Research · Verification · ToolsApplied AI Studio

Test the claim.
Build what works.

We test public AI and ML systems on real hardware, publish inspectable research, and turn verified methods into tools people can use.

No human-approved research release exists yet.

CAPSULE / TINY-MLP-SEED-SENSITIVITY-V1 LOCAL
  1. 01
    Protocolretrospective canary snapshot
  2. 02
    Execution12 seeds × 2 local trials
  3. 03
    Artifact checklocal self-asserted agent
  4. 04
    Human approvalnot received
bundle b49c7f93…14f9017 Risk tier 0
$495

founding offerone public target

48h

target turnaroundafter fit and access

02

executed research slicessynthetic, retained, draft

00

approved releaseshuman gate remains open

What we doEvidence into utility

Three ways to turn
claims into value.

Every service, experiment, and tool begins with the same discipline: define the claim, run a bounded test, preserve the evidence, and state exactly what the result does and does not prove.

Release disciplineEvidence before polish

A release is a sequence,
not a button.

  1. 01Freeze

    Question, protocol, seeds, limits, risk, and abort conditions.

  2. 02Execute

    Bounded runs with raw outputs, logs, environment, and hashes.

  3. 03Review

    A distinct reviewer checks the artifact and scopes every label.

  4. 04Approve

    One accountable human decides whether public release is warranted.

  5. 05Archive

    An immutable release makes the exact revision reachable.

Current boundary2026-07-21

The foundation is real.
The release is not.

The local harness, two retained bundles, evidence-derived Quarto report, trusted-source container canary, verified RO-Crates, five public procedures, machine-readable feed, and two browser labs exist. Accountable human approval, external reproduction, external learner observation, and an immutable public archive do not.

Read the exact labels