Of all the AI-assisted documents in this level, the code-compliance narrative is the one most likely to embarrass you, because it is the document where the model's two worst habits, fabricating citations and citing the wrong edition, meet the document type that a plan checker reads with the express purpose of finding exactly those errors. An AI code narrative looks authoritative and can be salted with sections that exist in no edition of the code, and the plan checker who knows Chapter 10 cold will bounce it within the hour. This lesson shows you how to use AI for the genuine value it offers on code work, a fast first-pass draft and a focused reading map, while applying the verification that this document, above all others, demands, because here the unverified citation does not just embarrass you, it gets your permit set rejected.

Why This Is the Most Dangerous Document for AI

The code narrative concentrates every hazard the program has warned about into one document. From the hallucination lesson, the model fabricates plausible-looking code citations, the egress section that appears confidently and exists nowhere. From the training-cutoff lesson, the model reaches for the code edition it read most of, citing the 2018 or 2021 IBC when your jurisdiction adopted the 2024. From the tier lesson, code interpretation for a stamped compliance determination is tier three, the licensed professional's act, never the AI's. And the audience is uniquely adversarial: a plan checker reviews the narrative specifically to find non-compliance and fabricated or wrong citations, so the document's reader is actively hunting exactly the errors AI is prone to produce.

This combination, the model's citation-fabrication and edition-drift habits meeting an expert reader looking for those very errors on a document that gates the permit, is why the code narrative demands the strictest verification in the level. It is also why the value, when captured correctly, is real: writing a code narrative from scratch is slow, and a fast first-pass draft plus a map of which code sections to read saves genuine time, as long as every citation is verified before the narrative leaves your hands. The document is dangerous and valuable in the same measure, and the entire skill is capturing the value without shipping the danger to the one reader most equipped to catch it.

What AI Does Well on Code Work, and What It Must Not

Mapped onto the tier system from Level 1, AI's legitimate role on code work is tiers one and two, and tier three is barred. Tier two, orientation, is truly useful: the model can summarize which code sections govern a condition so you know where to read, give you an overview of the egress requirements to focus your own review, and point you at the relevant chapters, which accelerates orienting yourself in a thousand pages of code. Tier one, drafting, is also useful: once you have done the analysis, the model can draft the code narrative's prose, turning your verified determinations into clear, well-organized language faster than writing from a blank page.

What is barred is tier three: the model never makes the compliance determination that the narrative supports, because whether the design actually complies is a licensed professional's judgment that ends in a stamp, and an AI answer to "does this comply" has no authority no matter how confident it sounds. The dangerous slide, named in the tier lesson, is when a tier-two orientation ("summarize the egress requirements") quietly becomes a tier-three determination ("the design complies with these requirements") because the model fluently produced both in one response and a rushed reviewer adopted the conclusion. So the code-narrative workflow keeps the determination human and uses AI only to orient the reading and draft the prose around determinations the professional made, which is the tier discipline applied to the single document where crossing the line is most tempting and most costly.

The dangerous slide is when "summarize the egress requirements" quietly becomes "the design complies with these requirements" in one fluent AI response, and a rushed reviewer adopts the conclusion. The model orients the reading and drafts the prose; the compliance determination, the stamped act, stays human.

The Citation Verification: Every One, Against the Published Code

The non-negotiable core of the workflow is that every code citation in the narrative is verified by hand against the published, adopted code before the narrative goes anywhere, and the verification checks three things from the hallucination and cutoff lessons. Does the section exist? Does it say what the narrative claims? Is it from the correct adopted edition for this jurisdiction? The model fails on all three, inventing sections, misstating what real sections require, and citing the wrong edition, so the verification hunts each citation and confirms it against the actual code, marking every fabrication and every wrong-edition citation for correction.

The IBC egress narrative is the canonical example because it is where the fabrication is most common and most catchable: the model salts a means-of-egress narrative with section numbers that sound like IBC Chapter 10 and do not exist, and the plan checker who knows Chapter 10 finds them immediately. So you open the published IBC 2024, you check every cited section against it, you confirm each says what the narrative claims and is from the 2024 edition your jurisdiction adopted, and you mark every section that does not exist or is misstated. The same applies across the codes the narrative touches: ASCE 7-22 for wind and seismic, NEC 2023 for working clearances, each cited provision verified against the actual published standard and edition. This is the citation-verification discipline from Level 1, applied at its highest stakes, and it does not pass until every citation in the narrative is confirmed real, correctly stated, and correctly editioned, because the one you do not check is the one the plan checker will.

The Plan-Check Response Packet

The code narrative usually exists in service of a plan-check submission, and the workflow extends to the plan checker's response, where AI helps again with the same discipline. When the plan checker returns comments, often a list of items where they found non-compliance or questioned a provision, you respond, and AI can draft the response to each comment, citing the provision and explaining the compliance, once you have made the determination of how the design actually complies. The plan-check round-trip is a high-friction, deadline-driven exchange where fast, well-organized responses move the permit, and AI's drafting speed helps, with the citations verified exactly as in the original narrative.

The discipline on the response packet is the same and the stakes are if anything higher, because you are now responding to a reader who already found problems and is reading your response skeptically. A response that cites a fabricated section in answer to a plan-check comment does not just fail to resolve the comment, it tells the plan checker your submission cannot be trusted, which slows every subsequent review. So the response packet gets the full citation verification, every provision cited in every response confirmed against the published code, and the compliance determinations behind the responses are the professional's, with AI drafting the language. Done right, the packet responds to every comment with verified citations and sound compliance reasoning in a form that moves the permit, and the AI accelerated the drafting of a response whose substance and citations the professional verified, which is the pattern of the entire level applied to the permit round-trip.

The Edition Trap and the Jurisdiction Question

The wrong-edition failure deserves its own attention because it is subtler than outright fabrication and just as fatal at plan check. Codes adopt on jurisdiction-specific cycles, so the code in force for your project is whatever edition your authority having jurisdiction has adopted, which may not be the newest published edition and is almost never the edition the model defaults to. The model, frozen at its training cutoff and weighted toward whatever editions dominated its training data, will confidently cite the 2018 or 2021 IBC provisions when your jurisdiction is on the 2024, and because a real provision is cited, just from the wrong edition, the error survives a check that only confirms the section exists somewhere.

This is why the edition is a distinct verification step, not folded into the existence check: a citation can be to a real section that exists in some edition and still be wrong because it is not the edition your AHJ enforces, and the requirements can differ between editions in ways that change compliance. So you establish, first, which edition your jurisdiction has actually adopted, a fact you confirm from the AHJ and not from the model, and then you verify every citation against that specific adopted edition. The jurisdiction question, which edition is in force here, is exactly the present-state-of-the-world fact the model cannot reliably know, so it is the human's to establish and the anchor against which all the citations are checked. A narrative perfect in every other respect but written against the wrong edition is still a rejected narrative, and the edition check is what prevents the model's edition-drift from quietly editioning your whole narrative wrong.

The Real Payoff, and the Floor Below Which AI Does Not Help

It is worth being precise about where AI truly accelerates code work and where it offers nothing, because the line protects you from over-relying on it. AI accelerates the orientation, finding which sections to read in a thousand pages, and the drafting, turning your verified analysis into clean prose, and across a code narrative those two are real time savings. What AI does not accelerate, and cannot, is the analysis itself, the determination of whether this specific building's egress, wind resistance, or electrical clearances actually comply, because that is the licensed judgment the narrative exists to document, and no orientation or drafting speed substitutes for it.

This matters because the temptation on a deadline is to let the fast orientation-and-drafting feel like the analysis was done, when the narrative reads complete and authoritative before any actual compliance determination was made. The floor below which AI does not help is exactly the compliance analysis: you can have a fully drafted, well-organized, fluent code narrative that is worthless because the determinations in it were never actually made by a professional examining the design against the verified code. So the discipline is to recognize that AI got you a faster draft of the narrative's form, and the substance, the determinations and the verified citations, is still entirely the work it always was, only now you spend your time on that substance instead of on the prose and the section-hunting. The payoff is real, hours saved on form, but the floor is firm, the analysis and verification are not accelerated, and confusing the fast form for done substance is the specific way this document's speed becomes its danger.

The Applied Problem: Egress Narrative and the Fabrication Audit

Here is the exercise. Take a real or representative building, for instance a four-story Type IIB building, and produce an AI-assisted means-of-egress code narrative against IBC 2024 Chapter 10, then audit it for fabrication and prepare the plan-check response approach. Run the workflow: use AI for tier-two orientation to map which Chapter 10 sections govern the egress for this building, make the compliance determinations yourself as the responsible professional, use AI to draft the narrative prose from your determinations, and then audit every citation against the published IBC 2024.

The audit is the heart of the exercise and it mirrors the Level 1 fabrication audit exactly: for every section the narrative cites, confirm against the published IBC 2024 that the section exists, says what the narrative claims, and is from the 2024 edition, and mark every fabrication with the disproving reference, "narrative cites 1029.6.4; no such section in IBC 2024; the governing egress-width provision is at the actual section." Tally the fabrications, because the count, as in Level 1, is the proof of why this audit is non-negotiable, and on a code narrative you will likely find at least one confident citation to a section that does not exist. Then note how you would handle the plan-check response, with the same verification on any cited provision.

The deliverable is the verified egress narrative, the fabrication audit marking every non-existent or wrong-edition citation, and the plan-check response approach. The lasting product is a code-narrative workflow that captures AI's drafting and orientation speed while guaranteeing that the document a plan checker reads to find your errors contains none of the citation errors AI is prone to produce. This is the highest-stakes application of the level's entire discipline, tier-respecting use, citation verification, human determination, and getting it right here means you never hand a plan checker the fabricated section that gets your permit set bounced within the hour. The professional who runs this drafts code narratives faster and submits them cleaner, which on the permit critical path is schedule protected and credibility preserved, achieved only because the determination stayed human and every citation was verified against the published code.

Key Takeaways

• The code narrative is the most dangerous AI document because it concentrates the model's worst habits, fabricating citations and citing the wrong edition, into the document a plan checker reads specifically to find those exact errors, on the document that gates the permit.
• AI's legitimate role is tier two (orientation: which sections govern, where to read) and tier one (drafting the prose from your verified determinations). Tier three, the compliance determination that ends in a stamp, is barred; an AI "does this comply" answer has no authority however confident.
• The dangerous slide is a tier-two orientation becoming a tier-three determination in one fluent response that a rushed reviewer adopts. The workflow keeps the determination human and uses AI only to orient and draft around it.
• Every citation is verified by hand against the published, adopted code, checking three things: does the section exist, does it say what the narrative claims, and is it the correct edition. The model fails on all three, so the verification hunts each citation and marks every fabrication and wrong-edition citation.
• The IBC egress narrative is the canonical case: the model salts it with Chapter 10-sounding sections that do not exist, and the plan checker finds them immediately. The same verification applies to ASCE 7-22 wind, NEC 2023 clearances, and every standard the narrative touches.
• The plan-check response packet gets the same discipline at higher stakes, because a fabricated citation in a response to a skeptical plan checker who already found problems tells them your submission cannot be trusted, slowing every subsequent review.
• The artifact: an AI-assisted egress narrative for a four-story Type IIB building, a fabrication audit marking every non-existent or wrong-edition citation against the published IBC 2024 with a tally, and the plan-check response approach, so the document a plan checker reads to find your errors contains none of AI's citation errors.