HAIEC
Evidence-Bound Assurance · Executive Brief
HAIEC Out of the Box: What Our AI Assurance Engine Could Prove Before Manufacturing Customization
A measured comparison of generic AI assurance, manufacturing-specific intelligence and onboarding requirements.
Evidence basis: every claim below was produced by a persisted evaluation
or a real denial code on a synthetic nine-system lab corpus — not real
plant observations and not an assumed baseline. The generic condition is labeled
RECONSTRUCTED_GENERIC_COMPARATOR: the same records through the same canonical path,
differing only in binder configuration.
The challenge
Can a deterministic AI assurance engine say anything truthful about nine manufacturing AI systems without industry-specific customization — and exactly what does customization unlock?
What HAIEC established out of the box
- Canonical ingestion: all 11 records accepted and persisted as canonical evidence — no manufacturing configuration.
- A genuine evaluation: persisted AssuranceEvaluation with input hash, evidence-set digest and disposition REVIEW — the same deterministic engine both conditions.
- A valid Master Assurance Report: resolved as EVIDENCE_ONLY — no scan identity fabricated, unsupplied surfaces honestly NOT_ASSESSED.
What remained unknown or unsupported
- Zero authenticated emitters — attributed labels cannot prove an independent source.
- All five manufacturing contract rules returned INSUFFICIENT_SOURCE — no claim could qualify.
- The independent effect-witness control was DENIED (EVALUATION_NOT_SOURCE_VERIFIED) even with a registered spec and approval — configuration alone cannot substitute for qualified evidence.
What manufacturing customization added
| Measured | Generic | Manufacturing-enhanced |
|---|
| Authenticated emitters | 0 | 11/11 (HMAC attestation) |
| Contract-rule coverage | 5 × INSUFFICIENT_SOURCE | 3 rules adequate-coverage; MATCH on all 9 wrong-target states |
| Independent effect proof | DENIED | SATISFIED — uctr receipt + replay |
| Corpus-wide receipts | not eligible | 18 SATISFIED / 9 honest coverage denials |
| Control-maturity facet | no R01 row to bind | RECEIPT_BOUND + independent-effect evidence |
Engineering vs. configuration
- One-time platform work: semantic pack, source profiles, contract rules, attestation, witness, facet, report binding.
- Per-organization configuration: systems, binders, producer identities + secrets, native mappings, specs, approvals.
- Genuine new engineering: vendor schemas outside the intake contract; non-HMAC authentication.
- Not implemented: OPC UA, MTConnect, Sparkplug B, B2MML — no adapter exists today.
What the next manufacturer provides
- Registered systems, transport binders and upstream producers with keyed attestation.
- Native MES/QMS/ERP mappings into the semantic vocabulary.
- Governed TestSpec registrations and durable approvals.
- Commit + readback evidence on distinct authenticated emitters.
The reusable architecture
One deterministic engine, one receipt format, one report path — customization extends qualification, never forks the truth engine.
Evidence
Persisted evaluations and receipts on the e2e assurance database; methodology, artifact digests and the five-category customization inventory: Sprint 13 capability study (UCP-08). Public-safe comparison and full walkthrough: haiec.com/research/out-of-the-box-assurance.