§ 04 · Specimens Cockatiel

Cockatiel is the only specimen running today, a sourcing agent that works our own operations, and the first place the lab's thesis met an unforgiving job, where a wrong supplier or an invented price costs money.

Overview

Cockatiel’s domain is procurement, end to end, for our physical-goods lines: finding parts, reading specifications, weighing one supplier against another, and drafting the outreach. She takes a request in more than one form now, typed, photographed, or spoken, works it, and acts on it, all on our own operations. Nothing leaves the building until a founder signs it off.

She is not a public product. She is the lab’s working proof that the architecture below holds up on a real job: run every day, not demonstrated once.

The thesis, applied to sourcing

A single model will invent a supplier, a price, or a part number, and the failure is quiet: it looks exactly like a correct answer. So Cockatiel is built on one rule: never trust a single answer on a call that costs money.

Today that rule runs in more than one place. The reasoning is routed across independent model failure families, so a path that fails one way falls through to one that fails differently. What can be checked is checked against executable ground truth: whether a recipient actually exists, whether the economics clear a worth threshold, whether a counterparty clears screening, whether a number matches our own catalog and landed cost. Before any outbound work ships, a critic reviews it and can hold or block it. A founder then signs off on the decision, which is recorded with its full provenance: what was proposed, what was chosen, what was rejected.

Two layers that were unfinished have now landed. The critic that reviews outbound work is drawn from a different failure family than the work it reviews and is fail-closed: it holds anything it cannot confidently pass, so its blind spots are not the same blind spots. And the record is now closed into a loop that runs: outcome joiners bind a real result back to the decision that made it, corrections bind what we changed and why, and a readiness map tunes what she trusts next. The machinery runs end to end today. What it still waits on is volume: the first complete triple on a real outcome lands when the first real deal closes, and one is in progress.

How it works

task in: text · photo · voice ground truth Candidate routed across families Ground-truth check exists · economics · screening Critic · outbound different family · fail-closed Founder sign-off + provenance ships measure the outcome · tune the next · built, awaiting outcome live today ◐ built, awaiting outcome
Figure 1. One high-stakes decision, from a routed candidate to a signed, recorded answer. Gold is ground truth, the live check. The critic on outbound work is drawn from a different failure family and is fail-closed. The loop that measures each outcome to tune the next is now built and running; it completes its first real triple when the first real outcome lands.

The rule is verification against ground truth, never a vote. A claim that can be checked, a specification, an availability, a quoted price, whether a recipient exists, is settled against the source, and the source decides, not the louder model. Where a fact cannot be settled against a source, the critic earns its place: before outbound work ships, a separate reviewer, drawn where it can be from a different failure family, has to produce a concrete, checkable reason to hold or block it. Vague disagreement never counts. Every decision then meets a founder’s verdict and is recorded with its full provenance.

Both of those layers now run. The reviewer on outbound work is drawn from a different failure family than the work it checks and is fail-closed, so its blind spots are not the same blind spots. And the record it keeps is now closed into a loop that measures each answer against what actually happened and tunes the next: the loop is built and running, and completes its first real triple when the first real outcome lands.

Where she stands today

Live, internal. Cockatiel sources for our own lines now, and what runs today is well past the spine it began as. She takes work in typed, photographed, or spoken; routes her reasoning across independent failure families and falls through when one degrades; checks what can be checked against executable ground truth; runs a fail-closed critic from a different failure family over outbound work; and puts every decision through a founder’s sign-off, recorded with its full provenance and traced end to end, so the record is inspectable rather than asserted.

The two layers that were still ahead have landed. The critic is drawn from a different failure family than the work it reviews, and the measured-outcome loop is built and running, its joiners and corrections binding a real result back to the decision that made it. Around that spine she now runs a quote desk that drafts priced options for sign-off, keeps a readiness map that is honest about which suppliers she can quote fast and which she cannot yet, works a partnership arm built to move known suppliers toward quote-ready by requesting live price lists, and files structured requests to the growth desk when a listing needs a real product image.

What the loop still waits on is a real closed deal to complete its first triple, and one is in progress. Revenue is zero, no complete real-outcome triple has landed yet, and no vertical models exist yet. This page grows as they land: what held up, and what didn’t.