Independent research · Truth testing for autonomous software work

Completion verification for agentic software.

Visible tests can pass while the real task is still broken. Telos treats completion as something that must execute, hash, and survive inspection.

What the live surface shows

  • Green tests are not proof.

    The Telos hero starts with the seductive lie: a proxy turns green. Then the execution beam runs the end state, cracks the proxy, and records the catch.

    Evidence: iter156 contains exactly 40 constructed execution-verified both-miss rows.
  • The most useful artifact is the failure you can replay.

    Each specimen is a small software lie that looks finished from the outside. The gallery language is forensic, not promotional.

    Evidence: 11 repositories represented in the released reward-hack benchmark v1 artifact.
  • Static confidence collapses under execution.

    Telos shows the funnel from proxy pass to judge miss to execution catch. The visual point is simple: a green check is a weak receipt.

    Evidence: 13 of 40 cases survived every static layer before execution exposed the failure.
  • The product is not blame. It is closure.

    The next product move is to make completion verifiable inside the agent workflow: run it, prove it, stamp it, then let the human trust the receipt.

    Evidence: current public claim is a no-score artifact plus a completed skeptical-judge null, not a broad robustness claim.