AMCouncil · Human–AI Boundary Framework · Foundational Research (Independent)
A theory-forward research pack built from first principles and primary sources — a rigorous, shared floor for members and volunteers to form their own view. Knowledge-gathering, not a webinar design and not a settled position.
AI amplifies whatever you already have — it can make an inferred condition look measured, an optimised proxy look like real performance, and a quantified risk look safe. Which way it resolves is a governance choice, not a property of the model. The irreducibly human core that recurred across every stream: define the purpose · validate the evidence · own the value judgement · accept & record the residual risk · retain the ability to stop · stay accountable by name · watch for drift · keep the skill to work without it.
The whole argument in one read: the triad re-grounded, what the classics say is irreducibly human, the boundary science, the standards gap, the case lesson — plus a provisional per-concept boundary table, a draft decision-rights sketch, and nine open questions for the volunteers.
Condition is a latent state — never observed, only inferred from proxies. Why that fact forces a human to validate the inference.
Performance is relational (function, standard, context) — and an AI optimiser is a Goodhart machine that maximises the proxy, not the purpose.
Risk is the effect of uncertainty on objectives. Knight's risk-vs-uncertainty line is exactly where AI's false confidence begins.
Hume, Aristotle, Polanyi, Wang Yangming, Zhuangzi and Sun Tzu converge — millennia apart — on judgement, tacit knowledge and knowing-as-acting as the human core.
Seventy years of human-factors research — Ironies of Automation, calibrated trust, automation bias, the responsibility gap — distilled into eight non-delegable human duties.
The accountability scaffolding already exists (ISO 55001, ISO 31000) — but no asset-management standard yet governs AI oversight. That gap is what AMC can fill.
Cross-sector cases (aviation, rail, grid, water, oil & gas, mining, structures) where prediction of condition, performance or risk succeeded, failed or nearly failed.
Contested claims and vendor/operator-reported figures (PdM market sizes, platform accuracy %, savings %) are flagged as claims throughout, and headline quotes were verified against primary sources. One error found & fixed: Nowlan & Heap's age-related-failure share is ~11% (≈89% not), not 9%. EU AI Act high-risk obligations (incl. Art. 14 human oversight) apply from 2 August 2026. Foundational research for member validation — not the AMCouncil's settled position. (A superseded open-source-tooling dossier is parked in Archive/ and is not part of this pack.)