AI Founder Hub

Review before reach

A public standard for work you can inspect.

Founder Hub separates a compelling idea from a reviewed operating skill. Maturity follows evidence, not presentation quality.

Current registry state

Every skill record is Planned.

Planned records expose a job, interface, limits, and lineage so critique can begin. They are not runnable artifacts, recommendations, or readiness claims.

Review criteria

The questions every kit must answer.

A strong result cannot compensate for missing authority, unverifiable evidence, or an interface that fails outside a happy path.

01

Generality

Useful beyond the venture for which it was first written.

02

Portability

Another founder's agent can run it without the author in the room.

03

Repairability

Legible, testable, and maintainable by someone other than its author.

Lifecycle

Maturity is earned in visible stages.

  1. 01

    Planned specification

    Publish the bounded job, proposed contract, evidence lineage, and unresolved dependencies without implying a runnable artifact.

  2. 02

    Architecture ratified

    Freeze the shared Stoa, RUN_RECORD, promotion, and write-ownership contracts before candidate work begins.

  3. 03

    Candidate built

    Implement one bounded capability against pinned inputs, frozen fixtures, and explicit human gates.

  4. 04

    Adversarially tested

    Exercise ordinary, boundary, negative, composition, and deliberately broken cases with executable checks.

  5. 05

    Independently graded

    A different-vendor reviewer grades the exact frozen bytes; any subsequent edit voids that verdict.

  6. 06

    Published

    Release only the accepted version, preserve its evidence packet, and reopen review for every repair.

Evidence packet

What a review needs to see.

Contract

Versioned purpose, users, inputs, outputs, gates, permissions, failure modes, and stop conditions.

Lineage + RUN_RECORD

Shared venture state with sources, claims, calculations, decisions, actions, exceptions, and retained artifacts attached.

Executable adversarial suite

Ordinary, boundary, negative, non-transfer, and deliberately broken checker-teeth fixtures with recorded exits.

Verdict + repair record

A different-vendor verdict over exact hashes, named limitations, change notes, and the route back through review.

Review posture

Signals that stop a record from advancing.

Authority by implication

The agent can act beyond the named job, tool boundary, budget, or human approval.

Evidence by confidence

Claims cannot be traced to sources, observations, or a declared inference.

Success by anecdote

A polished run substitutes for a test set, rubric, known failures, or comparison.

Lineage without license

Source material, data, or adaptations cannot be used and credited on the stated terms.

Self-certification

The producer or integration owner is presented as the independent grader of the same candidate bytes.

Verdict after byte drift

A candidate changes after its hash is frozen but retains the prior review result or publication status.