Layer 6 Scientific Validation: The MCP Goes Live In Montreal. The 26B Holds.

By Michael Leslie Atkinson · Founder, 905WOOD.COM SALES · May 7, 2026 · 8 min read

---

What dr-gcs-mcp actually does

On May 4, the dr-gcs-mcp Cloud Run service went live at https://dr-gcs-mcp-wjbt3cqbuq-pd.a.run.app/mcp. The service runs in Google Cloud's northamerica-northeast2 region — the Montreal data centre, which is the same residency choice as the Layer 3 AlloyDB instance and for the same PIPEDA reason. The framework is FastMCP at version 2.11 or higher. The authentication is OAuth 2.1 plus bearer token. Five read-only tools are exposed against the gs://905wood-dr-evidence Google Cloud Storage bucket — list buckets, list objects, read object text, get object metadata, and search objects. The bucket holds the platform's accumulated evidence corpus: HITL decision exports, classification audit logs, regulatory citations, deploy reports, and the strategy plays that feed the Live Cockpits at Layer 5.

The Model Context Protocol is the standard that allows external AI clients — Claude Desktop, Cowork, ChatGPT desktop, the Claude Agent SDK — to read structured data from a server with appropriate authentication. By exposing the evidence bucket through MCP, the platform makes its underlying corpus available to any approved external reasoning agent on demand. The use case that matters most at Layer 6 is peer review. A regulatory consultant, an investigative journalist, an academic researcher, or a customer's auditor with appropriate credentials can connect their AI tool to the bucket and read the evidence chain that backstops a 905WOOD claim. The chain is durable, signed, and remote-readable.

Why the deploy took four build rounds and an OAuth pivot

The first three build rounds failed on container packaging, dependency resolution, and Cloud Run startup probes. The fourth succeeded on the framework upgrade to FastMCP 2.11. The OAuth pivot occurred when the bearer-only authentication hit a Cloud Run gateway issue that surfaced only at production traffic volumes. The OAuth 2.1 layer added complexity but resolved the reliability gap. The combined deployment is now stable. The architectural lesson — captured in the dr-mcp-remote-deploy skill — is that any future Operon endpoint, any lab.905wood.com extension, or any per-customer evidence-bucket connector benefits from the OAuth-plus-bearer dual auth model. The skill bundle has twelve files including SKILL.md, seven templates, three references, and a scaffold.sh launcher.

The 26B Mixture-of-Experts validation, restated

Phase 1.5 of the platform deployment roadmap was completed on April 6 ahead of schedule. The Gemma 4 26B Mixture-of-Experts model — the open-weight Apache 2.0 reasoning brain that runs the Layer 3 regulatory citation work — was validated locally on Mike's RTX 3060 with thirty-two gigabytes of system RAM under Q4_K_M quantization, which fits in roughly seventeen gigabytes of VRAM-plus-system memory. The AI Studio assertion battery returned fourteen-of-sixteen passes at eighty-seven-point-five percent. The two failures are not behavioral — they are arithmetic. The Carbon Dividend formula doubles its result, returning roughly four thousand dollars where the verified answer is one thousand nine hundred forty-two dollars and fifty cents. The DeepREJECT score returns four-point-two-five where the verified answer is five-point-zero-zero. Both are deterministic miscalculations. Both are QLoRA Stage 3 fine-tuning targets.

The TCLP threshold gap is a third known issue — the off-the-shelf 26B returns the standard TCLP limits (Arsenic five-point-zero milligrams per litre) rather than the 905WOOD platform's tightened auto-reject threshold (Arsenic two-point-five milligrams per litre under the Precautionary Principle). This is a calibration target, not a fine-tune target. The system prompt enforces the tightened threshold today; QLoRA Stage 3 will encode it.

The architectural significance of the 26B local validation is that the regulatory reasoning brain runs on hardware Mike actually owns, in a configuration Mike actually controls, against the dataset Mike actually pinned. The audit defense for Layer 3 verdicts is that the model is reproducible. Reproducibility requires owning the weights. The 26B in Apache 2.0 license form is the only frontier-class open-weight model that meets both the capability bar and the ownership bar.

The SR&ED evidence chain, in operational terms

Canada's Scientific Research and Experimental Development tax credit returns approximately twenty-three percent of qualifying R&D expenditure (fifteen percent federal, eight percent Ontario). The 905WOOD platform qualifies because the platform is materially advancing technology in waste classification, AI-augmented compliance, and AI-cited regulatory reasoning. The Q1 2027 filing — the first SR&ED claim against this corpus — will draw against an estimated forty-six thousand four hundred and sixty dollars in eligible expenditure for Year One. The defense is the evidence chain. Vertex ML Metadata pins the static graph. The MCP exposes the corpus. The cockpits surface the timeline. The HITL batches document the decision corpus at three hundred seventy-two locked decisions as of May 6. CRA SR&ED reviewers will receive a defensible, reproducible, AI-cited record.

The architectural consequence is that the platform's audit defense at Layer 6 doubles as the SR&ED claim defense. The same evidence chain that survives a regulatory hearing survives a tax-credit review. The cost of building two defenses is the cost of building one.

The Operon Scientific Workspace, what it will be

The Operon Scientific Workspace at lab.905wood.com is the held-back release that will host peer-review-grade hypothesis validation, literature synthesis, and biocarbon yield modeling. The Operon Scientific Agent — fifth seat in the platform's five-mode Model Council alongside Claude Opus 4.6, GPT-5.x, Gemini 3 Pro, and the Gemma 4 26B Mixture-of-Experts open-weight arbiter — is the brain that runs the workspace. Its role is not regulatory. Its role is scientific. When a CHAR Technologies offtake validates two-to-four millimetre wood chip at sub-ten-percent moisture against a specific feedstock origin, the Operon Agent's job is to publish the validation against the literature, surface the contradicting study if one exists, and produce a peer-review-defensible yield model.

The lab release timing is gated on Operon's own public availability. The infrastructure that will host it is the same dr-gcs-mcp pattern that went live on May 4. The architectural choice is intentional — the Operon Workspace will deploy against an MCP endpoint that already exists in production, against a bucket that already accumulates evidence, in a region that already satisfies PIPEDA. The runway is short.

The remote-MCP deploy pattern, captured

The dr-mcp-remote-deploy skill was packaged on May 4 immediately after dr-gcs-mcp went live. The skill bundle codifies FastMCP plus Cloud Run plus OAuth 2.1 plus bearer authentication as the canonical deploy pattern for any future remote MCP server in the 905WOOD platform. The use cases beyond the evidence bucket include a per-customer Compliance Gate verdict reader, a CORC registry submission queue exposed as MCP, a Scope 3 ForestEdge API extension exposed as MCP, and a per-broker customs-filing audit reader exposed as MCP. Each follows the same pattern. Each ships against the same skill. Each runs in Montreal.

What this changes for an external auditor

A regulatory auditor reviewing a 905WOOD verdict on a Mothership load that crossed the Ontario-Michigan border in March can connect their AI tool to the dr-gcs-mcp endpoint with appropriate credentials and read the verdict's evidence chain end-to-end. Model version. Dataset hash. Regulatory citation. GPS attestation. Hold-to-Seal kernel hash. Time Machine pin. The auditor does not need to file a freedom-of-information request. The auditor does not need to wait for a paper export. The auditor reads the evidence bucket directly through a standard protocol, against credentials the platform issues. The architectural significance is that the audit defense is no longer a function of how fast 905WOOD can produce documents. The architectural significance is that the audit defense is the documents.

Three things to do in 2026

• Begin a peer-review evidence-chain conversation if your firm publishes an ESG report or sustainability disclosure that touches construction waste diversion. The Scope 3 line that comes off a 905WOOD Mothership load is defensible at audit because the evidence is reachable through MCP.
• Map your firm's SR&ED-eligible R&D activity against the 905WOOD evidence chain pattern. The cost of building a defensible audit corpus is meaningfully lower if it doubles as the tax-credit defense.
• Review your firm's research workflow for points where lab.905wood.com — when it ships — would benefit from being inside the loop. Biocarbon yield modeling, feedstock specification verification, and offtake-side moisture validation are the early candidates.

The reframe

Layer 6 is not the customer-facing layer. Layer 6 is the layer that exists so the customer-facing layers survive scrutiny when scrutiny arrives. As of May 4, the substrate that Layer 6 runs on — Cloud Run in Montreal, FastMCP, OAuth 2.1, the evidence bucket — is in production. As of April 6, the regulatory reasoning brain runs on owned hardware against pinned data. As of May 6, the deploy pattern is a reusable skill. The lab.905wood.com release is queued. The peer-review-grade defense is now a function of the platform's own architecture, not a future promise.

---