Research Brief — R03: Legal & Illicit Cannabis Production

Report ID: R03 (Wave 1, closing the macro foundation) · Brief version: v1.0 — FROZEN September 1, 2026 on Danny’s confirmation. Both research runs execute this version and must cite it; the critique stage records it. Date drafted: August 31, 2026 (v0.95, Sept 1: ChatGPT editorial review folded in — reconstruction-chain standard, observation→estimate transformation field, measurement-asymmetry requirement, causal-evidence ladder, data-vs-inference-vs-motivation separation, practitioner-layer boundary sharpened; v1.0, Sept 1: three consistency/pre-judgment fixes from ChatGPT’s read of the full draft — §B reference, incentives-as-context rewording, rung-1 expectation softened) · Drafted by: Claude (editor) · Protocol: v1.3

Methodological spine (one sentence): define the populations → identify how each population becomes observable → reconstruct the estimate provenance → grade the transformation assumptions → compare only commensurable quantities → evaluate policy causality separately. Assigned to: ChatGPT deep research + Gemini deep research (independently, same frozen brief; both runs cite the frozen version)

Assignment

Produce an evidence review of the legal/illicit structure of cannabis production: how cultivation is organized across the licit–illicit spectrum, what is actually known (versus estimated, versus asserted) about illicit production and its persistence after legalization, what drives growers into or out of licensed systems, and what harms — of illicit production and of enforcement — the evidence actually documents. This is not a policy-advocacy piece in either direction: the goal is to establish what the record supports, where it is estimate-lore, and where the honest answer is unknown.

Keep three claims-classes strictly separate throughout:

1. Documented structure — who grows, where, and under what legal status (administrative records, surveys, ethnography, prosecutions).
2. Measured magnitudes — hectares, tonnage, market shares, harm rates. These are rare; R02-C002 already established that no defensible global production figure exists, and the same epistemics apply at national and state levels.
3. Causal claims — “legalization reduces X,” “enforcement suppresses Y,” “taxes drive Z.” These require causal evidence, not before/after narration.

Grower discourse, journalism, and policy advocacy collapse these constantly; this report must not.

Methodology requirements

Estimate-provenance discipline (the methodological spine — extends R02-C002/C005/C014). Illicit production is, by definition, not administratively measured. Every quantitative claim about it must state how the number was made: eradication/seizure extrapolation, consumption back-calculation, wastewater analysis, remote sensing, grower/consumer survey, expert elicitation, or administrative residual (total demand minus legal sales). Each method carries a named bias — enforcement statistics measure enforcement activity, not production (R02’s “denominator laundering” finding); seizures are not supply; license counts are not output (R02-C022); legal-capture percentages of spending are not percentages of farms (R02-C030 note). A precise-looking illicit-market figure with unreconstructable provenance is estimate-lore and must be labeled as such (failure-mode #4; the Paraguay precedent, R02-C014). Unknown is an acceptable answer and often the correct one.

The reconstruction-chain standard. Every major illicit-production number carries its full chain: who estimated it, what raw observations it used, what assumptions converted observations into production, what geography/population/period it covered, what uncertainty was reported, and whether later sources independently re-derived it or merely repeated it. Ranking rule: a precise number with an opaque origin ranks below a wide but reconstructable estimate. Expect to find that many supposedly independent market-size figures descend from a small number of shared transformations and inherited assumptions — mapping that descent is itself a deliverable of §B, Data epistemology.

Measurement asymmetry — production variables vs. visibility-to-the-state variables. Legal markets generate production-side data (licensing, track-and-trace inventory, taxation, testing, canopy, sales, destruction records). Illicit markets are observed only through detection systems. Keep the two variable classes explicitly separate: a change in seizures or eradicated plants can reflect more production, more enforcement effort, better detection, geographic displacement, reporting changes, or any combination — never treat visibility metrics as production metrics without an argued bridge.

Causal-evidence ladder (pre-registered). For every policy-effect claim (“legalization reduced illegal cultivation,” “taxes caused illicit persistence,” “enforcement displaced production”), classify the best available evidence on a three-rung ladder — (1) descriptive association / before-after narration; (2) controlled observational or quasi-experimental designs; (3) credible causal inference (natural experiments, difference-in-differences with defensible counterfactuals) — and make the prose match the rung actually achieved. Expect many circulating policy claims to resolve only to rung 1; classify each independently and say so where they do.

Legal status is a regulatory category, not a cultivation system. Do not attribute cultivation practices, product quality, or environmental behavior to legality itself (R01-C033 discipline). “Illicit” spans at least: legacy smallholders declining or excluded from licensing; unlicensed commercial warehouse operations; trespass/public-land grows; house grows in prohibition states; licensed-operator diversion; home grows exceeding plant limits (gray zone); cross-border production; informal cultivation in weak-enforcement jurisdictions; and traditional producing regions under prohibition (Rif, Bekaa, Cauca, Paraguay — R02-C010/C012/C014/C015). Complete this typology before any synthesis — it is the frame every subsequent claim hangs on. These populations differ in practices, harms, and policy response; claims must name which population they cover, and evidence about one does not transfer to another.

Symmetry without false symmetry. Drug-war-era, legalization-advocacy, and industry sources can each carry institutional incentives or framing pressures that favor particular narratives about illicit scale, harms, legal-market success, enforcement harms, or illicit competition. Record those incentives as context, then independently evaluate data quality and inferential quality. Label the institutional position of every load-bearing source — but separate three things that affiliation alone does not settle: data quality, inferential quality, and rhetorical motivation. An enforcement agency can hold excellent primary operational data inside institutional framing; an advocacy organization can surface genuinely neglected populations while selecting evidence strategically. Affiliation is context, never validation or disqualification by itself. Harm claims in both directions require denominators and comparators — “trespass grows poison wildlife” (documented: R02-C030) and “enforcement devastated communities” both meet the same evidentiary standard.

Jurisdiction- and date-stamping. This domain moves fast (rescheduling processes, new state markets, Thailand’s reversals, Germany’s CanG). Every regulatory and market claim carries its jurisdiction and as-of date. Prefer primary administrative sources (state cannabis agencies, Health Canada, EMCDDA/EUDA, UNODC — with UNODC’s own method caveats per R02) over journalism, and journalism over advocacy.

Causal humility on policy effects. Legalization arrived simultaneously with price collapse (R02-C025), interstate dynamics, COVID, and enforcement changes. Before/after comparisons are not causal identification. Where genuine causal work exists (difference-in-designs, credible natural experiments), rank it above narrative; where it doesn’t, say the causal question is open.

Provenance for quantitative estimates — 11 fields for R03. Every major quantitative estimate carries the standard 10-field provenance record (geography, period, population, metric, functional unit, system, product boundary, method, directional bias, confidence) plus an eleventh field for this report: “observation → estimate transformation” — the explicit arithmetic that turned raw observations into the number (e.g., “eradicated plants × assumed mature-plant yield × assumed detection fraction”). If the transformation cannot be stated, the estimate is by definition unreconstructable and is labeled estimate-lore.

Key questions

A. Taxonomy & structure

1. Build the operational taxonomy of illicit-spectrum production above, with the best available structural evidence for each population (who, where, scale class, typical practices). Name which are documented vs. inferred.
2. What does the record actually establish about organized-crime involvement in North American production (trespass networks, labor trafficking prosecutions, “cartel grow” claims) — distinguishing prosecuted cases from asserted patterns?

B. Data epistemology (required output category)

3. Catalog the estimation methods used for illicit cultivation/production and their documented biases. Which jurisdictions have any defensible illicit-scale estimate, by what method? Where do widely cited figures (e.g., “X% of California cannabis is illicit,” national illicit-market dollar sizes) actually come from — trace 3–5 of the most-circulated numbers to their origins.
4. What happened to survey-based crop monitoring (UNODC Morocco/Afghanistan-style) for cannabis, and why does nothing comparable exist for the Americas’ herb production (extend R02-C035)?

C. The legal transition

5. What is documented about who enters licensed systems and who doesn’t: compliance cost structures, local bans, license caps, taxes, testing costs, capital requirements — anchored on California’s contraction (R02-C027/C030) and compared across at least Canada, Oregon/Colorado/Oklahoma-style variants, and one or two non-US arcs (Thailand, Germany, Uruguay).
6. Legacy-grower outcomes specifically: what do surveys and ethnographies (e.g., the Berkeley Cannabis Research Center work already verified in R02, S-085–S-087) establish about barriers, participation, and exit? Equity-program outcomes where measured.
7. What share of consumption/spending do legal markets capture, by jurisdiction and method — kept strictly separate from shares of production (the Canada ~70–80% spending figure is the calibration example, R02-C030 note)?

D. Illicit persistence & interaction

8. What does evidence support about why illicit production persists where it does: price/tax wedges, prohibition-state demand (export arbitrage from legal states), municipal bans, testing/compliance costs, cultural continuity? Rank documented mechanisms above asserted ones.
9. How do legal and illicit sectors interact: diversion from licensed grows (documented cases), burner licenses, inversion (illicit product entering legal channels), and enforcement displacement effects. What is prosecuted fact vs. industry lore?
10. Does enforcement measurably suppress production? What does the eradication-era record actually show about supply effects (as opposed to activity counts)?

E. Harms — both directions (required output category)

11. Environmental: trespass-grow toxicology (the verified Gabriel cluster, S-088; banned carbamates/anticoagulant rodenticides; water diversion R05-C011), with denominators where possible (what fraction of illicit production do trespass grows represent — if unknown, say so). Do NOT let trespass-grow evidence characterize all unlicensed production (population discipline, §Methodology).
12. Product safety: what does testing of illicit vs. licensed product actually show (pesticides such as myclobutanil, molds, heavy metals) — and what are the licensed market’s own documented failures (lab shopping, THC inflation, recalls)? Neither market gets a halo.
13. Human costs: documented labor exploitation/trafficking in illicit grows; and, symmetrically, documented enforcement harms (criminalization patterns, racial disparities in cannabis enforcement, asset forfeiture) — each at the same evidentiary standard.
14. Failure modes of legalization design: which regulatory choices (tax levels, license caps, vertical integration rules, home-grow allowances) have documented effects on illicit persistence, and which claimed effects remain unmeasured?

F. Synthesis targets

15. The three-class map: documented structure / measured magnitudes / open causal questions — and the conditions under which legal markets do or don’t absorb illicit production. Absence of a defensible estimate is itself a substantive finding (the R02-C002 lesson, held aggressively): do not manufacture an illicit/legal production ratio because readers expect one. “Not defensibly estimable from available evidence” is a better answer than false precision, for any jurisdiction and period where the reconstruction chain fails.
16. The 5–10 most repeated public claims on this topic and what the evidence says about each (see pre-registered myths).

Pre-registered myth checks

Word all verdicts so absence of evidence remains absence of evidence (“unsupported” ≠ “the opposite is true”; unknown is acceptable):

• M1. “Legalization eliminates the illicit market” — and the inverse, “legalization did nothing to the illicit market.”
• M2. “X% of cannabis sold in [state] is still illicit” — trace the circulating percentages to their actual method; most are suspected estimate-lore.
• M3. “Illicit cannabis is commonly laced (fentanyl etc.).”
• M4. “Legal means tested means safe” — evaluate against documented lab-shopping, potency inflation, and recall records.
• M5. “Cartels run most illegal US grows.”
• M6. “Illicit growers are environmental villains” / “legacy growers are land stewards” — both tribal framings; the evidence covers populations, not the moral categories.
• M7. “Eradication and enforcement reduce supply.”
• M8. “Seizure and eradication totals track production trends.”
• M9. “Home-grow allowances feed the black market.”

A practitioner-record passage is withheld from the public copy.

Standing source & citation requirements (PROTOCOL v1.3 — verbatim)

• Every citation supporting a substantive claim must carry a resolvable DOI, stable URL, or independently retrievable identifier — an unresolvable citation makes the claim unsourced.
• Citation-identity integrity: for every substantive source, the authors, title, publication year, venue, volume/pages where applicable, and DOI or stable identifier must resolve to the same bibliographic object — a DOI that resolves to a different real paper fails identity verification. If you cannot fully resolve the tuple, flag it explicitly as a partial citation rather than completing it from memory.
• Exact claim support: a source supporting the general direction of a finding does not automatically support its stated magnitude, population, geography, time period, or causal wording — preserve these distinctions in candidate claims.
• Corrections inherit the complete verification burden: a replacement citation or corrected statistic remains unverified until its existence, identity, and claim support are checked.
• When fact-checking public claims, “the myth is unsupported” and “the opposite is true” are different findings — do not overcorrect; unknown is an acceptable verdict.
• For each major claim, tag the population it covers (which illicit-spectrum category, which jurisdiction, which period) and the estimation method behind any magnitude.
• Report ranges and disagreements between sources, not averaged-away conclusions; list data gaps explicitly.

Candidate claims for editorial review

Conclude with 15–30 concise, testable candidate claims with supporting citation(s), each carrying its population/jurisdiction/period tags and, for magnitudes, the estimation method. Do not assign final evidence status or confidence; the editor will.

Search-stopping rationale

End with an explicit statement of why searching concluded: databases/queries covered, saturation observed, and named question areas where searching stopped because the literature is genuinely absent — distinguishing “not searched” from “searched and empty.”

What “done” looks like

A structured report answering A–F with cited, population-tagged evidence; the data-epistemology section (B) and both-directions harms section (E) as distinct outputs; calibrated myth verdicts; a top-10–15 strongest-sources list with one-line reliability and institutional-position notes; an explicit data-gap list; candidate claims; and the search-stopping rationale. Uncertainty findings are as valuable as positive findings.

Editor’s note (repo workflow): outputs land in raw-inputs/chatgpt-YYYY-MM-DD.md and raw-inputs/gemini-YYYY-MM-DD.md; then verification & critique (critique.md), claim extraction (ledger.md, IDs R03-C001+), synthesis, and content extraction per PROTOCOL v1.3. Corpus anchors: R02-C002 (no defensible global figure — the epistemic keystone), R02-C005 (enforcement-data epistemology), R02-C014 (Paraguay estimate-lore precedent), R02-C015 (Colombia dual structure), R02-C022 (licenses ≠ output), R02-C025 + R05-C029 (price collapse/margin compression), R02-C027 (California license contraction), R02-C030 (legalization nowhere eliminated illicit cultivation; trespass-grow toxicology via S-088), R02-C035 (largest-producer claims), R05-C011 (outdoor environmental impacts are site-specific).