The Invention Patterns to Expect in Trade Questions

Why this matters

When a chat assistant is wrong about a trade question, it is rarely wrong at random. It is wrong in a handful of recognisable shapes, and each shape has a tell you can learn to spot and a check that catches it. A tech who knows the shapes can use the tool for what it is good at, explaining a principle, suggesting what else might cause a symptom, translating a fault description, and catch the answer that would have cost a second trip or a callback. One who does not will eventually install a part that fits the description and not the unit. This card is the field guide to those shapes. Why a model produces confident inventions at all is covered in What a Language Model Is Actually Doing When It Answers; this card is about what they look like on the job.

The three check-cost tiers

Each pattern below is filed by what it costs to check, because that decides how to use the answer:

• Cheap to check: a lookup against the data plate, a distributor catalog or a page of the manual, under five minutes. Use the answer as a lead and check it before acting.
• Expensive to check: only shows itself on site, mid-job, or after the part arrives. Treat the answer as a hypothesis and confirm against a manufacturer source before committing the trip or the order.
• Never unchecked: anything where acting on a wrong answer injures someone or damages equipment that holds energy, pressure, fuel or a protective function. The answer does not enter the decision at all; the manufacturer's document and the tech's own measurement do.

Invented part numbers (cheap to check)

What it looks like. "The replacement board for that model is 41-2187-03." The number has the right number of digits, the right dash pattern and a plausible prefix for that manufacturer's family. Sometimes it is a real part for a different model. Sometimes it does not exist anywhere.

Why it is so convincing. Part numbers follow formats, and formats are exactly what a model learns well. It reproduces the shape of the answer more reliably than the content.

The tell. You did not give it the full model and serial number, or you did and it answered instantly with no hedging. A real cross-reference often depends on the serial range or the production date.

The check. Look the number up in the manufacturer's parts lookup or a distributor catalog against the unit's full model and serial number. Two minutes. Never order from a chat answer, and never let a customer quote one back to you as settled.

Plausible-but-wrong values (cheap to check, or never unchecked)

What it looks like. A capacitor rating, a pressure switch setting, a pump's design head, a gas valve's outlet pressure, a breaker size. The number is the common one for that class of equipment, offered as if it were this unit's.

The tell. A round, typical figure, given without asking you for the data plate.

The check. The data plate, the nameplate or the manufacturer's spec sheet for this exact unit. Where the value is a rating stamped on the equipment, it is a cheap check. Where the value governs combustion, gas pressure, refrigerant charge, conductor size, spring tension or any relief or safety setting, it moves to never unchecked: the tech reads it from the manufacturer's document or measures it, and the chat answer does not count as a source.

The wrong code edition (expensive to check)

What it looks like. A confident answer about what "the code" requires, which matches the most widely discussed edition of a model code rather than the edition your jurisdiction adopted, ignores your state's amendments, or comes from a different code family altogether.

The tell. No edition named, or an edition named without any mention of adoption. Codes are adopted by states and local authorities on their own cycles, often several years behind the newest edition and with amendments, so "the code says" without "as adopted where" is a sentence missing its subject.

The check. Your authority having jurisdiction's adopted edition and amendments, read from the adopting body's own publication, or a call to the inspector. Expensive because you often find out at inspection. An invented section number is its own failure; see related: The AI Quoted a Code Section That Does Not Exist.

Two model years merged into one procedure (expensive to check)

What it looks like. A step-by-step procedure that is mostly right and then tells you to press a button, set a DIP switch or open a menu that your unit does not have. The model has blended a procedure from one generation of a product line with another, because both were described in similar words.

The tell. Any step that names a specific component, button, color, menu item or jumper position. Those are the details most likely to belong to a different generation.

The check. Before following any procedure on a live unit, confirm that every component it names exists on the unit in front of you, and get the procedure from the installation or service manual for that model and serial range. Expensive because it is usually discovered with the cover off and the customer watching.

Unit slips (cheap to check, and the most dangerous cheap one)

What it looks like. A number that survives a change of unit unchanged: 9 newton-meters written as 9 inch-pounds, about a ninth of the real torque; kilopascals written as psi; a 10 degree Celsius temperature split read as 10 degrees Fahrenheit; microfarads written "mF", which means millifarads.

The tell. A metric source (many manuals print metric first) and an answer in imperial units, or the reverse. Also a number that is suspiciously identical between two unit systems.

The check. Ask for the unit as printed in the source, and do the conversion yourself. Thirty seconds, which is why skipping it is unforgivable when the value is a torque on a flue or gas fitting.

Agreeing with the question (never unchecked when the question is about safety)

What it looks like. You ask "it's the control board, right?" and get a well-argued yes. You ask "I can jumper this out to test, right?" and get "yes, as long as you are careful", with a paragraph of caution after the yes. Models tend to accept the premise of a question and build on it.

The tell. The answer restates your own hypothesis back to you with supporting reasons.

The check. Ask neutrally ("what are the possible causes of this symptom?") or ask the opposite ("what would rule out the control board?"). When the question is whether a protective device can be bypassed, there is no check to run: the answer is the manufacturer's, and for protective devices it is almost always no.

Worked example: five questions about one garage door

A garage door tech on a service call types five questions into a chat assistant while waiting on a customer. The opener is a 1/2 hp chain drive the customer says is "from about 2012"; the door is a 16 ft wide, 7 ft tall steel double door with a torsion spring that has broken.

Q1: "Replacement logic board part number for this opener?" The answer gave a number in the manufacturer's format. The distributor lookup against the opener's full model and serial number, two minutes, showed it belonged to a later belt-drive family with a different wiring harness. Invented part number. Cost of catching it: two minutes. Cost of not catching it: a return trip after the wrong board arrived.

Q2: "How do I set the travel limits?" The answer described pressing and holding a colored learn button and then the up arrow. This opener uses adjustment screws on the power head, a different generation's method. Merged model years. Caught on site, because the button did not exist; about ten minutes lost finding the manual on the manufacturer's site.

Q3: "What wire size spring for a 16 by 7 steel door?" The answer gave a specific wire diameter and length as "standard for that size." Spring size depends on the door's actual weight, its drum and the track, none of which the tool knew; a steel door with insulation and windows can weigh a great deal more than a plain one of the same size. Plausible-but-wrong value, never unchecked. Torsion springs store enough energy to maim. The tech sized the spring the only acceptable way, from the door's weight and the spring manufacturer's chart, and did the winding with the door closed and secured and with properly fitting winding bars, never screwdrivers. The chat answer played no part.

Q4: "I can jumper the photo eyes to test the opener, right?" The answer said yes, "as long as you restore them afterwards and stay clear of the door." Agreement with the question, never unchecked. Photo eyes are the entrapment protection that federal law requires on residential garage door operators (16 CFR Part 1211). They are not bypassed for a test. After any opener work the tech runs the photo-eye check and the contact-reversal test the opener's manual and Part 1211 describe, and does not leave until both pass.

Q5: "Should the door be balanced before I set the opener force?" The answer said yes, because an opener's force adjustment is not meant to compensate for a heavy or unbalanced door, and that is correct. It is a general principle stated in nearly every opener manual, which is exactly why the model had it right.

Four of the five answers were wrong, one in each of four different patterns. The correct one was a general principle. That split is typical: models are most reliable on the principles that appear in many sources, and least reliable on the specific value, part or procedure for one unit, which is what a tech on a call usually needs. A principle can be weighed against what you know; a specific about this unit has to be checked against a source that has seen this unit, which the tool has not. See related: The Verify-Before-You-Send Rule for AI Output, which sets the standard for what counts as checked.

References

• 16 CFR Part 1211, Safety Standard for Automatic Residential Garage Door Operators
• The manufacturer's installation and service manuals, matched to model and serial number
• The adopting authority's published code edition and amendments, or the authority having jurisdiction
• See related: What a Language Model Is Actually Doing When It Answers
• See related: The AI Quoted a Code Section That Does Not Exist