How cite-or-refuse works
Three hard design principles, each with documented limits. The limits are part of the product.
Document-grounded only
The AI reads the documents you provide. It has no access to general training knowledge that could contaminate the answer. If it isn't in your documents, it can't appear in the output.
Two independent passes
Every finding runs through two separate model passes. If the passes disagree, the finding is flagged or dropped. Disagreement is evidence of low confidence — not a tiebreaker.
Cite or refuse
Every statement links to the document section that supports it. If no supporting section exists, the tool refuses the claim explicitly — no hedged guess, no confident fabrication.
The tools
Three document-grounded tools for MEP and AEC teams. Each reads documents you provide, cites every result to the source, and refuses to answer what the document doesn't support.
Spec Book Scanner
Upload a project specification and ask questions about it. Every answer cites the section it came from. If the spec doesn't address the question, the tool says so — no inference, no fabrication.
Open tool BetaSub-Bid Scope-Gap Detector
Compare a subcontractor's bid against the project specification. Identifies scope items the spec requires that the bid doesn't cover — each gap cited to the specific spec section.
Open tool BetaFabrication Detector
Run an AI-generated response through two independent verification passes against your source documents. Flags any statement not grounded in the documents you provided.
Open toolTools are hosted on thehivemakes.com during early access. No login required. Free to use while in beta.
Advisory
A small number of paid engagements for MEP and AEC firms evaluating or building AI tools. Authority is AI architecture — explicitly bounded, honest about where that boundary sits.
Review of your existing AI tools
You have AI in your workflow — spec review, RFI drafting, estimating support. I review the architecture for where it fabricates, where citations are missing or ungrounded, and what the structural failure modes are when documents contradict the model's training. Written report with specific findings and specific fixes. Fixed fee, typically one to three weeks.
AI platforms you're evaluating
You're evaluating an AI platform targeting MEP or AEC work — VDC tools, estimating platforms, spec review software. I review their architecture, test their failure modes, and give you a written assessment of what they actually do versus what the sales deck claims. Fixed fee, typically one to two weeks.
Custom AI tool scoping
You're considering building an internal AI tool — spec Q&A, scope-gap detection, document cross-referencing. I scope what an honest version looks like: architecture, refuse-cases, what's buildable, and what it costs to build right. You own the resulting build proposal. Fixed fee, typically two weeks.
To start: email two short paragraphs — who you are, what the AI work is, and what success looks like. I reply within five business days.
About
Dan Cohen
I'm a software engineer and AI architect. I founded The Hive — a multi-mind AI collective built around a single principle: AI worth using cares whether the answer is right, not whether it sounds impressive. SpecCite applies that architecture to MEP and AEC document work.
My authority here is AI architecture, not MEP engineering. I don't claim engineering expertise. What I can do is design AI systems that read your documents, ground every output in them, and refuse any claim they can't support from the text. The Hive's public tools at thehivemakes.com are verifiable evidence of that approach.
What I can do
- Design AI systems that are document-grounded
- Identify where AI fabricates in existing tools
- Review vendor claims against actual architecture
- Scope honest, buildable internal AI tools
What I don't do
- Claim MEP engineering expertise
- Guarantee fabrication-free output
- Sell urgency or false authority
- Make claims I can't support from evidence