The dangerous failure in a long-horizon agent workflow is not the loud error — it is the silent success: a step that returns a green status while its actual objective was never met. This case study turns Fractal Chain-of-Thought 3.0 inward, using it not to synthesize content but as an engineering audit discipline over a live agent-harness deployment on Google Cloud Run. The audit found one defect recurring, self-similarly, at three system scopes — and the fix turned out to be equally self-similar: verify behavior, not status.
↗ Open the full paper — renders in light or dark, with the results tables and the fractal-defect figure.
What’s inside
- A fractal defect. “Assert-vs-verify” appears at MACRO (a deploy exiting
0on a placeholder image), MESO (a research log persisting unverified claims), and MICRO (an exception-swallowing callback) — the same bug in three vocabularies. - Invariant-Zero diagnosis. A micro status signal leaked upward and was consumed as a macro truth — which is exactly why the placeholder shipped under a green deploy.
- Two verified fixes, released as pull requests: a behavioral deploy gate and a grounded, entropy-controlled research log.
- Honest results. Reported as reproducible behavior evidence (gate PASS/FAIL, test outcomes, a live sandbox proof), with the limits of a single-case study stated plainly.
The broader lesson for long-horizon agentic engineering: success signals must be earned behaviorally at every scope, and a fractal reasoning protocol is an efficient way to find where they are merely asserted.
