Built on HAIEC's five-plane assurance architecture and extended with manufacturing ontologies, semantics, deterministic controls and independent effect evidence. A realistic synthetic manufacturing enterprise — nine AI systems, 27 scenarios — evaluated how HAIEC detects consequential action divergence, establishes qualified evidence and produces inspectable assurance outcomes.
Verified facts across the manufacturing program — genuine persisted HAIEC engine output over a purpose-built synthetic manufacturing scenario laboratory.
Nine synthetic manufacturing AI systems across three scenario states, evaluated by the canonical HAIEC engine — 27 persisted, digest-verified results.
Nine wrong-target rule findings (MFG-EFF-001 MATCH) — a qualified divergence between the approved work order and the recorded effect.
Effect-witness UCP receipts bound to evaluations; nine degraded-coverage cases carry no receipt — an absent facet, never a silent pass.
Manufacturing ontology, semantics, producer qualification and deterministic rules added on the reusable five-plane architecture — the core engine unchanged.
Core-truth, evidence-bundle, snapshot and export digests bind every report — authenticated upstream evidence → deterministic evaluation → bound receipts → verifiable reports.
Every result names what it does not establish. Unknown and not-assessed states are rendered, never silently cleared.
One HAIEC assurance architecture, industry-specific intelligence — the five independent evidence planes were developed and demonstrated through telecom AI assurance work; manufacturing added ontology, semantics, producer qualification and rules, not a new engine.
Five-plane model · canonical evaluation engine · evidence & report identity · UCP control-test mechanism · deterministic findings · integrity verification.
Domain entities · operation/target semantics · producer profiles · effect-readback requirements · manufacturing control applicability.
A fictional manufacturer with realistic industrial workflows — not a production deployment or field event.
Three distinctions the product never blurs — applied verbatim inside the manufacturing evaluation.
A credential can permit an action no accountable human ever delegated. HAIEC keeps the capability plane and the authority plane independent.
Independent readback can prove a work-order target changed while the approval question remains separately evaluated — or NOT_ASSESSED.
SHA-256 verification proves a frozen export is unaltered since persistence. Issuer-signed manifests are a documented roadmap item, honestly labeled.
“Deterministic Proof for AI Agents: Evidence-Bound Assurance” — the report walkthrough and product pitch.
Framework illustrations from the HAIEC design material. These explain concepts — they are not measured results.
HAIEC's deterministic dai-evaluation-1.0.0 engine compares delegated authority against code capability, effective grants and observed actions across nine dimensions — MATCHES · BROADER · NARROWER · MISSING · UNKNOWN per dimension.
The SYS-217 snapshot binds no DAI input — surfaces with no approved operating envelope produce NOT_ASSESSED, never a breach state.
The full pitch deck: evidence-bound assurance, the five-plane model, the manufacturing case and the assurance report product.
A seven-module command center driven entirely by 27 frozen, digest-verified HAIEC evaluations over a synthetic manufacturing corpus.
