The Reference Device

Why Hardware Characterization Is a Precondition for Cannabis Efficacy Research

Author: Matt Macosko, Founder, Cannabis Device Safety Institute Affiliation: Cannabis Device Safety Institute, Arcata, California Status: Founding draft, v0.1 (July 2026) Suggested citation: Macosko, M. The Reference Device. CDSI Working Paper P-011. 2026.

Abstract

Cannabis is most commonly consumed by inhalation, and inhalation requires a device. That device is heated, and heating it releases compounds of its own into the same stream that carries the cannabis. Yet in nearly every published study of inhaled cannabis — analytical or clinical — the delivery device is described in a sentence, or not at all, and is never characterized. This paper argues that the omission is not a matter of incomplete reporting but a structural defect in the inferential chain. When a device’s chemical contribution is unknown, every measured outcome has at least two candidate sources: the cannabis material and the hardware that heated it. We name this condition device confounding, and we argue it is an attribution failure rather than a statistical one — it is not reduced by increasing sample size, randomizing subjects, or blinding investigators, because it is not noise. The remedy is a reference device: hardware whose own contribution has been characterized across its operating envelope and quantified below a defined attribution floor for a named analyte panel. We argue that reference devices do not currently exist for cannabis, that constructing them is a tractable analytical problem requiring no access to controlled substances, and that until they exist, the question “does this cannabis medicine help this person?” cannot be cleanly answered — only approximated, with an uncontrolled variable in the middle. The paper closes by specifying what reference-device qualification would require, and by identifying an unresolved problem: reference status may decay with device use.

Keywords: cannabis, inhalation, delivery device, confounding, attribution, reference material, method validation, efficacy research

1. Introduction

There is a question that a great many people want answered, and it is not really a hardware question. It is: does this help?

Does inhaled cannabis reduce this person’s pain, or their nausea, or their seizure frequency, or their intrusive thoughts at three in the morning? That is the question patients ask, and it is the question that justifies most of the public money that has ever gone into cannabis research. It is a medical question, and I am not a physician. I build vaporizers.

I want to make an argument that connects those two things, because I think the connection is load-bearing and I think it is being missed.

The argument is this. You cannot answer does this help until you can say what “this” was. And in inhaled cannabis, “this” is never only the cannabis. It is the cannabis plus whatever the device contributed on the way to the lungs. If you do not know what the device contributed, you do not know what you administered. And if you do not know what you administered, then whatever you measured afterward — relief, harm, a compound in the blood, a shadow on a scan — you cannot say with confidence what produced it.

This is not a complaint about sloppy science. The people doing this work are careful. It is a claim about a missing instrument. Every measurement science eventually has to confront the fact that its apparatus contributes to its readings, and every one of them solves it the same way: characterize the apparatus first, then measure through it. The analytical chemist runs a blank before running samples. The inhalation pharmacologist validates the nebulizer before dosing a subject. The spectroscopist calibrates against a certified reference material. This is not advanced practice; it is the first week.

Cannabis research skips it. Not out of carelessness — out of the institutional condition described in P-001 as the Hardware Vacuum. No one was ever responsible for the device, so no one characterized it, so there is nothing to characterize it against. The blank was never run because there is no blank.

From where I sit, that is the real reason a hardware institute has to exist. Not primarily to protect consumers from bad coils, though it should do that. But because hardware characterization is upstream of the medical question everyone actually cares about, and someone has to go stand upstream.

2. Device confounding

2.1 The structure of the problem

Consider a study — analytical or clinical, it does not matter yet — in which a subject or a sampling train receives inhaled cannabis through a vaporizer, and something is measured downstream. A carbonyl in the condensate. A metal in the blood. A change in reported pain.

Ask the simplest possible question about any positive finding: where did it come from?

There are two answers available, and the study design does not distinguish them:

The material. The cannabis, its cannabinoids and terpenes, its residual solvents, its pesticide load, its adulterants.
The machine. The heating element, its alloy or ceramic or quartz, its binder and glaze, its solder joints, its wick, its airpath, its adhesives — each at temperature, each doing chemistry of its own (Lexicon §1.1).

The second source is invisible in the write-up. It is generally represented by a product name and, if the reader is lucky, a wattage.

We name this condition device confounding (Lexicon §2A.1): the systematic error introduced into any study of inhaled cannabis when the delivery device’s own chemical contribution to the inhaled stream is uncharacterized.

2.2 Why this is not a statistics problem

It is tempting to file this under “limitations” and move on — one more source of variance, to be swamped by a larger n. That instinct is wrong, and the reason matters.

Statistical machinery handles noise: variation that is random with respect to the thing you are studying. Add subjects, and noise averages out.

Device confounding is not noise. It is structured, directional, and correlated with the exposure itself — you only get the device’s contribution when you dose, because dosing is heating the device. The hardware’s signal arrives in lockstep with the material’s signal, from the same puff, into the same lung, at the same moment. It does not average out with more subjects. It averages in, and it recruits more of itself with every draw.

Worse, it is often correlated with the very variables a study cares about. Higher dose means more draws means more device contribution. Higher temperature means more cannabinoid volatilization and more off-gassing from the heater (Lexicon §2.3). A dose-response curve run on an uncharacterized device is measuring two dose-response curves added together, and reporting the sum as one.

So this is not a precision problem to be beaten with sample size. It is an attribution failure: two causes, one observation, no way to separate them from the data as collected. No amount of n resolves it, because the information required to resolve it was never captured. Randomization does not help — you can randomize subjects, but every arm still breathes through an uncharacterized machine. Blinding does not help — the device does not care whether the investigator knows what it is.

The only fix is upstream, in the apparatus. This is why the remedy is an instrument, not a statistic.

2.3 The device-blind study

We call the current default a device-blind study (Lexicon §2A.4): work that is valid-looking, competently executed, honestly reported — and whose delivery vehicle is an uncontrolled variable.

I want to be careful here, because it would be easy and wrong to read this as an attack on the existing literature. It is not. Device-blind studies have produced real knowledge, and their authors are not at fault for the vacuum they inherited. A device-blind study is not false. It is under-determined: its findings are consistent with more than one explanation, and it lacks the information to choose between them.

That distinction should be stated in the limitations section of a great many papers, and currently is not — usually, I suspect, because the field has not been given the vocabulary to state it. Supplying that vocabulary is part of why this paper exists.

3. What a reference device is

3.1 Not a safe device — an instrument

The remedy is a reference device (Lexicon §2A.3): hardware whose contribution to the inhaled stream has been characterized across its operating envelope and quantified below the attribution floor (§3.2) for a defined analyte panel.

The most important thing to say about a reference device is what it is not. It is not a good device, or a safe device, or a device that passed anything. Those are certification concepts and they belong to a different paper (P-006).

A reference device is a scientific instrument. Its virtue is not that it is harmless but that its own signal is known and bounded, so that signal appearing downstream can be attributed to the material rather than the hardware. It is the vaporizer equivalent of an assay blank, a certified reference material, or a validated delivery system.

A reference device could, in principle, be an unpleasant thing to smoke. That is not the point. The point is that when a researcher measures formaldehyde downstream of it, they know the device did not put it there.

3.2 The attribution floor — and why “zero” is not a thing

Here is where I have to correct my own instinct, because the phrase I reach for naturally is “a device that tests at zero,” and that phrase cannot survive contact with a chemist.

Zero is not a measurable quantity. It is an artifact of an insufficiently sensitive instrument. Every material at temperature emits something; push the detection limit low enough and you will find it. A device that “tests at zero” is a device that was tested with equipment that could not see what it emitted, which is a statement about the equipment, not the device.

So the honest construct is the attribution floor (Lexicon §2A.2): the concentration below which a device’s own contribution cannot be distinguished from background, for a specified analyte, under a specified operating envelope.

A device is never clean. A device is characterized, and quantified below the attribution floor, for a named analyte panel, at a stated limit of quantitation, within a stated envelope. That is a longer sentence and a weaker-sounding claim, and it is the one that will still be standing after review. I state it plainly here because I expect to be tempted by the shorter version for the rest of my life, and I would rather be on record.

This is not pedantry with practical stakes attached — the stakes are practical. “Zero” invites a single number on a marketing page. The attribution floor forces three disclosures — panel, LOQ, envelope — and each of them is a place where a reader can check the work. That is the difference between a certification mark that means something and one that does not.

3.3 The device-qualified study

A device-qualified study (Lexicon §2A.4) is one conducted on a reference device, such that observed effects are attributable to the material under study.

This gives the field something it currently lacks: a name for the thing that is missing, and therefore a thing to ask for. A journal can require it. An IRB can ask about it. A program officer can fund it. A reviewer can note its absence. None of that is possible while the gap has no name — which is the entire theory of the Lexicon.

4. The instrument-first sequence

Everything above collapses into an ordering principle (Lexicon §2A.5):

Characterize the device → establish a reference device → then, and only then, ask what the medicine does to the human.

Each step is a precondition for the next, and the sequence cannot be run out of order. This is a stronger claim than “the steps are best done in this order.” It is that efficacy findings generated on uncharacterized hardware are not merely noisy — they are unattributable, and no downstream cleverness recovers the attribution that was never captured.

This is the answer to a question I get asked in various forms: why start with devices, when what people actually need to know is whether the medicine works?

Because starting anywhere else forecloses the answer. The device is not a detail to be handled later. It is the instrument through which every other measurement is taken. You do not get to the human question by skipping the machine question; you get there through it.

There is a practical corollary worth stating for the institutions that would fund this work. The first step — characterizing devices — requires no access to controlled substances. Off-gas analysis, heavy-metals migration, VOC speciation, and leachables testing are performed on the device: dry-fire, or loaded with a non-controlled surrogate matrix. This is materials chemistry, not drug research. It needs no scheduling registration, no cannabis license, and no human subjects.

Which means the least legally encumbered step in the entire chain is also the one that unblocks all the others. The bottleneck in cannabis efficacy research is a materials-characterization problem wearing a drug-policy costume.

5. What qualification would require

This paper is an argument, not a protocol; CDSI-002 is where the protocol would live. But the argument owes the reader a sketch of what it is asking for, and an honest account of what is unresolved.

5.1 The analyte panel

Reference status is meaningless without a named panel — a device below the floor for carbonyls tells you nothing about the metals it sheds. A first-pass panel, to be justified rather than asserted in the protocol document:

Carbonyls — formaldehyde, acetaldehyde, acrolein. The canonical thermal-degradation markers, and the compounds around which the EVALI-era literature organized itself (P-004).
Heavy metals — lead, nickel, chromium, cadmium, from alloys, solders, and ceramic glazes (Lexicon §4.2).
VOCs and semi-volatiles — the broader device volatilome (Lexicon §2.7), by GC-MS.

The panel must be defended on the basis of what the hardware is known to emit, not what is convenient to measure.

5.2 The operating envelope

A device below the floor at 350 °F may be nowhere near it at 650 °F (Lexicon §2.3). Reference status is therefore not a property of a device; it is a property of a device within a stated range, and the range must cover the conditions the downstream study will actually use. A reference device qualified at three temperatures and cited by a study that ran at a fourth has been misused.

5.3 The unresolved problem: reference status may decay

This is the hard part, and I would rather flag it than finesse it.

The pyrolytic reservoir effect (Lexicon §2.5; P-003) is the observation that residual concentrate chars onto the heating element over successive sessions and re-cooks on later draws, producing an off-gas profile measurably different from a fresh load — and growing with use.

If that is right, then a device below the attribution floor at draw 1 may be above it by draw 500. Reference status would decay with use. And that implies reference status attaches not to a model, and perhaps not even to a unit, but to a unit in a stated draw-history state.

The consequences are not small. A clinical study running for six months on the same hardware may begin device-qualified and end device-blind, without anyone noticing, because nothing about the device visibly changed. Qualification would then need to specify a valid draw count, a re-qualification interval, or a cleaning protocol that provably returns the device to floor — and “provably” is doing a great deal of work in that sentence.

I do not know the answer. I know the question is real, because it falls directly out of a phenomenon we have already named and have already committed to measuring. It probably deserves its own paper once P-003 has data behind it.

5.4 Position within existing metrology

Nothing here should be a new vocabulary if an old one will do. The concepts above map onto established practice, and the mapping should be made explicit rather than reinvented:

Assay blanks and matrix blanks — analytical chemistry’s version of the same move.
Certified reference materials — the closest structural analogue; cannabis has CRMs for the plant and none for the device.
ISO/IEC 17025 method validation — the framework any real qualification would have to live inside.
FDA combination products — the regulatory recognition that a drug plus a delivery device is a single product, and that the device is part of what is being evaluated. Inhaled cannabis is a combination product that has never been evaluated as one.

If this paper’s argument is correct, it is not novel science. It is the application of ordinary metrology to a field that was never given the institutional means to apply it. That is a less thrilling claim, and a more defensible one.

6. What this asks for

Three things, in ascending order of difficulty.

From researchers: state the device. Not the brand — the materials, the temperature, the draw regime, the unit’s history. If it has not been characterized, say so in the limitations, in the language of attribution rather than the language of noise. This costs nothing and it is where the field starts becoming able to see the problem.

From journals, IRBs, and funders: ask whether the delivery device is characterized, and treat “it is a commercially available vaporizer” as the non-answer it is. A reviewer who would never accept an uncharacterized reagent should not accept an uncharacterized device.

From somebody with a bench: build the reference devices. Someone has to do the unglamorous materials work that lets everyone else ask the interesting question. That is the job CDSI was incorporated to do, and it is the reason a hardware institute belongs in a conversation about medicine.

7. Conclusion

The argument of this paper is narrow, and I want to end by keeping it narrow rather than inflating it.

I am not claiming that inhaled cannabis does not work. I am not claiming that existing studies are wrong, or that their authors were careless. I am claiming something more specific and, I think, harder to dismiss: a device-blind study of inhaled cannabis cannot cleanly separate the effect of the medicine from the effect of the machine — and that this is a property of the study design, not of the investigators.

The fix is not more subjects. It is an instrument that does not currently exist: a device whose own contribution is characterized and bounded, so that what happens downstream can be honestly attributed to what was administered.

Building that instrument requires no cannabis, no scheduling registration, and no human subjects. It requires someone to take the machine seriously as a machine.

It all starts with the devices. Not because the devices are the most important thing — the person is the most important thing — but because the device is the only part of this you have to understand first in order to understand any of the rest of it honestly.

Acknowledgments

The argument in this paper is a formalization of a position the author has held informally since commissioning the 2016 ALS Environmental off-gas analysis (service request P1605022), and owes its structure to the observation — obvious in hindsight, and not original to this field — that no one would run an experiment through an uncalibrated instrument. This draft is open for scientific review, and §5.3 in particular is offered as an unsolved problem rather than a result.

References (preliminary)

To be expanded in v1.0. This draft is an argument; the citation apparatus required for journal submission is not yet built, and the claims in §5.4 in particular require sourcing to the specific standards named.

Cannabis Device Safety Institute. The CDSI Lexicon, v. 2026-07, §2A. 2026.
Macosko, M. The Hardware Vacuum. CDSI Working Paper P-001. 2026.
Macosko, M. Loaded-State Off-Gas: A Proposed Methodology for Cannabis Concentrate Vaporizer Characterization. CDSI Working Paper P-002. 2026.
Macosko, M. The Pyrolytic Reservoir: Why Single-Draw Testing Underestimates Real-World Hardware Risk. CDSI Working Paper P-003. 2026.
Macosko, M. Adulterant Amplification: How EVALI Could Have Been Caught. CDSI Working Paper P-004. 2026.
ALS Environmental. Off-Gas Analysis Report, Service Request P1605022. November 30, 2016. Archived at CDSI document repository.
Cannabis Device Safety Institute. CDSI-001 Protocol: Off-Gas Analysis of Cannabis Concentrate Vaporizers, v1.0 (Founding Draft). April 2026.
International Organization for Standardization. ISO/IEC 17025:2017 — General requirements for the competence of testing and calibration laboratories. 2017.
U.S. Food and Drug Administration. Combination Products Guidance. [citation to be completed]

Comments and corrections to: matt@ineedhemp.com (until papers@cdsi.click is provisioned).