The Attribution Chain

The Four Characterizations That Precede an Honest Human Answer

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 Attribution Chain. CDSI Working Paper P-012. 2026.

Abstract

P-011 argued that hardware characterization is a precondition for valid research into inhaled cannabis, because an uncharacterized device makes every downstream measurement unattributable. This paper extends that argument to its end. If the goal is a defensible answer to does this help this person, then the device is the first of four characterizations that have to happen, in order: (1) the device alone, (2) the material alone, (3) the two together under heat, and (4) the human outcome. We call this the attribution chain, and we argue it has a property that makes it harder to satisfy than it first appears: Link 3 is not the sum of Links 1 and 2. Heating a characterized material on a characterized device produces compounds attributable to neither in isolation — interaction products — of which the already-named phenomenon of adulterant amplification is one documented special case. Because the system is non-additive, the interaction cannot be inferred by subtracting two independent characterizations; it must be measured directly, as a system, under realistic operating conditions. We describe each link, state what it can and cannot establish, and name the regulatory gate that separates it from the next — a gradient running from no license required at Link 1 to human-subjects oversight at Link 4. We argue that the chain is only as attributable as its weakest link, that the field has been attempting Link 4 with Links 1 through 3 open, and that the institutional consequence is unglamorous: someone has to fund and staff the upstream links before the downstream question can be honestly asked.

Keywords: cannabis, inhalation, attribution, confounding, thermal degradation, interaction products, non-additivity, research infrastructure

1. Introduction

I want to begin by repeating a question from P-011, because this paper is only an attempt to follow it further than that one did.

The question is does this help? Does inhaled cannabis reduce this person’s pain, or their nausea, or their seizures. That is the question patients ask. It is the question that justifies most of the public money that has ever gone into this field. And it is a medical question, which I am not qualified to answer. I build vaporizers.

P-011 made a narrow argument about that question: you cannot answer it on an uncharacterized device, because the device contributes its own chemistry to the same stream, and so every measured outcome has at least two possible sources. That is device confounding (Lexicon §2A.1), and it is an attribution failure rather than a statistical one — no sample size touches it.

What that paper left implicit, and what I want to make explicit here, is that the device is not the only thing standing between the question and the answer. It is the first thing. There are four.

You have to know what the machine does. You have to know what the medicine is. You have to know what happens when the machine heats the medicine — which, and this is the part I think is genuinely underappreciated, is not knowable from the first two. And then, with those three closed, you can put it in a person and believe what you measure.

That ordering is what I am calling the attribution chain. This paper describes its four links, argues that they cannot be run out of order or skipped, identifies the property that makes the third link irreducible, and names the regulatory gate standing at each step — because those gates explain a great deal about why the chain has never been completed by anyone.

From where I sit, the chain is the most useful thing I have to offer this field, and it is not a discovery. It is bookkeeping. It is the observation that four things have to be known and only some of them are.

2. The chain

2.1 Link 1 — The device alone

What it measures: the chemistry the hardware itself contributes at operating temperature. The heating element, its alloy or ceramic or quartz, its glaze, its binder, its solder joints, its wick, its airpath, its adhesives — each at temperature, each doing chemistry of its own (Lexicon §1.1). Measured dry-fire, and with a non-controlled surrogate matrix to supply realistic thermal mass and airflow.

What it establishes: the attribution floor (Lexicon §2A.2) for a named analyte panel across a stated operating envelope, and — if the device is quantified below that floor — a reference device (§2A.3): an instrument whose own signal is known and bounded.

What it cannot establish: anything about cannabis. A reference device is not a safe device and not a tested product. It is a blank.

The gate: none. This is the link that matters most institutionally, and the one people find hardest to believe. Off-gas analysis, metals migration, VOC speciation, and leachables testing on a device require no scheduling registration, no cannabis license, and no human subjects. It is materials chemistry. As P-011 put it, the bottleneck in cannabis efficacy research is a materials-characterization problem wearing a drug-policy costume.

2.2 Link 2 — The material alone

What it measures: everything that went in. Cannabinoid profile, the full terpene profile, and the contaminant panel — pesticides, residual solvents, heavy metals, microbials, and any adulterant or cutting agent present.

What it establishes: the composition of the input. Without it you cannot say what was administered even in principle, because “cannabis” is not a specification — it is a category containing an enormous range of actual chemistry.

What it cannot establish: what any of that becomes at 650 °F. A terpene profile is a description of the material at room temperature. It is not a prediction of the aerosol.

The gate: a cannabis license. This link touches the plant. In California that means a testing-laboratory license from the Department of Cannabis Control and ISO/IEC 17025 accreditation for the methods used. This is the first hard gate in the chain, and it is where CDSI’s current charter stops: the Institute’s organizing documents prohibit it from handling cannabis at all. Link 2 therefore requires either a licensed subsidiary or a licensed partner — a structural decision, not a purchase.

2.3 Link 3 — The two together, under heat

What it measures: what actually comes off when the characterized material is loaded onto the characterized device and heated under realistic use — the loaded-state off-gas (P-002), the thermal-degradation and pyrolysis products, the aerosol as inhaled rather than as formulated.

What it establishes: the thing that was actually administered. Not the input. Not the apparatus. The output of the system.

What it cannot establish: what the output does to a person.

The gate: both of the above, plus a bench. Link 3 needs the license from Link 2 and the emissions-capture rig from Link 1, operating together. It is the most demanding link that does not involve a human being, and it is the one nobody owns.

This is the link the 2016 ALS Environmental analysis (P1605022) reached for on a single device, and it is the reason that test remains this Institute’s founding artifact — not because it was comprehensive, but because it asked the right question and then nothing followed it for a decade.

2.4 Link 4 — The human

What it measures: the outcome. Relief, harm, a compound in the blood, a change in a symptom score.

What it establishes: the answer — but only if Links 1 through 3 are closed. Otherwise it establishes that something happened when a person inhaled something.

The gate: the highest one. Human-subjects research requires IRB review and informed consent. Research administering a controlled substance to human subjects carries further federal requirements — DEA registration and, where the work is investigating a drug’s effects, FDA involvement. The specifics are in motion: the partial rescheduling that took effect in April 2026 moved FDA-approved and state-licensed medical marijuana products to Schedule III, while other categories remain Schedule I, and the practical consequences for research protocols are still being worked out. Anyone approaching this link needs counsel, not a working paper’s summary. What matters here is only the shape: this is the most encumbered link in the chain and the least likely to be run by a hardware institute.

I want to be direct about what that implies for CDSI, because it would be easy to let the ambition of this paper outrun the honesty of it. I do not expect this Institute to run Link 4. The stronger role — and the one consistent with everything CDSI is structurally built to be — is to supply Links 1 through 3 as the instrument layer, so that somebody else’s Link 4, run by people qualified to run it, is finally interpretable. The clinic does not need us in the room. It needs the machine in the room to be known.

3. Why the order is forced

The links are not four topics that happen to be related. Each one supplies a term that the next one requires.

Link 4 asks what the administered aerosol did to the person. It cannot be answered without Link 3, which says what the aerosol was. Link 3 cannot be interpreted without Links 1 and 2, which say which parts of that aerosol came from the machine and which from the material. And Link 1 is answerable on its own — it depends on nothing upstream, which is exactly why it is where the chain has to start.

Run the chain backwards and each step dissolves into the one before it. Run it forwards and each step is tractable. That asymmetry is the whole argument, and it is the reason the instrument-first sequence (Lexicon §2A.5) is stated as an ordering principle rather than a preference.

It also produces a rule worth stating plainly, because it is the practical form of everything here:

A study is only as attributable as the weakest link in its chain.

A perfectly designed clinical trial — randomized, blinded, adequately powered, competently analyzed — that runs on an uncharacterized device with an uncharacterized product has a chain that is open at three of four links. Its statistics are impeccable. Its attribution is absent. The field has been attempting Link 4 with the upstream links open, not out of carelessness, but because of the institutional vacuum described in P-001: nobody was ever responsible for the upstream links, so nobody built them.

4. The non-additivity problem

This is the part of the paper I would most like to be wrong about, and I do not think I am.

4.1 The subtraction that does not work

Here is the natural shortcut, and I want to name it because it is what a careful reader of P-011 will reach for first. If you have characterized the device (Link 1) and characterized the material (Link 2), why do you need Link 3 at all? Measure the aerosol, subtract the device’s known contribution, subtract the material’s known contribution, and whatever remains is a rounding error. Two characterizations, one arithmetic step, chain closed.

That would work if the system were additive — if the compounds arriving at the mouthpiece were the compounds that were already in the device and the material, merely relocated by heat.

They are not. Heat is not a transport mechanism. It is a reagent.

4.2 Interaction products

At operating temperature, the material and the hardware are not two inventories sitting next to each other. They are a reaction vessel. The material decomposes, and its decomposition products contact hot metal, hot ceramic, hot glaze, and hot residue — and react there. The device’s surfaces are not inert bystanders; some of them are catalytic, and all of them are hot.

The consequence is that the inhaled stream contains compounds attributable to neither input in isolation. We name this class interaction products (Lexicon §2B.1): compounds present in the inhaled aerosol that are produced by the reaction of the material with the device at operating temperature, and are therefore absent from a characterization of either one alone.

This is not a speculative category. CDSI has already named one of its special cases: adulterant amplification (Lexicon §4.3) is precisely the observation that an adulterant in the concentrate — vitamin E acetate, a glycol, an untested terpene additive — becomes a different and more dangerous thing when it reacts with hot heater material, and that the resulting harm is device-mediated. The adulterant alone does not predict it. The device alone does not predict it. The pairing does. Adulterant amplification is what interaction chemistry looks like when it goes badly enough to make the news; the general phenomenon does not require an adulterant to be present.

The pyrolytic reservoir effect (Lexicon §2.5; P-003) is a second instance, and a nastier one, because it means the interaction has a memory. Residue chars onto the element and re-cooks on later draws, so the reaction vessel is not the same vessel it was two hundred draws ago. The interaction term is not even constant for a given device-plus-material pairing — it drifts with use.

4.3 What follows

If interaction products exist, three things follow, and none of them are convenient.

First, Link 3 must be measured, not derived. The subtraction in §4.1 fails because the quantity it is trying to isolate was generated by the measurement conditions themselves. There is no arithmetic that recovers it from two independent characterizations. The loaded, heated system has to be sampled as a system — which is exactly what loaded-state off-gas analysis (P-002) exists to do, and why CDSI-001 specifies concentrate present, at temperature, under realistic draw regimes rather than a dry-fire proxy.

Second, Links 1 and 2 do not become redundant — they become the only way to read Link 3. This is the point I want to be careful not to overstate in the other direction. If you only run Link 3, you have measured the aerosol from a black box: you know what came out and you cannot say what put it there. The upstream links are what let you partition the Link 3 result into “the machine did this,” “the material did this,” and “these two made this together.” Link 3 without Links 1 and 2 is a measurement without an interpretation. Links 1 and 2 without Link 3 are an interpretation without a measurement.

Third, the interaction term is pairwise. Characterizing device A and material X does not characterize device A with material Y. If the interaction is real, then some of what a study measures belongs to the combination, and combinations do not inherit each other’s results. How badly this bites is an empirical question and an important one — if interaction products turn out to be modest and predictable across common pairings, the chain gets much cheaper to run; if they are large and idiosyncratic, reference status may have to be stated per pairing. I do not know which world we are in. Nobody does, because the measurement has not been made systematically. That is not a rhetorical point. It is the actual reason to build the bench.

5. What this does not claim

The argument is easy to inflate, so I want to bound it explicitly, as P-011 did.

It does not claim inhaled cannabis is dangerous. Nothing here says the interaction term is large, or harmful, or even clinically detectable. It says it is unmeasured, and that unmeasured is not the same as zero — a distinction this Institute has committed to in the strongest terms it knows how (Lexicon §2A.2).

It does not claim the existing literature is false. Device-blind studies (Lexicon §2A.4) have produced real knowledge. They are not wrong; they are under-determined — consistent with more than one explanation, and lacking the information to choose. That belongs in a limitations section, in the language of attribution rather than the language of noise. Most of the field has not been given the vocabulary to write that sentence. Supplying it is most of what these papers are for.

It does not claim novelty. None of this is new science. Characterizing your apparatus, characterizing your inputs, and measuring your system as a system rather than inferring it from parts is ordinary metrology. The claim is that a field was never given the institutional means to apply the ordinary thing — which is a less thrilling assertion and a more defensible one.

It does not claim CDSI should run the whole chain. See §2.4. The Institute is built to own Links 1 through 3 and to hand Link 4 to people qualified for it.

6. The regulatory gradient

Lay the four gates side by side and a pattern appears that I find genuinely clarifying.

The gradient runs the same direction as the chain. The link with no legal gate at all is the link everything else depends on — and it is the one that has never been systematically built.

That is the whole tragedy of the hardware vacuum in one row of a table. The unblocking step was always available. It required no policy change, no rescheduling, no act of Congress, no permission from anybody. It required a bench, a method, and somebody whose job it was. There was no such body, so there was no such job, so the easiest link in the chain is the one still missing — and its absence propagates all the way to the patient.

I find that maddening and I also find it encouraging, because a missing thing that requires no one’s permission is a thing you can go build.

7. The institutional path

A working paper is not a business plan, and I am not going to pretend the last three sections were secretly about money. But the chain does have institutional consequences, and a reader deciding whether this is worth funding or joining deserves to see them stated rather than implied.

Link 1 is buildable now, by CDSI, under its existing charter. It needs a thermal and emissions bench: controlled temperature, controlled airflow, sorbent capture, and analysis — initially at accredited outside laboratories, which is the pay-the-lab discipline (Lexicon §3.1) and also, conveniently, the cheapest way to start. This produces publishable data immediately and it produces the first reference devices. It requires no license and no partner.

Link 2 requires a decision, not a purchase. CDSI cannot touch cannabis. Closing Link 2 means either taking the authorized-but-unformed testing subsidiary all the way to a testing-laboratory license and ISO/IEC 17025 accreditation, or partnering with a laboratory that already holds both. Both routes are real; they are different organizations at the end of them, and that choice belongs to a board rather than a paper.

Link 3 is the Institute’s actual destination. It is where the founding artifact pointed in 2016 and it is the link nobody owns. It needs Links 1 and 2 in the same building, or in two buildings with a chain of custody between them.

Link 4 is somebody else’s, and should be. The contribution CDSI can make to Link 4 is to hand its investigators an instrument and a characterized dose, so that the study they were going to run anyway becomes a device-qualified study (Lexicon §2A.4) instead of a device-blind one.

Read in order, that is a sequence in which each link earns the credibility that makes the next one fundable. It is also, I notice, the same order the science requires. I do not think that is a coincidence. Work that has to be done first tends to be work that is worth doing first.

8. Unresolved problems

Offered as open questions rather than results, in the manner of P-011 §5.3.

How large is the interaction term? Everything in §4 establishes that interaction products exist as a class and must be measured directly. None of it says how much they matter. This is the first empirical question a bench should be pointed at, and the answer determines whether the chain is expensive or ruinous to run.

Does the interaction term generalize across pairings? If device-plus-material interactions are broadly similar across common hardware and common formulations, reference status can attach to a device. If they are idiosyncratic, qualification may have to be stated per pairing — a combinatorial problem that would change what this Institute is for.

Does reference status decay? Carried forward unresolved from P-011 §5.3 and sharpened here: if the pyrolytic reservoir gives the interaction a memory, then a device qualified at draw 1 may be unqualified at draw 500, and a six-month clinical study could begin device-qualified and end device-blind with nothing visibly changing.

What is the right surrogate matrix for Link 1? Characterizing a device under realistic thermal and flow conditions without cannabis requires a non-controlled stand-in with defensible thermal behavior. Choosing it badly imports an interaction term of its own — the surrogate’s chemistry with the hot element — into what is supposed to be a device blank. This is a real methodological problem sitting inside the one link everyone can legally run, and I do not think it has been solved.

9. Conclusion

The chain is four links long. The device, the material, the two of them under heat, and then the person.

Everyone wants the fourth link. The fourth link is where the patient is, and it is the only one that answers the question anybody actually asked. But it is also the only link that cannot be run first, and the field has spent a great deal of money and care attempting it with the upstream links open — which is why so much of what we know about inhaled cannabis is under-determined rather than wrong.

The first link needs no permission from anyone. That is the strangest fact in this paper, and the one I would most like a reader to leave with. The step that unblocks the rest of the chain has always been available, has never been anybody’s job, and therefore has never been done.

It all starts with the devices. Not because the devices matter most — the person matters most, and I would like that on the record in every paper this Institute publishes. But because the machine is the part you have to understand first in order to understand any of the rest of it honestly, and someone has to go stand upstream and do the unglamorous part.

That is the job. It is available. We took it.

Acknowledgments

This paper is a formalization of an argument made informally by the author and sharpened in conversation during July 2026; it extends the instrument-first sequence of P-011 into the full chain and owes its central claim — the non-additivity of Link 3 — to a phenomenon this Institute had already named without noticing it was the general case. The 2016 ALS Environmental off-gas analysis (service request P1605022) was an attempt at Link 3 with Links 1 and 2 open, which is both why it was worth commissioning and why it was not sufficient. This draft is open for scientific review. §8 is offered as a list of things the author does not know.

Conflict of interest disclosure: The author founded and holds ownership interests in cannabis-hardware businesses (Divine Tribe / Nice Dreamz LLC). CDSI discloses this interest on the cover page of every document it publishes, including this one. The Institute’s Conflict of Interest Policy (Article VIII) governs.