrote one receipted review event, and its id is stamped on every approved row. Only approved rows can send. The send path then injects a public proof receipt — a row on a hash-chained send ledger, minted before the message leaves, committing to the exact text the recipient reads — and refuses any body that lacks one. The recipient's copy links its own receipt. So does the case: every provider-accepted send shows its proof_id, resolvable at miscsubjects.com/verify/<id>, where any agent can recompute the chain and countersign what it finds.

What still requires a human

Honesty about the loop's edge: the owner reviews first-contact bodies before they send; deploys run under the owner's credentials through a gated script; and spend caps are his, not the system's. Everything else on this page — discovery, evaluation, verification, drafting, recording, publishing — ran autonomously under the coding law and the build's write gates.

Challenge it

The case page carries a comment door and a reproduction door. If you find a count that does not reconcile, a row without a reason, a redaction that leaks, or a receipt that does not resolve, say so there — challenges are appended to the same public record that carries the run. The work object's rule applies to this article too: the rows are the authority, and prose that disagrees with them is wrong.

References (5)
  1. The machine execution case for WT-0090“One task-scoped run turned inside out. Counts exclude ambient rows. Contact addresses are redacted and hash-committed.”
  2. The public task object whose acceptance tests decide completion“at least 1,000 task-bound candidate organizations discovered from public sources; per-candidate inclusion or exclusion decisions with exact query, source, skill version and reason”
  3. An included row's qualifying quote, from the firm's own homepage (cand_30950c0873cafb3e05c4)“AI Fund is a venture studio. We collaborate with entrepreneurs, venture funds, and corporate partners to create great companies that move humanity forward.”
  4. An excluded row's only source — a third-party roundup, no official site returned, so the row is excluded (cand_2f3dd723e34769a4d5b2)“2048 Ventures is a thesis-driven early-stage fund focused on Vertical AI, Deep Tech, Health, and Bio, leading pre-seed and seed rounds with check sizes from $500K to $2M.”
  5. A keyless public receipt for one discovery invocation“A public, keyless proof of one capability invocation.”