How this research was produced, why transparency is critical for credibility, and why AI-assisted analysis produces more rigorous results, not less.
The research on this site was produced using a purpose-built research system constructed on top of Claude (an AI assistant made by Anthropic), running locally via the Claude Code interface.
It is a structured set of instructions, guidelines, domain knowledge, and persistent memory that shapes how the AI behaves when asked to analyze energy procurement documents, environmental assessments, and regulatory filings. The system includes:
The result is an AI agent that reads like a domain-aware research assistant. The quality comes from the system's design and the researcher's attention and verification, not from the AI alone.
I'm up-front about this because I believe AI can fundamentally change who gets to participate in public policy debates. A 149-page tolling agreement, two 200-page environmental assessments, seven RFP appendices, and a shelf of federal regulations is a volume of material designed for law firms and engineering consultants, not community members with day jobs. AI makes it possible for a person to read, cross-reference, and analyze that volume. I see AI-assisted research as giving folks access to the same types of tools and resources larger organizations have.
Source documents are loaded directly into the AI's context window — the full text, not summaries. For this project, that includes:
The AI reads the actual documents. It works from the primary sources rather than searching the internet for summaries or relying on training data
The researcher asks specific questions: "What does the Tolling Agreement say about fuel cost allocation?" or "Extract the water demand figures from Section 3.3 of the Marshdale EARD." The AI extracts the relevant passages and presents findings with page/section references.
This is not "ask AI a question and publish the answer." The researcher already has an idea of what to look for and uses the AI to process volume efficiently. A single Environmental Assessment Registration Document can be 200+ pages. The AI reads all of it and the researcher directs where to look and what matters.
All quantitative claims on this site — emission intensities, annual CO2 tonnage, cost projections, capacity factor analysis — are computed by Python scripts, not generated by the AI's language model. LLMs are powerful text processors, but they are not calculators. They can get arithmetic wrong, especially with multi-step formulas or unit conversions.
For example, the emission intensity figure of 531 t CO2/GWh is calculated from EPA AP-42 emission factors and heat rate data using a Python script that shows its formula, inputs, and intermediate values. The AI writes the script while the researcher reviews the formula and checks the output against known benchmarks.
Every factual claim is checked against the source document. When the AI says "the Tolling Agreement allows IESO to extend the term to 27 years," the researcher verifies that specific clause exists in the agreement.
The AI is explicitly instructed: never guess. If a fact isn't in the source documents, the AI says so rather than filling in plausible-sounding information. This is enforced through standing instructions built into the research system and verified through spot-checking.
Every claim on this site links back to its source. Not "according to reports" but specific documents, specific sections, specific data tables. Where possible, I link directly to the publicly available source so readers can verify independently.
Sources include government filings on the NS Environment EA registry, federal regulations published in the Canada Gazette, corporate technical specifications, and peer-reviewed engineering data. I only cite social media and opinion pieces if they are the sole source and directly relevant. I do not cite anonymous sources. (other than EA comment submissions)
Before criticizing any aspect of the proposed projects, the research process requires constructing the strongest possible version of the proponent's argument. What is the best case for these gas plants? What legitimate grid needs do they serve? What are the real limitations of alternatives?
This site acknowledges that Nova Scotia faces real grid stability challenges as coal retires. It acknowledges that batteries have duration limits. It acknowledges that grid-forming and stability is a genuine engineering need. The critique is that these specific plants, procured this specific way, are not the best answer, not that the problem doesn't exist.
A single person cannot carefully read a 149-page tolling agreement, two 200-page environmental assessments, a draft RFP with seven appendices, and comparable documents from four other provinces, and cross-reference them all, in a reasonable timeframe. AI handles the volume while the human handles the judgment.
AI doesn't get tired on page 140 and start skimming. It applies the same level of attention to the last appendix as the first page. It catches the clause buried in Exhibit T that contradicts the summary in the RFP overview.
The AI has no financial interest in the outcome. It doesn't work for a gas company, a battery manufacturer, or an environmental group. It reports what the documents say. The human researcher has a perspective, which is disclosed, but the underlying data extraction is neutral.
Transparency requires acknowledging limitations:
The source documents referenced on this site are linked to the Sources page which contains links to the online version of the resource. Deep-linking directly to the relevant text is challenging so the relevant document sections are displayed with the source reference. The methodology is described above. Anyone with access to the same documents and a capable AI tool can reproduce these findings. The point is to build research that is verifiable, not proprietary.
If you want to build a similar research system for your own community issue, I'm happy to share the prompts, instructions, and markdown source files that power the system. Get in touch and I'll help you get started.
This research is conducted by Jason Hurst, a systems analyst with a focus on data and maps, and a Pictou County resident. I have no financial interest in any energy company, battery manufacturer, or consulting firm involved in these projects. My interest is ensuring that Nova Scotia's energy decisions are made with full public information and proper oversight.