Arc · 6 posts

No Model Reviews Its Own Homework

The factory's adversarial-review practice didn't start as doctrine, it became doctrine because the same failure kept recurring: a fix, itself unreviewed by a second perspective, quietly introduced the next bug. This arc is that doctrine's origin story plus four demonstrations that it holds outside its original context, closing on the sharpest evidence available: two real, live-exploit-confirmed bugs a second reviewer caught that the first pass missed.

Reading order

  1. 1
    No Model Reviews Its Own Homework

    agentic-sdlc-patterns

    Four documented review cycles, one recurring shape: a fix, left unreviewed by a second perspective, quietly introduces the next bug. The sentence the factory arrived at after proving it the hard way, four times.

  2. 2
    The Fixes That Held Changed a Rule

    agentic-sdlc-patterns

    Three adversarial review rounds on the same new codebase, and the diagnostic that separated a durable fix from a fragile one: did it change a rule, or just a line?

  3. 3
    Doubt Before the Code Exists

    agentic-sdlc-patterns

    Adversarial review, applied at intent-declaration time instead of code-review time, caught real defects before a single commit existed to review, proving the doctrine is a decision-making technique, not just a code technique.

  4. 4
    It Wasn't a Matched-Cast Comparison

    agentic-sdlc-patterns

    A live-model coordination comparison looked decisive: peer-delegation succeeding, central-orchestrator starving the colony. An independent review asked whether it was actually a fair fight. It wasn't.

  5. 5
    Shipping With a Watch, Not a Verdict

    agentic-sdlc-patterns

    Five doubt cycles on one small registry guard, and the finding rate never reached zero. What actually caught the headline bugs was mutation testing and fresh-context review, self-inspection caught none of them.

  6. 6
    Two Reviewers, One Exploit, Two Real Bugs

    agentic-sdlc-patterns

    Two AI reviewers, independently, with real build and openssl access, found a bypassable revocation path and a spoofable certificate identity field in device-CA code. Both were confirmed with a live exploit, before and after the fix.

Interactive · Arc-level · No Model Reviews Its Own Homework

Six real incidents, six different products and people. Every one collapses to the same shape: a first pass believed something was settled, and a differently-angled second pass, aimed at exactly what the first pass just changed, caught what it structurally could not see. Pick an incident.

First pass

A device-certificate authority's enrollment path shipped without checking a device's current status before issuing it a fresh certificate.

Second pass caught

Two independent reviewers, given real build, Docker, and OpenSSL access and told to attack rather than read, separately reproduced a live exploit: replaying an already-processed enrollment request against a just-revoked device minted it a brand-new, fully valid certificate.

Why the first pass missed it

A review that never runs the code it's reviewing is trusting the code to describe itself honestly. This class was very plausibly invisible to static reading alone.

Proof

The fix was verified by revoking a real device, then replaying the identical exploit again: confirmed rejected with the correct error, and confirmed the device's status stayed revoked, not silently reset by the attempt.

Read the full post →