Dank by Definition, Narrowed.

There is a version of cannabis criticism that only knows how to find fraud, and it would be useless here. Dank by Definition publishes hundreds of certificates of analysis. They come from an ISO-accredited laboratory. They carry Metrc package identifiers, manifest numbers, harvest lots and revision histories. We pulled twenty-seven of them, transcribed every cannabinoid and every terpene by hand, and ran the arithmetic on all of it.

Twenty-seven for twenty-seven. Not one number on one certificate fails to reconcile with the numbers beside it. That is the best result this series has recorded, and it is better than the brand it is about to be unfavourably compared with.

The problem is not what is on the certificates. It is what stopped being on them.

What it is, and who actually holds the licence

Dank by Definition is a brand, not a licensee. We swept New York's licensee register for every entity and trade name containing "Dank" — ten rows come back, and none of them is this company. "Definition" returns nothing at all.

The licensee is HPI, in Poughkeepsie, and the relationship is the healthy kind: Edward Kirkham, named on the brand's own About page as a founder and owner, is the primary contact of record on every HPI licence. This is not a rented licence. It is a house brand of the company that makes it — the documented opposite of the pattern New York spent 2025 prosecuting.

NY OCM licensee dataset (data.ny.gov, jskf-tt3q), queried 15 August 2026. Edward Kirkham is primary contact on all four. The three HPI Cultivation licences carry the SEE category "Women-Owned Business, Distressed Farmer."

A note on that last column, stated carefully. New York's social and economic equity certification requires that "sole control of the applicant" be held by persons matching the qualifying category. The register publishes a primary contact, which is an administrative field and not an ownership field, and New York does not publish the officers or shareholders of a business corporation. We therefore cannot tell you who controls HPI Cultivation Inc., and we are not going to imply anything from a name. We note the certification exists, note what it requires, and note that the record is silent on the rest.

Three licence numbers, one of which is dead

Every certificate in our set — all twenty-seven, from August 2023 to April 2026 — prints the same licensee: HPI Canna Inc, 886 Noxon Road, Poughkeepsie. The licence number underneath it changes three times.

Neither of these is fraud and we want to say so plainly. The AUCP series was the 2022 conditional processor programme; those licences lapsed and their holders moved to annual licences, which is exactly what HPI did in December 2024. "HPI Canna" and "HPI Cultivation" are near-certainly the same operation under two corporate names at one address.

But a certificate of analysis is a document whose entire function is to let a stranger check something. Eight of the certificates currently published as proof of testing carry a licence number that resolves to nothing in the state's own database. The same dead number is on the company's website today: "HPI Canna, License #OCM-AUCP-22-000022, is your comprehensive partner in plant-to-product cannabis manufacturing." A reader who does the thing the document invites them to do — look it up — finds an empty result and has no way to know why.

The arithmetic, which is impeccable

The standing test is Total THC = 0.877 × THCa + Δ9. Across twenty-seven batches spanning three certificate templates and nearly three years, the largest deviation we found is 0.86%, and twenty of the twenty-seven land inside a hundredth of a percentage point.

The apparent outliers are not outliers. On the 2023 and early-2024 template, Total THC also includes Δ8-THC — add Floor 13's 0.1349% Δ8 to the calculation and the printed 26.8236% reproduces to four decimal places. Same for Boss Drank. The laboratory changed its formula and the certificates are internally consistent under whichever one applied at the time. We mention it because a reader running the check themselves on an older certificate will get a mismatch and should know why.

Twenty-four current-template batches, January 2025 – April 2026. Three legacy-template batches (Floor 13, Boss Drank, Velvet Vampire) are excluded from ratio comparison because their panel width differs. Computed column is ours. Laboratory: Kaycha Labs, Albany NY, OCMCPL-24-00006-L1.

Where the terpenes go when the THC goes up

Across the whole set there is no inversion — the correlation between THC and total terpenes is r = −0.15, which on twenty-four samples is nothing. Median ratio 1:17.7, comparable to the farm we praised in Volume 11.

The average conceals the shape. Sort by potency and the pattern arrives at the very top:

Three of the four highest-potency cultivars occupy three of the four worst terpene ratios in the catalogue. Gas Berry, the strongest lot at 34.09% THC, is second-worst on aromatics. Baby Yoda, second-strongest, is worst. The cuts a shopper is steered toward by the number on the jar are the ones carrying the least of everything else.

Meanwhile the best ratios in the set belong to cultivars nobody is putting on a billboard: Permanent Marker at 1:10.5 with 2.59% total terpenes, Afghani at 1:11.5 with 2.57%. Those are genuinely good numbers — better than most flower we have measured anywhere. They sit at 27–29% THC, which in this market reads as unremarkable.

Two cultivars, one fingerprint

Something else surfaced when we plotted the individual terpenes rather than the totals. Gas Berry and Baby Yoda — different cultivars, different lots, sampled twelve days apart — return terpene profiles that are, to the precision the certificate reports, the same flower.

Treated as vectors, those two profiles have a cosine similarity of 0.998. Across all 253 cultivar pairs in the current-template set, thirty-nine pairs exceed 0.95 — LA Kush Cake and Gushers reach 0.995.

We want to be careful about what that means, because the temptation to over-read it is strong. Cosine similarity on a twelve-analyte vector is a blunt instrument; the median pair in this set already scores 0.82, so high values are not by themselves remarkable. Reporting to two decimal places compresses real differences. And a house growing under one light recipe, one nutrient regime and one cure will legitimately converge — terroir is real and it flattens genetic distinctions.

What we can say is narrower and still worth saying. Twenty-one analytes reported to two decimals is a coarse enough instrument that two different cultivars can come out indistinguishable on it, and here two of them did. Under the thirty-nine-analyte panel this brand's certificates carried in early 2024, the same two lots would have been measured on nerolidol, eucalyptol, terpinen-4-ol, 3-carene and fourteen other compounds — exactly the minor analytes that separate cultivars from one another. That data no longer exists. The narrower panel does not create the similarity. It removes the means of telling whether the similarity is real.

The third of the business with no lab data at all

The certificate page has one heading: FLOWER. Nothing else.

Per Headset's point-of-sale panel for July 2026, pre-rolls are 28.8% of this brand's revenue and eight of its ten best-selling SKUs. The brand operates live product pages for pre-rolls and for edibles. Weedmaps carries more than forty infused pre-roll listings — infused joints and infused blunts advertising potencies from roughly 38% to 45% THC — plus two gummy SKUs at 10 mg.

There is no published certificate for any of it. Not one pre-roll, not one infused blunt, not one gummy. The infused-blunt potency figures — the highest numbers the brand puts in front of a customer anywhere — are unverifiable from any public document.

This is the same gap we documented in Volume 09, and it keeps appearing for the same structural reason: brands publish flower certificates because flower buyers ask for them, and stop there. New York requires a scannable code linking to a downloadable certificate on every package under §128.5(b)(9) — so the documents exist. They are simply not where anyone can read them before deciding what to buy.

Thirty claims, and a rule that changed in December

The brand's Strains page carries symptom-relief language on thirty distinct cultivar entries. Four invoke medical patients explicitly. Verbatim, live as of 15 August 2026:

One point of fairness that belongs here rather than in a footnote. Part 129 binds licensees. Dank by Definition holds no licence. The regulated party is HPI, and any exposure is HPI's, not the brand's as such. We also make no claim that OCM has taken or will take any action; we found none, and we say below why the absence of a record proves little.

What the certificate page itself gets wrong

A publication that audits disclosure has to audit the disclosure mechanism, and this one is unreliable as a map from strain to document.

• Humboldt Pound Cake links to Earthshaker OG's certificate. The page spells it "Humbolt Poundcake" and points to lab ID AL40917003-006 — byte-identical to the Earthshaker OG entry, batch code DANK-26324-FLW-3.5G-ESOG. There is no Humboldt Pound Cake certificate anywhere on the site.
• Skywalker OG's second link is LA Kush Cake's certificate (AL60121006-019). Cherry Pie's first link is Boss Drank's (AL30809002-012).
• Three entries are dead — Gushers 1g, Alaskan Thunderfuck DK03724FLW-ATF03 and Vanilla Kush print batch numbers as plain text with no link at all. No Vanilla Kush certificate is published anywhere.
• Gas Face's link is labelled with Permanent Markers' batch number. The destination is correct; the label is not.
• One certificate contradicts itself about the cultivar. AL50308002-004 carries a printed revision note changing the product name from "Permanent Marker" to "Sharpie" — the header now says Sharpie, the strain line still says Permanent Marker, the batch code still ends -PM. The correction note quotes the original typo, "Sugrown," inside itself.
• Two batches were remediated and retested — Cake Boss ("the batch was remediated and retested with passing results") and Afghani ("sample passed for Micro"). Both revision notes name the retest lab IDs. Neither retest is published on the certificate page, and neither failing original is either. What a reader sees is a passing certificate with a note referring to a document they cannot open.
• The "Label Claim Verification" panel is blank on twelve of fifteen current-template certificates — status prints PASSED, method prints N/A, and the analyte table has no rows.

There is a real story in that footnote, but it is about the regulator. New York mandates a test, sets no standard for it, prescribes no vocabulary for reporting it, and permits six licensed laboratories to use five different undefined tokens — two of which print an affirmative "PASS" against a limit that does not exist. Meanwhile OCM's own consumer guide tells New Yorkers the answer will be either PASSED or FAILED. We will take that up separately.

Claims against documents

The number nobody in that ecosystem is reporting

Headset's July 2026 snapshot, published 4 August, is the one genuinely independent measurement available. It is a point-of-sale panel with a stated margin of error, not a census, and we treat it as such.

Headset's own reporting has this brand at #1 in New York overall in twenty of twenty-four months from March 2024 to February 2026. Five months later it is third in its own core category, its year-over-year decline is steeper than its two-year decline — meaning the fall is accelerating — and its flower is selling for half what it did a year ago.

We are not going to pretend a terpene panel caused that. New York's market doubled its licensed retail footprint over the same period and prices fell across the board. But it is worth noticing what the brands that passed it have in common, and worth noticing that not one of the dozens of retailer pages describing this brand as New York's finest mentions any of these numbers.

What we could not verify

• Who controls the licensee. New York does not publish officers or shareholders of a business corporation, and OCM's true-party-of-interest disclosures are not publicly searchable. George Karavias is independently and extensively documented as founder of Dream Hospitality Group, and Honeysuckle reports Dank as the product of a partnership with that company — but no cannabis coverage names him personally, and neither man's ownership stake in any licensed entity is verifiable from public record. Edward Kirkham was identified on camera by News 12 Hudson Valley in October 2022 as chief operating officer of HPI Canna Inc.
• Twenty-seven batches out of hundreds. The certificate library is far larger than our sample. We selected for breadth across cultivars and recency, not randomly, and a random sample could differ.
• The panel-narrowing is Kaycha's decision, not demonstrably the brand's. We found no evidence either way about whether HPI requested, was consulted on, or even noticed the change.
• Cross-year terpene comparison is invalid and we have not attempted it. Three different panel widths across the period mean a 2023 total and a 2026 total are not the same measurement.
• Pricing is thin. Two live menu data points plus Headset's averages. Weedmaps geolocation suppressed most New York pricing.
• No public record of enforcement — which is not a clean record. No recall names this brand or either HPI entity; we read the current and past recall registers in full. But OCM publishes no licensee-level enforcement register at all. Its enforcement pages cover the illicit market and report aggregates. Notices of violation, fines and administrative decisions against licensees are not published anywhere searchable, and the delinquent-payment list is restricted to licensed suppliers. Absence of a public record here means very little.

The shape of it

Dank by Definition did the hard part. It built a real certificate library at a real laboratory, it prints Metrc identifiers and manifest numbers, it publishes revision notes admitting when a batch was remediated, and across three years and three document templates its numbers have never once failed to add up. In a market where we have documented brands with no laboratory at all, that is not a small thing and it should not be flattened into a takedown.

And every year the window has gotten smaller. Thirty-nine analytes became twenty-one. The catalogue grew to 438 SKUs while the certificate page stayed at one word: flower. The best-selling products in the line — infused blunts advertising forty-five percent THC — have no published chemistry at all. The flower a shopper is pushed toward by the biggest number is the flower carrying the least aroma. And thirty strain pages still tell readers what the plant will do for their pain, eight months after New York made that a freestanding prohibition.

None of that is fraud. All of it is a company that used to show more of its work than it shows now, at exactly the moment its sales fell by half. We will publish any response in full, and we will revise this piece the day a pre-roll certificate appears.