Explainable Compliance: Citing O.Reg 347 By Section, Not By Vibe

By Michael Leslie Atkinson · Founder, 905WOOD.COM SALES · May 1, 2026 · 9 min read

---

The problem with most AI compliance products

Most AI-powered compliance tools deliver a confidence score and a verdict. The vendor markets the model accuracy. The buyer trusts the score. Then a regulator asks for the basis of the verdict, and the vendor cannot reproduce it because the model has updated, the training data has shifted, and the regulatory text has been amended since the verdict was issued.

This is not a theoretical risk. It is the dominant audit-failure mode for AI deployments in regulated industries since 2023. The pattern is now well-documented: confidence-score AI works for marketing and operational efficiency; it fails for regulatory defensibility because there is no chain of evidence.

The 905WOOD architecture — every verdict shows its work

The Digital Refinery is engineered around a different premise: every classification verdict is reproducible, citable, and auditable. The architecture has five components, each picked specifically for the audit chain rather than for raw model performance.

1. Google Gemma 4 26B MoE — the regulatory reasoning brain

Apache 2.0 licensed, 26-billion-parameter mixture-of-experts model with 256K context window and constrained decoding. Configured with native function calling so the model can pull live regulatory citations from the database rather than hallucinate them from training memory. The constrained decoding is critical — it forces the model to output JSON-schema-compliant responses with mandatory fields for the regulatory citation, the confidence band, and the lineage hash.

2. AlloyDB in northamerica-northeast2 (Montreal) — the regulatory corpus

AlloyDB is provisioned in Montreal for one reason: PIPEDA. The Personal Information Protection and Electronic Documents Act requires that data touching certain regulatory records (including RPRA Hazardous Waste Program Registry data) cannot leave Canadian jurisdiction. AlloyDB's Montreal availability zone satisfies the requirement. AWS US-East does not. This is not a preference — it is a structural compliance constraint.

The corpus AlloyDB stores includes the full text of Ontario Regulation 347 (Waste Management) at multiple revision snapshots, the RPRA HWPR specifications, USMCA Chapter 4 Rules of Origin schedules, federal CBP Customs Bulletins, and 905WOOD's own classified-decision lineage going back to platform inception.

3. Vertex AI Pipelines — the orchestration and audit trail

Vertex AI Pipelines orchestrates every inference. More importantly, Vertex ML Metadata creates what 905WOOD internally calls the Time Machine — every inference is pinned to (a) the specific model version that produced it, (b) the cryptographic hash of the dataset that retrieved the supporting citations, (c) the snapshot of regulatory text in effect at the moment of inference, and (d) the kernel hash and timestamp of the device that produced the verdict.

This means a verdict issued in May 2026 can be replayed in February 2030 against an audit request, and the replay will produce the identical output — even if the model has been updated, the database has expanded, and O.Reg 347 has been amended in the intervening years.

4. ScaNN over the regulatory corpus — sub-100ms retrieval

Google's ScaNN (Scalable Nearest Neighbors) library indexes the AlloyDB corpus for sub-100-millisecond semantic retrieval. When the 26B model receives a classification query, ScaNN surfaces the most relevant O.Reg 347 schedules, RPRA precedents, and prior 905WOOD decisions for the model to cite. This is what makes the cited citations real — they are pulled from the live corpus, not memorized.

5. The 9-Point Certificate of Origin — the output container

Every Compliance Gate verdict that survives the consensus check (E2B + E4B for Visual Fallacy, escalating to Hub XRF/NIR if disagreement) is packaged into a 9-Point Certificate of Origin. The CoO PDF is fax-ready, officer-friendly, and contains the full audit chain as embedded metadata — the model version, the regulatory citations, the lineage hash, and the kernel attestation.

What an explained verdict looks like

A Compliance Gate verdict for a 144-yd Walking Floor Mothership classified as clean wood produces a structured output that, in narrative form, reads:

That single verdict is reproducible by the regulator three years later. The model can be queried with the exact dataset and regulatory snapshot that informed the original call, and the output will be identical. This is what audit defensibility looks like in practice. This is what most AI compliance products do not provide.

Why this matters for SR&ED and CRA review

Beyond regulatory audit, the same Time Machine architecture supports Scientific Research and Experimental Development tax credit defense. CRA SR&ED reviews require evidence that R&D activity involved technological uncertainty and systematic experimentation. The Vertex ML Metadata pinning creates a verifiable record of every model iteration, every fine-tuning experiment, and every methodology revision. SR&ED filings backed by Time Machine evidence have a structurally easier path through CRA review than filings backed by descriptive narrative alone.

The Visual Fallacy guard — when sight lies

Edge AI inference is fast and cheap, but it can be fooled. Weathered gray pine looks identical to clean construction wood on visual inspection but tests positive for chromated copper arsenate (CCA) under XRF. The Visual Fallacy doctrine, codified in 905WOOD's HITL Decision Engine, requires that any visual classification with confidence between 70% and 85% be escalated to the E4B (4.5-billion-parameter) model for second-pass evaluation, with audio environmental cues (sawmill provenance, demolition site sound signature) overriding visual cues in ambiguous cases.

If E2B and E4B disagree, the load is escalated to the Hub XRF / NIR sensor for physical verification. The Precautionary Principle applies: when in doubt, default to mixed classification — the higher CIRCIL Levy is preferable to a contaminated clean-wood load reaching a BioHub.

The bottom line

Compliance is not a marketing claim. It is a chain of evidence. 905WOOD's Layer 3 Regulatory architecture — Google Gemma 4 26B on Vertex AI Pipelines, AlloyDB in Montreal, ScaNN-indexed O.Reg 347 corpus, Time Machine audit trail — is engineered to produce that evidence at every classification. Every verdict cites its source. Every source is cryptographically pinned. Every classification is replayable years later under audit conditions.

This is what 'AI-powered compliance' should mean. This is what we mean when we say it.

---