It computes with a citation,
or it refuses.
Most AI in the trades has one setting: confident. Ask it what size system a house needs and it will tell you, wrongly, in a tone that gets a compressor killed and a contractor blamed.
The Register OS has two settings. It computes — from your numbers or from a published standard, showing the arithmetic and naming the clause. Or it refuses. There is no third mode where it guesses.
Runs entirely on your device. No signup, no card, no data leaving the browser to do arithmetic.
“What size AC does a 2,000 sq ft house need?”
Building-specific. Needs a Manual J with envelope, orientation, infiltration and duct losses. Any confident number here is a fabrication, and it is the kind that gets installed.
“12 × 14 × 9 ft room, wall-mounted, R-454B, 7.4 lb charge.”
Not a matter of opinion. UL 60335-2-40 Annex GG: m_max = 2.5 · LFL5/4 · h₀ · √A
= 2.5 × (0.293)1.25 × 1.8 m × √15.61 m²
= 4.30 kg → 9.48 lb
7.4 lb sits at 78% of the screening limit. Then it tells you the OEM instructions govern and are frequently stricter — because they do, and they are.
The old line was technical versus business. That was the wrong line.
Refusing everything technical is safe and nearly useless — it walks away from the compliance and diagnostic work where the real money and the real liability both live. The honest line is not what the question is about. It is whether the answer can be derived or only generated.
Generated → refused
Equipment sizing. Duct design. “Why is it short-cycling?” Anything where the answer depends on a building, an install, or a component the software cannot see. A language model will produce a fluent answer to every one of these. That is the problem, not the feature.
Derived → computed
A2L charge limits from LFL and room geometry. Conductor ampacity with the 110.14(C) termination cap applied in the right order. EPA 608 annualised leak rate. NETA thermographic classification. Every one shows its arithmetic and names its clause.
Yours → computed
Loaded cost per billable hour. What a 10% discount does to gross profit. When the account goes negative and by how much. Arithmetic on your own numbers, which needs no citation beyond showing the working.
Where a value is jurisdiction-dependent, it says so and refuses to pretend otherwise. Code adoption lags publication. States amend. Workers' comp rates vary several-fold. OEM instructions override the standard and are frequently stricter. Every one of those is flagged in the output as a hard caveat you cannot dismiss — because a screening result presented as an approval is worse than no screening result at all.
Wages are paid on every hour. They come back on the hours you invoice.
That sentence is the whole business. A technician on $32 an hour is paid for 2,236 hours a year and can only be invoiced for about 1,300 of them. Divide burdened cost by the wrong number and every job in the price book is underwater by construction — which is how a shop can be flat out all summer and finish the year with nothing.
Move the sliders. This is the real engine from the app, not a marketing approximation.
Of available hours. Paid, not billable.
High-efficiency shops run 65–75% billable. Every point is worth — per technician per year.
| Component | Annual | Per billable hr |
|---|
Fifteen workbenches. One profile. Nothing entered twice.
The previous version of this product was five isolated calculators. You entered a labour rate in one and a different one in the next, so nothing agreed with anything, so nothing could be trusted. Now there is one profile: change the comp rate and it moves the price book, the callback cost, the hiring case and the valuation.
The ledger — what the shop earns
- Loaded rates. Statutory burden with the wage-base caps applied properly, comp per $100, benefits, the van, and the honest billable-hour denominator. Two multipliers, because the one you have heard of excludes the truck.
- Pricing. Overhead absorbed per billable hour rather than as a percentage of parts. Reserve and card fees solved out of price instead of added to cost. Good/better/best that holds margin. And the discount ladder — at 45% margin, 10% off costs 22% of the profit and needs 29% more work to stand still.
- Callbacks. Prices the displaced job, not just the truck roll. That term is usually the largest and it never appears on a P&L, which is why a 12% callback rate feels survivable.
- Lead channels. Contribution ROAS on gross profit, not revenue. Wilson confidence bounds on the booking rate, so no channel gets scaled or killed on nine jobs of noise.
- Cash & season. A 52-week ledger with collections lagged through your ageing profile, variable cost that actually varies, supplier terms, the revolver, and the predictable lumps. Tells you which week you go negative.
- Capacity, agreements, software, enterprise value. Hire versus overtime versus subcontract versus decline — each charged for the demand it leaves uncovered. Plan pricing against real delivery cost. Termination traps in your stack. And what a buyer would pay, driver by driver.
The field — what you are standing in front of
- A2L charge limits. UL 60335-2-40 Annex GG, worked line by line, in both unit systems. Minimum area by mounting option for a fixed closet, and maximum charge by refrigerant when you are choosing equipment.
- Circuit readings. Separates restricted airflow from low charge. Both depress suction pressure; air ΔT and static pressure move in opposite directions. Ranked hypotheses, the confirming test, and what acting wrongly costs — never “add refrigerant”.
- Electrical. Ampacity derated from the insulation column then capped at the termination column, in that order. 240.4(D) with the Article 440 exception. Working clearance. Subpanel bonding.
- Leak rate. The §82.157(b) annualising formula, the trigger for your appliance class, the 30-day clock — and the repair-versus-feed-it payback that closes the job.
- Plumbing & hydronic. Fixture-unit tally with sizing. Flow from load with the glycol constants that stop a loop being under-pumped. Velocity limits both ways. Thermal expansion on a closed system.
All field output is screening. It does not replace the OEM installation instructions, the authority having jurisdiction, or a licensed professional — and it says so on every result rather than in a footer nobody reads.
Every number carries a label saying how much it deserves to be trusted.
Not a footnote. A structural field on every result, because the difference between a code clause and a forum consensus is the difference between something you can stake a licence on and something you cannot.
The engines are pure functions with 180 tests pinning the arithmetic — including the ones that caught two real bugs during the rebuild: a burden multiplier conflating the van with statutory cost, and a hiring comparison that rewarded covering less of the demand.
Written down in a standard, code or statute. Reproducible by anyone holding the document. NEC Table 310.16. UL 60335-2-40 Annex GG. 40 CFR §82.157. NETA MTS.
A specific named third-party dataset or filing. IRS Publication 15 wage bases. IPCC GWP values. KFF employer contribution survey.
An industry range compiled from operator reports. Directional, never authoritative for your shop. First-time fix bands. Transaction multiples. Technician replacement cost.
Computed from your own inputs. As good as your data and no better — which is why the app pushes you to replace the sample profile before trusting anything.
A default chosen so the tool works before you have typed anything. Always overridable, always labelled, never quietly load-bearing.
The calculators are free. Permanently.
Every engine on this page runs in your browser. There is no server to pay for, so there is nothing to charge for. A product built on the claim that it will not mislead you should not start by putting the arithmetic behind a card form.
- ✓ All ten ledger workbenches
- ✓ All five field screening workbenches
- ✓ Full loaded-rate and margin engine with worked arithmetic
- ✓ Citation and confidence labelling on every number
- ✓ Profile persisted locally — nothing transmitted
- ✓ Works with no signal, in a basement, on a phone
The paid tier is for the things that genuinely need a server, and it is honest about not existing yet.
- · Multi-user shops with a shared price book
- · Asset history — plate captures, readings and compliance records per unit
- · Bidirectional sync with your existing field-service platform
- · Warranty claim packaging from the readings already captured
- · Board-ready exports for lenders and acquirers
The ones worth asking.
Didn't this product used to refuse all technical questions?
Yes, and the line was drawn in the wrong place. Refusing everything technical is safe and it walks away from the compliance and diagnostic work where both the money and the liability actually are. The distinction that holds up is not what the question is about — it is whether the answer can be derived from a published relation and shown, or only generated. A2L charge limits are arithmetic on constants published in a standard. Equipment sizing is not. One gets computed with the clause named, the other gets refused. The trust mechanism is stronger, not weaker: “here is the formula, here are the constants, here is the clause, and here is what overrides me” beats a blanket no.
Is the compliance screening safe to rely on?
It is safe to rely on as a screening step and it is not an approval, which is exactly what every result says on its face. Three things override it, in order: the OEM installation instructions for the listed equipment, the standard as adopted and amended by your AHJ, and the designer of record on engineered systems. What it removes is not professional judgement — it is the situation where a technician is standing in a mechanical room with no signal and no manual, and guesses.
How is this different from asking a general-purpose AI?
A general model generates the most plausible-sounding continuation. On NEC ampacity that means it will usually cite the right table and then apply the derating and the termination cap in the wrong order, confidently, and you will not be able to tell. These engines are deterministic functions with 180 tests pinning them. Same inputs, same outputs, arithmetic shown, and the tests fail if anyone breaks them.
Does it replace ServiceTitan, Jobber or Housecall Pro?
No. Those handle dispatch, scheduling and invoicing and this does none of that. It sits beside them and answers what they do not: what an hour actually costs you, what a discount actually costs you, which week the account goes negative, and whether the charge in front of you is inside the limit. The software workbench will also tell you, uncomfortably, whether the platform you are paying for returns more than it costs — including the termination clause you signed.
Where does my data go?
Nowhere. The engines are pure functions that run in your browser and your profile is stored in IndexedDB on your device. A shop's wage rates and margin targets are the most commercially sensitive information it holds, and there is no reason for them to travel to run arithmetic. The practical trade-off is real: clearing site data clears your profile, and it does not follow you to another device. The waitlist form above is the only thing on this site that transmits anything, and it says so.
Is any of this legal, engineering or financial advice?
No. It is a business-judgement and code-screening tool. It is not a licensed engineering, code-compliance, safety, legal, accounting or financial advisory service. Financial outputs are estimates for planning. You are responsible for verifying trade-specific requirements against your own jurisdiction, carrier, and the manufacturer's documentation.
Find out what an hour actually costs you.
Ten minutes with your real wage, comp rate and overhead. No account, no card, nothing leaves the device.
Open the app