Chapter 19 - What We Can and Cannot Yet Claim
Loop Engineering by Fabio Correa, the Applied AI foundation volume on working with AI through framing, feedback, verification, memory, and human judgment.
Last updated 2026-10-25. Chapter 19 weighs the book's own claims against their evidence and names the open questions the field has not solved. This page keeps the register: each claim group with the chapter's verdict, its current status, and the date it was last checked. The arguments stay in the book; what changes as evidence arrives is recorded here.
The question this chapter takes up
Which of the framework's claims can the evidence carry today, which are optimistic, and which cannot yet be verified?
Put it to work
Before relying on a claim from the book, check it against the chapter's verdicts: where the framework is firmest, where it is optimistic, and what should be tested next. If you find evidence that moves a verdict, send it in.
Claim-status register
No verdict has changed since publication. Evidence that changes a claim's status can be sent to errata@correax.com; corrections to the second edition are collected on the Errata page.
| Claim | The chapter's verdict | Status | Last checked |
|---|---|---|---|
| The convergence claim: seven primitives recur across three public practices and one personal practice (Chapters 1 and 3) | Firmest ground | Unchanged since publication | 2026-10-23 |
| The materiality dial applies cleanly in practice (Chapter 13) | Optimistic: medium-stakes classification is where errors concentrate | Unchanged since publication | 2026-10-23 |
| Calibrated reliance (Chapter 16) | Directionally well grounded, instrumentally underdeveloped | Unchanged since publication | 2026-10-23 |
| Reliance calibration metrics | Open gap, as of October 2026: no validated composite measure | Open | 2026-10-23 |
| Autonomy-budget calibration | Open gap, as of October 2026: formal frameworks remain weak | Open | 2026-10-23 |
| Cross-instance error propagation rates | Open gap, as of October 2026: no cross-system metric | Open | 2026-10-23 |
| Faithfulness of visible reasoning traces | Open problem, as of October 2026: no field-grade instruments yet show when a trace reflects actual computation | Open | 2026-10-25 |
| Verifier-against-world checks | Open problem, as of October 2026: side effects that never return telemetry stay outside the loop's native observability | Open | 2026-10-25 |
| Moral patiency of AI systems | Open problem: the book remains agnostic; human accountability is assigned now, ahead of the evidence | Open | 2026-10-25 |
What should be tested next
The chapter specifies three evaluations. None has been published yet; when one appears, the register rows it affects change here.
- A prospective evaluation comparing the seven-primitive architecture with a simpler baseline on matched tasks and equal budgets.
- A materiality classification test: independent operators classifying the same tasks, their verdicts checked against field outcomes.
- Fleet-learning uptake instrumentation: a lesson counts as learning only when a recipient retrieves it, applies it, and improves a defined outcome.
Key sources
The chapter's verdicts rest on three decades of human-factors research on trust in automation:
- Bainbridge, L. (1983). Ironies of automation. doi.org/10.1016/0005-1098(83)90046-8
- Lee, J. D., and See, K. A. (2004). Trust in automation: Designing for appropriate reliance. doi.org/10.1518/hfes.46.1.50_30392
- Parasuraman, R., and Manzey, D. H. (2010). Complacency and bias in human use of automation. doi.org/10.1177/0018720810376055