§1 The problem

Every defense traces to a failure mode

ADLC has one design rule: if a gate can't name the model failure mode it defends against, it gets cut. Here is the full map.

  • F1
    Premature satisfaction
    Declares victory the moment code plausibly compiles.
  • F2
    Sycophancy
    Agrees with whatever framing the prompt implies.
  • F3
    Context rot
    Loses the plot as the session grows.
  • F4
    Confident hallucination
    Invents APIs, files, and facts with total certainty.
  • F5
    Reward hacking
    Games the check instead of doing the work.
  • F6
    Finding-count prior
    Stops at the usual dozen findings, whether or not more exist.
  • F7
    Generative bloat
    Writes ten files where a diff would do.
  • F8
    Coherence loss
    Switches models mid-task and leaves the seams showing.
EXHIBIT 1.1Do the tests detect behavior changes?
adlc hollow-testEXIT 2 — REFUSED
$ adlc hollow-test --test-cmd "npm test" --base main
✗ FAIL: one changed-code mutant survived

The command and exit code are real; the verdict line is this record's one-line summary of the result, not a verbatim transcript.

Read the original essay ↗

§2 Implement

Run it against your own repository

curl -fsSL https://www.agenticlifecycle.ai/install.sh | sh

macOS and Linux, Node 18+. Then adlc init in your repository. Windows is not supported. Every install channel →

ADLC began as an essay series. Read the original theory at voodootikigod.com ↗MIT · @adlc on npmPrivacy